[
{
"path": ".gitignore",
"content": "a# Byte-compiled / optimized / DLL files\n__pycache__/\n*.py[cod]\n*$py.class\n\n# C extensions\n*.so\n\n# Distribution / packaging\n.Python\nbuild/\ndevelop-eggs/\ndist/\ndownloads/\neggs/\n.eggs/\nlib/\nlib64/\nparts/\nsdist/\nvar/\nwheels/\nshare/python-wheels/\n*.egg-info/\n.installed.cfg\n*.egg\nMANIFEST\n\n# PyInstaller\n# Usually these files are written by a python script from a template\n# before PyInstaller builds the exe, so as to inject date/other infos into it.\n*.manifest\n*.spec\n\n# Installer logs\npip-log.txt\npip-delete-this-directory.txt\n\n# Unit test / coverage reports\nhtmlcov/\n.tox/\n.nox/\n.coverage\n.coverage.*\n.cache\nnosetests.xml\ncoverage.xml\n*.cover\n*.py,cover\n.hypothesis/\n.pytest_cache/\ncover/\n\n# Translations\n*.mo\n*.pot\n\n# Django stuff:\n*.log\nlocal_settings.py\ndb.sqlite3\ndb.sqlite3-journal\n\n# Flask stuff:\ninstance/\n.webassets-cache\n\n# Scrapy stuff:\n.scrapy\n\n# Sphinx documentation\ndocs/_build/\n\n# PyBuilder\n.pybuilder/\ntarget/\n\n# Jupyter Notebook\n.ipynb_checkpoints\n\n# IPython\nprofile_default/\nipython_config.py\n\n# pyenv\n# For a library or package, you might want to ignore these files since the code is\n# intended to run in multiple environments; otherwise, check them in:\n# .python-version\n\n# pipenv\n# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.\n# However, in case of collaboration, if having platform-specific dependencies or dependencies\n# having no cross-platform support, pipenv may install dependencies that don't work, or not\n# install all needed dependencies.\n#Pipfile.lock\n\n# PEP 582; used by e.g. github.com/David-OConnor/pyflow\n__pypackages__/\n\n# Celery stuff\ncelerybeat-schedule\ncelerybeat.pid\n\n# SageMath parsed files\n*.sage.py\n\n# Environments\n.env\n.venv\nenv/\nvenv/\nENV/\nenv.bak/\nvenv.bak/\n\n# Spyder project settings\n.spyderproject\n.spyproject\n\n# Rope project settings\n.ropeproject\n\n# mkdocs documentation\n/site\n\n# mypy\n.mypy_cache/\n.dmypy.json\ndmypy.json\n\n# Pyre type checker\n.pyre/\n\n# pytype static type analyzer\n.pytype/\n\n# Cython debug symbols\ncython_debug/\n\n\n\n__pycache__\nbuild\ndist\n*.egg-info\n*.vscode/\n*.pth\ntests\ncheckpoints\ndatasets\nruns\ncache\n*.out\n*.o\ndata\nfigures/*.pdf\n\n\n"
},
{
"path": ".gitmodules",
"content": "[submodule \"thirdparty/lietorch\"]\n\tpath = thirdparty/lietorch\n\turl = https://github.com/princeton-vl/lietorch\n[submodule \"thirdparty/eigen\"]\n\tpath = thirdparty/eigen\n\turl = https://gitlab.com/libeigen/eigen.git\n"
},
{
"path": "LICENSE",
"content": " GNU GENERAL PUBLIC LICENSE\n Version 3, 29 June 2007\n\n Copyright (C) 2007 Free Software Foundation, Inc. \n Everyone is permitted to copy and distribute verbatim copies\n of this license document, but changing it is not allowed.\n\n Preamble\n\n The GNU General Public License is a free, copyleft license for\nsoftware and other kinds of works.\n\n The licenses for most software and other practical works are designed\nto take away your freedom to share and change the works. By contrast,\nthe GNU General Public License is intended to guarantee your freedom to\nshare and change all versions of a program--to make sure it remains free\nsoftware for all its users. We, the Free Software Foundation, use the\nGNU General Public License for most of our software; it applies also to\nany other work released this way by its authors. You can apply it to\nyour programs, too.\n\n When we speak of free software, we are referring to freedom, not\nprice. Our General Public Licenses are designed to make sure that you\nhave the freedom to distribute copies of free software (and charge for\nthem if you wish), that you receive source code or can get it if you\nwant it, that you can change the software or use pieces of it in new\nfree programs, and that you know you can do these things.\n\n To protect your rights, we need to prevent others from denying you\nthese rights or asking you to surrender the rights. Therefore, you have\ncertain responsibilities if you distribute copies of the software, or if\nyou modify it: responsibilities to respect the freedom of others.\n\n For example, if you distribute copies of such a program, whether\ngratis or for a fee, you must pass on to the recipients the same\nfreedoms that you received. You must make sure that they, too, receive\nor can get the source code. And you must show them these terms so they\nknow their rights.\n\n Developers that use the GNU GPL protect your rights with two steps:\n(1) assert copyright on the software, and (2) offer you this License\ngiving you legal permission to copy, distribute and/or modify it.\n\n For the developers' and authors' protection, the GPL clearly explains\nthat there is no warranty for this free software. For both users' and\nauthors' sake, the GPL requires that modified versions be marked as\nchanged, so that their problems will not be attributed erroneously to\nauthors of previous versions.\n\n Some devices are designed to deny users access to install or run\nmodified versions of the software inside them, although the manufacturer\ncan do so. This is fundamentally incompatible with the aim of\nprotecting users' freedom to change the software. The systematic\npattern of such abuse occurs in the area of products for individuals to\nuse, which is precisely where it is most unacceptable. Therefore, we\nhave designed this version of the GPL to prohibit the practice for those\nproducts. If such problems arise substantially in other domains, we\nstand ready to extend this provision to those domains in future versions\nof the GPL, as needed to protect the freedom of users.\n\n Finally, every program is threatened constantly by software patents.\nStates should not allow patents to restrict development and use of\nsoftware on general-purpose computers, but in those that do, we wish to\navoid the special danger that patents applied to a free program could\nmake it effectively proprietary. To prevent this, the GPL assures that\npatents cannot be used to render the program non-free.\n\n The precise terms and conditions for copying, distribution and\nmodification follow.\n\n TERMS AND CONDITIONS\n\n 0. Definitions.\n\n \"This License\" refers to version 3 of the GNU General Public License.\n\n \"Copyright\" also means copyright-like laws that apply to other kinds of\nworks, such as semiconductor masks.\n\n \"The Program\" refers to any copyrightable work licensed under this\nLicense. Each licensee is addressed as \"you\". \"Licensees\" and\n\"recipients\" may be individuals or organizations.\n\n To \"modify\" a work means to copy from or adapt all or part of the work\nin a fashion requiring copyright permission, other than the making of an\nexact copy. The resulting work is called a \"modified version\" of the\nearlier work or a work \"based on\" the earlier work.\n\n A \"covered work\" means either the unmodified Program or a work based\non the Program.\n\n To \"propagate\" a work means to do anything with it that, without\npermission, would make you directly or secondarily liable for\ninfringement under applicable copyright law, except executing it on a\ncomputer or modifying a private copy. Propagation includes copying,\ndistribution (with or without modification), making available to the\npublic, and in some countries other activities as well.\n\n To \"convey\" a work means any kind of propagation that enables other\nparties to make or receive copies. Mere interaction with a user through\na computer network, with no transfer of a copy, is not conveying.\n\n An interactive user interface displays \"Appropriate Legal Notices\"\nto the extent that it includes a convenient and prominently visible\nfeature that (1) displays an appropriate copyright notice, and (2)\ntells the user that there is no warranty for the work (except to the\nextent that warranties are provided), that licensees may convey the\nwork under this License, and how to view a copy of this License. If\nthe interface presents a list of user commands or options, such as a\nmenu, a prominent item in the list meets this criterion.\n\n 1. Source Code.\n\n The \"source code\" for a work means the preferred form of the work\nfor making modifications to it. \"Object code\" means any non-source\nform of a work.\n\n A \"Standard Interface\" means an interface that either is an official\nstandard defined by a recognized standards body, or, in the case of\ninterfaces specified for a particular programming language, one that\nis widely used among developers working in that language.\n\n The \"System Libraries\" of an executable work include anything, other\nthan the work as a whole, that (a) is included in the normal form of\npackaging a Major Component, but which is not part of that Major\nComponent, and (b) serves only to enable use of the work with that\nMajor Component, or to implement a Standard Interface for which an\nimplementation is available to the public in source code form. A\n\"Major Component\", in this context, means a major essential component\n(kernel, window system, and so on) of the specific operating system\n(if any) on which the executable work runs, or a compiler used to\nproduce the work, or an object code interpreter used to run it.\n\n The \"Corresponding Source\" for a work in object code form means all\nthe source code needed to generate, install, and (for an executable\nwork) run the object code and to modify the work, including scripts to\ncontrol those activities. However, it does not include the work's\nSystem Libraries, or general-purpose tools or generally available free\nprograms which are used unmodified in performing those activities but\nwhich are not part of the work. For example, Corresponding Source\nincludes interface definition files associated with source files for\nthe work, and the source code for shared libraries and dynamically\nlinked subprograms that the work is specifically designed to require,\nsuch as by intimate data communication or control flow between those\nsubprograms and other parts of the work.\n\n The Corresponding Source need not include anything that users\ncan regenerate automatically from other parts of the Corresponding\nSource.\n\n The Corresponding Source for a work in source code form is that\nsame work.\n\n 2. Basic Permissions.\n\n All rights granted under this License are granted for the term of\ncopyright on the Program, and are irrevocable provided the stated\nconditions are met. This License explicitly affirms your unlimited\npermission to run the unmodified Program. The output from running a\ncovered work is covered by this License only if the output, given its\ncontent, constitutes a covered work. This License acknowledges your\nrights of fair use or other equivalent, as provided by copyright law.\n\n You may make, run and propagate covered works that you do not\nconvey, without conditions so long as your license otherwise remains\nin force. You may convey covered works to others for the sole purpose\nof having them make modifications exclusively for you, or provide you\nwith facilities for running those works, provided that you comply with\nthe terms of this License in conveying all material for which you do\nnot control copyright. Those thus making or running the covered works\nfor you must do so exclusively on your behalf, under your direction\nand control, on terms that prohibit them from making any copies of\nyour copyrighted material outside their relationship with you.\n\n Conveying under any other circumstances is permitted solely under\nthe conditions stated below. Sublicensing is not allowed; section 10\nmakes it unnecessary.\n\n 3. Protecting Users' Legal Rights From Anti-Circumvention Law.\n\n No covered work shall be deemed part of an effective technological\nmeasure under any applicable law fulfilling obligations under article\n11 of the WIPO copyright treaty adopted on 20 December 1996, or\nsimilar laws prohibiting or restricting circumvention of such\nmeasures.\n\n When you convey a covered work, you waive any legal power to forbid\ncircumvention of technological measures to the extent such circumvention\nis effected by exercising rights under this License with respect to\nthe covered work, and you disclaim any intention to limit operation or\nmodification of the work as a means of enforcing, against the work's\nusers, your or third parties' legal rights to forbid circumvention of\ntechnological measures.\n\n 4. Conveying Verbatim Copies.\n\n You may convey verbatim copies of the Program's source code as you\nreceive it, in any medium, provided that you conspicuously and\nappropriately publish on each copy an appropriate copyright notice;\nkeep intact all notices stating that this License and any\nnon-permissive terms added in accord with section 7 apply to the code;\nkeep intact all notices of the absence of any warranty; and give all\nrecipients a copy of this License along with the Program.\n\n You may charge any price or no price for each copy that you convey,\nand you may offer support or warranty protection for a fee.\n\n 5. Conveying Modified Source Versions.\n\n You may convey a work based on the Program, or the modifications to\nproduce it from the Program, in the form of source code under the\nterms of section 4, provided that you also meet all of these conditions:\n\n a) The work must carry prominent notices stating that you modified\n it, and giving a relevant date.\n\n b) The work must carry prominent notices stating that it is\n released under this License and any conditions added under section\n 7. This requirement modifies the requirement in section 4 to\n \"keep intact all notices\".\n\n c) You must license the entire work, as a whole, under this\n License to anyone who comes into possession of a copy. This\n License will therefore apply, along with any applicable section 7\n additional terms, to the whole of the work, and all its parts,\n regardless of how they are packaged. This License gives no\n permission to license the work in any other way, but it does not\n invalidate such permission if you have separately received it.\n\n d) If the work has interactive user interfaces, each must display\n Appropriate Legal Notices; however, if the Program has interactive\n interfaces that do not display Appropriate Legal Notices, your\n work need not make them do so.\n\n A compilation of a covered work with other separate and independent\nworks, which are not by their nature extensions of the covered work,\nand which are not combined with it such as to form a larger program,\nin or on a volume of a storage or distribution medium, is called an\n\"aggregate\" if the compilation and its resulting copyright are not\nused to limit the access or legal rights of the compilation's users\nbeyond what the individual works permit. Inclusion of a covered work\nin an aggregate does not cause this License to apply to the other\nparts of the aggregate.\n\n 6. Conveying Non-Source Forms.\n\n You may convey a covered work in object code form under the terms\nof sections 4 and 5, provided that you also convey the\nmachine-readable Corresponding Source under the terms of this License,\nin one of these ways:\n\n a) Convey the object code in, or embodied in, a physical product\n (including a physical distribution medium), accompanied by the\n Corresponding Source fixed on a durable physical medium\n customarily used for software interchange.\n\n b) Convey the object code in, or embodied in, a physical product\n (including a physical distribution medium), accompanied by a\n written offer, valid for at least three years and valid for as\n long as you offer spare parts or customer support for that product\n model, to give anyone who possesses the object code either (1) a\n copy of the Corresponding Source for all the software in the\n product that is covered by this License, on a durable physical\n medium customarily used for software interchange, for a price no\n more than your reasonable cost of physically performing this\n conveying of source, or (2) access to copy the\n Corresponding Source from a network server at no charge.\n\n c) Convey individual copies of the object code with a copy of the\n written offer to provide the Corresponding Source. This\n alternative is allowed only occasionally and noncommercially, and\n only if you received the object code with such an offer, in accord\n with subsection 6b.\n\n d) Convey the object code by offering access from a designated\n place (gratis or for a charge), and offer equivalent access to the\n Corresponding Source in the same way through the same place at no\n further charge. You need not require recipients to copy the\n Corresponding Source along with the object code. If the place to\n copy the object code is a network server, the Corresponding Source\n may be on a different server (operated by you or a third party)\n that supports equivalent copying facilities, provided you maintain\n clear directions next to the object code saying where to find the\n Corresponding Source. Regardless of what server hosts the\n Corresponding Source, you remain obligated to ensure that it is\n available for as long as needed to satisfy these requirements.\n\n e) Convey the object code using peer-to-peer transmission, provided\n you inform other peers where the object code and Corresponding\n Source of the work are being offered to the general public at no\n charge under subsection 6d.\n\n A separable portion of the object code, whose source code is excluded\nfrom the Corresponding Source as a System Library, need not be\nincluded in conveying the object code work.\n\n A \"User Product\" is either (1) a \"consumer product\", which means any\ntangible personal property which is normally used for personal, family,\nor household purposes, or (2) anything designed or sold for incorporation\ninto a dwelling. In determining whether a product is a consumer product,\ndoubtful cases shall be resolved in favor of coverage. For a particular\nproduct received by a particular user, \"normally used\" refers to a\ntypical or common use of that class of product, regardless of the status\nof the particular user or of the way in which the particular user\nactually uses, or expects or is expected to use, the product. A product\nis a consumer product regardless of whether the product has substantial\ncommercial, industrial or non-consumer uses, unless such uses represent\nthe only significant mode of use of the product.\n\n \"Installation Information\" for a User Product means any methods,\nprocedures, authorization keys, or other information required to install\nand execute modified versions of a covered work in that User Product from\na modified version of its Corresponding Source. The information must\nsuffice to ensure that the continued functioning of the modified object\ncode is in no case prevented or interfered with solely because\nmodification has been made.\n\n If you convey an object code work under this section in, or with, or\nspecifically for use in, a User Product, and the conveying occurs as\npart of a transaction in which the right of possession and use of the\nUser Product is transferred to the recipient in perpetuity or for a\nfixed term (regardless of how the transaction is characterized), the\nCorresponding Source conveyed under this section must be accompanied\nby the Installation Information. But this requirement does not apply\nif neither you nor any third party retains the ability to install\nmodified object code on the User Product (for example, the work has\nbeen installed in ROM).\n\n The requirement to provide Installation Information does not include a\nrequirement to continue to provide support service, warranty, or updates\nfor a work that has been modified or installed by the recipient, or for\nthe User Product in which it has been modified or installed. Access to a\nnetwork may be denied when the modification itself materially and\nadversely affects the operation of the network or violates the rules and\nprotocols for communication across the network.\n\n Corresponding Source conveyed, and Installation Information provided,\nin accord with this section must be in a format that is publicly\ndocumented (and with an implementation available to the public in\nsource code form), and must require no special password or key for\nunpacking, reading or copying.\n\n 7. Additional Terms.\n\n \"Additional permissions\" are terms that supplement the terms of this\nLicense by making exceptions from one or more of its conditions.\nAdditional permissions that are applicable to the entire Program shall\nbe treated as though they were included in this License, to the extent\nthat they are valid under applicable law. If additional permissions\napply only to part of the Program, that part may be used separately\nunder those permissions, but the entire Program remains governed by\nthis License without regard to the additional permissions.\n\n When you convey a copy of a covered work, you may at your option\nremove any additional permissions from that copy, or from any part of\nit. (Additional permissions may be written to require their own\nremoval in certain cases when you modify the work.) You may place\nadditional permissions on material, added by you to a covered work,\nfor which you have or can give appropriate copyright permission.\n\n Notwithstanding any other provision of this License, for material you\nadd to a covered work, you may (if authorized by the copyright holders of\nthat material) supplement the terms of this License with terms:\n\n a) Disclaiming warranty or limiting liability differently from the\n terms of sections 15 and 16 of this License; or\n\n b) Requiring preservation of specified reasonable legal notices or\n author attributions in that material or in the Appropriate Legal\n Notices displayed by works containing it; or\n\n c) Prohibiting misrepresentation of the origin of that material, or\n requiring that modified versions of such material be marked in\n reasonable ways as different from the original version; or\n\n d) Limiting the use for publicity purposes of names of licensors or\n authors of the material; or\n\n e) Declining to grant rights under trademark law for use of some\n trade names, trademarks, or service marks; or\n\n f) Requiring indemnification of licensors and authors of that\n material by anyone who conveys the material (or modified versions of\n it) with contractual assumptions of liability to the recipient, for\n any liability that these contractual assumptions directly impose on\n those licensors and authors.\n\n All other non-permissive additional terms are considered \"further\nrestrictions\" within the meaning of section 10. If the Program as you\nreceived it, or any part of it, contains a notice stating that it is\ngoverned by this License along with a term that is a further\nrestriction, you may remove that term. If a license document contains\na further restriction but permits relicensing or conveying under this\nLicense, you may add to a covered work material governed by the terms\nof that license document, provided that the further restriction does\nnot survive such relicensing or conveying.\n\n If you add terms to a covered work in accord with this section, you\nmust place, in the relevant source files, a statement of the\nadditional terms that apply to those files, or a notice indicating\nwhere to find the applicable terms.\n\n Additional terms, permissive or non-permissive, may be stated in the\nform of a separately written license, or stated as exceptions;\nthe above requirements apply either way.\n\n 8. Termination.\n\n You may not propagate or modify a covered work except as expressly\nprovided under this License. Any attempt otherwise to propagate or\nmodify it is void, and will automatically terminate your rights under\nthis License (including any patent licenses granted under the third\nparagraph of section 11).\n\n However, if you cease all violation of this License, then your\nlicense from a particular copyright holder is reinstated (a)\nprovisionally, unless and until the copyright holder explicitly and\nfinally terminates your license, and (b) permanently, if the copyright\nholder fails to notify you of the violation by some reasonable means\nprior to 60 days after the cessation.\n\n Moreover, your license from a particular copyright holder is\nreinstated permanently if the copyright holder notifies you of the\nviolation by some reasonable means, this is the first time you have\nreceived notice of violation of this License (for any work) from that\ncopyright holder, and you cure the violation prior to 30 days after\nyour receipt of the notice.\n\n Termination of your rights under this section does not terminate the\nlicenses of parties who have received copies or rights from you under\nthis License. If your rights have been terminated and not permanently\nreinstated, you do not qualify to receive new licenses for the same\nmaterial under section 10.\n\n 9. Acceptance Not Required for Having Copies.\n\n You are not required to accept this License in order to receive or\nrun a copy of the Program. Ancillary propagation of a covered work\noccurring solely as a consequence of using peer-to-peer transmission\nto receive a copy likewise does not require acceptance. However,\nnothing other than this License grants you permission to propagate or\nmodify any covered work. These actions infringe copyright if you do\nnot accept this License. Therefore, by modifying or propagating a\ncovered work, you indicate your acceptance of this License to do so.\n\n 10. Automatic Licensing of Downstream Recipients.\n\n Each time you convey a covered work, the recipient automatically\nreceives a license from the original licensors, to run, modify and\npropagate that work, subject to this License. You are not responsible\nfor enforcing compliance by third parties with this License.\n\n An \"entity transaction\" is a transaction transferring control of an\norganization, or substantially all assets of one, or subdividing an\norganization, or merging organizations. If propagation of a covered\nwork results from an entity transaction, each party to that\ntransaction who receives a copy of the work also receives whatever\nlicenses to the work the party's predecessor in interest had or could\ngive under the previous paragraph, plus a right to possession of the\nCorresponding Source of the work from the predecessor in interest, if\nthe predecessor has it or can get it with reasonable efforts.\n\n You may not impose any further restrictions on the exercise of the\nrights granted or affirmed under this License. For example, you may\nnot impose a license fee, royalty, or other charge for exercise of\nrights granted under this License, and you may not initiate litigation\n(including a cross-claim or counterclaim in a lawsuit) alleging that\nany patent claim is infringed by making, using, selling, offering for\nsale, or importing the Program or any portion of it.\n\n 11. Patents.\n\n A \"contributor\" is a copyright holder who authorizes use under this\nLicense of the Program or a work on which the Program is based. The\nwork thus licensed is called the contributor's \"contributor version\".\n\n A contributor's \"essential patent claims\" are all patent claims\nowned or controlled by the contributor, whether already acquired or\nhereafter acquired, that would be infringed by some manner, permitted\nby this License, of making, using, or selling its contributor version,\nbut do not include claims that would be infringed only as a\nconsequence of further modification of the contributor version. For\npurposes of this definition, \"control\" includes the right to grant\npatent sublicenses in a manner consistent with the requirements of\nthis License.\n\n Each contributor grants you a non-exclusive, worldwide, royalty-free\npatent license under the contributor's essential patent claims, to\nmake, use, sell, offer for sale, import and otherwise run, modify and\npropagate the contents of its contributor version.\n\n In the following three paragraphs, a \"patent license\" is any express\nagreement or commitment, however denominated, not to enforce a patent\n(such as an express permission to practice a patent or covenant not to\nsue for patent infringement). To \"grant\" such a patent license to a\nparty means to make such an agreement or commitment not to enforce a\npatent against the party.\n\n If you convey a covered work, knowingly relying on a patent license,\nand the Corresponding Source of the work is not available for anyone\nto copy, free of charge and under the terms of this License, through a\npublicly available network server or other readily accessible means,\nthen you must either (1) cause the Corresponding Source to be so\navailable, or (2) arrange to deprive yourself of the benefit of the\npatent license for this particular work, or (3) arrange, in a manner\nconsistent with the requirements of this License, to extend the patent\nlicense to downstream recipients. \"Knowingly relying\" means you have\nactual knowledge that, but for the patent license, your conveying the\ncovered work in a country, or your recipient's use of the covered work\nin a country, would infringe one or more identifiable patents in that\ncountry that you have reason to believe are valid.\n\n If, pursuant to or in connection with a single transaction or\narrangement, you convey, or propagate by procuring conveyance of, a\ncovered work, and grant a patent license to some of the parties\nreceiving the covered work authorizing them to use, propagate, modify\nor convey a specific copy of the covered work, then the patent license\nyou grant is automatically extended to all recipients of the covered\nwork and works based on it.\n\n A patent license is \"discriminatory\" if it does not include within\nthe scope of its coverage, prohibits the exercise of, or is\nconditioned on the non-exercise of one or more of the rights that are\nspecifically granted under this License. You may not convey a covered\nwork if you are a party to an arrangement with a third party that is\nin the business of distributing software, under which you make payment\nto the third party based on the extent of your activity of conveying\nthe work, and under which the third party grants, to any of the\nparties who would receive the covered work from you, a discriminatory\npatent license (a) in connection with copies of the covered work\nconveyed by you (or copies made from those copies), or (b) primarily\nfor and in connection with specific products or compilations that\ncontain the covered work, unless you entered into that arrangement,\nor that patent license was granted, prior to 28 March 2007.\n\n Nothing in this License shall be construed as excluding or limiting\nany implied license or other defenses to infringement that may\notherwise be available to you under applicable patent law.\n\n 12. No Surrender of Others' Freedom.\n\n If conditions are imposed on you (whether by court order, agreement or\notherwise) that contradict the conditions of this License, they do not\nexcuse you from the conditions of this License. If you cannot convey a\ncovered work so as to satisfy simultaneously your obligations under this\nLicense and any other pertinent obligations, then as a consequence you may\nnot convey it at all. For example, if you agree to terms that obligate you\nto collect a royalty for further conveying from those to whom you convey\nthe Program, the only way you could satisfy both those terms and this\nLicense would be to refrain entirely from conveying the Program.\n\n 13. Use with the GNU Affero General Public License.\n\n Notwithstanding any other provision of this License, you have\npermission to link or combine any covered work with a work licensed\nunder version 3 of the GNU Affero General Public License into a single\ncombined work, and to convey the resulting work. The terms of this\nLicense will continue to apply to the part which is the covered work,\nbut the special requirements of the GNU Affero General Public License,\nsection 13, concerning interaction through a network will apply to the\ncombination as such.\n\n 14. Revised Versions of this License.\n\n The Free Software Foundation may publish revised and/or new versions of\nthe GNU General Public License from time to time. Such new versions will\nbe similar in spirit to the present version, but may differ in detail to\naddress new problems or concerns.\n\n Each version is given a distinguishing version number. If the\nProgram specifies that a certain numbered version of the GNU General\nPublic License \"or any later version\" applies to it, you have the\noption of following the terms and conditions either of that numbered\nversion or of any later version published by the Free Software\nFoundation. If the Program does not specify a version number of the\nGNU General Public License, you may choose any version ever published\nby the Free Software Foundation.\n\n If the Program specifies that a proxy can decide which future\nversions of the GNU General Public License can be used, that proxy's\npublic statement of acceptance of a version permanently authorizes you\nto choose that version for the Program.\n\n Later license versions may give you additional or different\npermissions. However, no additional obligations are imposed on any\nauthor or copyright holder as a result of your choosing to follow a\nlater version.\n\n 15. Disclaimer of Warranty.\n\n THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY\nAPPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT\nHOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM \"AS IS\" WITHOUT WARRANTY\nOF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,\nTHE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR\nPURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM\nIS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF\nALL NECESSARY SERVICING, REPAIR OR CORRECTION.\n\n 16. Limitation of Liability.\n\n IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING\nWILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS\nTHE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY\nGENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE\nUSE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF\nDATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD\nPARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),\nEVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF\nSUCH DAMAGES.\n\n 17. Interpretation of Sections 15 and 16.\n\n If the disclaimer of warranty and limitation of liability provided\nabove cannot be given local legal effect according to their terms,\nreviewing courts shall apply local law that most closely approximates\nan absolute waiver of all civil liability in connection with the\nProgram, unless a warranty or assumption of liability accompanies a\ncopy of the Program in return for a fee.\n\n END OF TERMS AND CONDITIONS\n\n How to Apply These Terms to Your New Programs\n\n If you develop a new program, and you want it to be of the greatest\npossible use to the public, the best way to achieve this is to make it\nfree software which everyone can redistribute and change under these terms.\n\n To do so, attach the following notices to the program. It is safest\nto attach them to the start of each source file to most effectively\nstate the exclusion of warranty; and each file should have at least\nthe \"copyright\" line and a pointer to where the full notice is found.\n\n \n Copyright (C) \n\n This program is free software: you can redistribute it and/or modify\n it under the terms of the GNU General Public License as published by\n the Free Software Foundation, either version 3 of the License, or\n (at your option) any later version.\n\n This program is distributed in the hope that it will be useful,\n but WITHOUT ANY WARRANTY; without even the implied warranty of\n MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the\n GNU General Public License for more details.\n\n You should have received a copy of the GNU General Public License\n along with this program. If not, see .\n\nAlso add information on how to contact you by electronic and paper mail.\n\n If the program does terminal interaction, make it output a short\nnotice like this when it starts in an interactive mode:\n\n Copyright (C) \n This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.\n This is free software, and you are welcome to redistribute it\n under certain conditions; type `show c' for details.\n\nThe hypothetical commands `show w' and `show c' should show the appropriate\nparts of the General Public License. Of course, your program's commands\nmight be different; for a GUI interface, you would use an \"about box\".\n\n You should also get your employer (if you work as a programmer) or school,\nif any, to sign a \"copyright disclaimer\" for the program, if necessary.\nFor more information on this, and how to apply and follow the GNU GPL, see\n.\n\n The GNU General Public License does not permit incorporating your program\ninto proprietary programs. If your program is a subroutine library, you\nmay consider it more useful to permit linking proprietary applications with\nthe library. If this is what you want to do, use the GNU Lesser General\nPublic License instead of this License. But first, please read\n.\n"
},
{
"path": "README.md",
"content": "# DBA-Fusion\n\n>Tightly Integrating Deep Dense Visual Bundle Adjustment with Multiple Sensors for Large-Scale Localization and Mapping\n\n\n\n
![\"\"](\"./assets/abstract.png\")
\n
\n\n
![\"\"](\"./assets/Hv.svg\")
\n
\n\n
![\"\"](\"./assets/0005.gif\")
\n
\n
![\"\"](\"./assets/outdoors6.gif\")
\n
\n
![\"\"](\"./assets/lidar.jpg\")
\n
\n
![\"\"](\"./assets/postprocessing.png\")
\n
\n\n
![\"\"](\"./assets/GREAT.png\")
\n
\n\n
![\"\"](\"./assets/whu.png\")
\n
\n\nThis work is based on [DROID-SLAM](https://github.com/princeton-vl/DROID-SLAM) and [GTSAM](https://github.com/borglab/gtsam). We use evaluation tools from [evo](https://github.com/MichaelGrupp/evo) and 3D visualization tools from [Open3d](https://github.com/MichaelGrupp/evo).\n"
},
{
"path": "batch_kitti360.py",
"content": "import os\nimport subprocess\n\nfor i in ['0000','0002','0003','0004','0005','0006','0009','0010']:\n p = subprocess.Popen(\"python demo_vio_kitti360.py\" +\\\n \" --imagedir=/home/zhouyuxuan/data/2013_05_28_drive_%s_sync/image_00/data_rgb\"%i +\\\n \" --imagestamp=/home/zhouyuxuan/data/2013_05_28_drive_%s_sync/camstamp.txt\"%i +\\\n \" --imupath=/home/zhouyuxuan/data/2013_05_28_drive_%s_sync/imu.txt\"%i +\\\n \" --gtpath=/home/zhouyuxuan/data/2013_05_28_drive_%s_sync/gt_local.txt\"%i +\\\n # \" --enable_h5\" +\\\n # \" --h5path=/home/zhouyuxuan/DROID-SLAM/%s.h5\"%i +\\\n \" --resultpath=results/result_%s.txt\"%i +\\\n \" --calib=calib/kitti_360.txt\" +\\\n \" --stride=2\" +\\\n \" --max_factors=48\" +\\\n \" --active_window=12\" +\\\n \" --frontend_window=5\" +\\\n \" --frontend_radius=2\" +\\\n \" --frontend_nms=1\" +\\\n \" --inac_range=3\" +\\\n \" --visual_only=0\" +\\\n \" --far_threshold=-1\" +\\\n \" --translation_threshold=0.5\" +\\\n \" --mask_threshold=1.0\" +\\\n \" --skip_edge=[-4,-5,-6]\" +\\\n \" --save_pkl\" +\\\n \" --pklpath=results/%s.pkl\"%i +\\\n \" --show_plot\",shell=True)\n p.wait()\n"
},
{
"path": "batch_subt.py",
"content": "import os\nimport subprocess\n\nfor i in [\\\n # 'Handheld1_Folder',\\\n 'Handheld2_Folder',\\\n ]:\n p = subprocess.Popen(\"python demo_vio_subt.py\" +\\\n \" --imagedir=/mnt/e/subt/%s/cam_0\"%i +\\\n \" --imagestamp=/mnt/e/subt/%s/cam_0/timestamps.txt\"%i +\\\n \" --imupath=/mnt/e/subt/%s/imu/imu_data.csv\"%i +\\\n \" --resultpath=results/result_%s.txt\"%i +\\\n \" --calib=calib/subt.txt\" +\\\n \" --stride=8\" +\\\n \" --max_factors=48\" +\\\n \" --active_window=12\" +\\\n \" --frontend_window=5\" +\\\n \" --frontend_radius=2\" +\\\n \" --frontend_nms=1\" +\\\n \" --far_threshold=0.02\" +\\\n \" --inac_range=3\" +\\\n \" --visual_only=0\" +\\\n \" --translation_threshold=0.2\" +\\\n \" --mask_threshold=-1.0\" +\\\n \" --skip_edge=[-4,-5,-6]\" +\\\n \" --save_pkl\" +\\\n \" --pklpath=results/%s.pkl\"%i +\\\n \" --show_plot\",shell=True)\n p.wait()\n"
},
{
"path": "batch_tumvi.py",
"content": "import os\nimport subprocess\n\nfor i in [\\\n 'outdoors1',\\\n 'outdoors2',\\\n 'outdoors3',\\\n 'outdoors4',\\\n 'outdoors5',\\\n 'outdoors6',\\\n 'outdoors7',\\\n 'outdoors8',\\\n 'magistrale1',\\\n 'magistrale2',\\\n 'magistrale3',\\\n 'magistrale4',\\\n 'magistrale5',\\\n 'magistrale6'\n ]:\n p = subprocess.Popen(\"python demo_vio_tumvi.py\" +\\\n \" --imagedir=/mnt/z/tum-vi/dataset-%s_512_16/mav0/cam0/data\"%i +\\\n \" --imagestamp=/mnt/z/tum-vi/dataset-%s_512_16/mav0/cam0/data.csv\"%i +\\\n \" --imupath=/mnt/z/tum-vi/dataset-%s_512_16/mav0/imu0/data.csv\"%i +\\\n \" --gtpath=/mnt/z/tum-vi/dataset-%s_512_16/dso/gt_imu.csv\"%i +\\\n # \" --enable_h5\" +\\\n # \" --h5path=/home/zhouyuxuan/DROID-SLAM/%s.h5\"%i +\\\n \" --resultpath=results/result_%s.txt\"%i +\\\n \" --calib=calib/tumvi.txt\" +\\\n \" --stride=4\" +\\\n \" --max_factors=48\" +\\\n \" --active_window=12\" +\\\n \" --frontend_window=5\" +\\\n \" --frontend_radius=2\" +\\\n \" --frontend_nms=1\" +\\\n \" --far_threshold=0.02\" +\\\n \" --inac_range=3\" +\\\n \" --visual_only=0\" +\\\n \" --translation_threshold=0.2\" +\\\n \" --mask_threshold=-1.0\" +\\\n \" --skip_edge=[-4,-5,-6]\" +\\\n \" --save_pkl\" +\\\n \" --pklpath=results/%s.pkl\"%i +\\\n \" --show_plot\",shell=True)\n p.wait()\n"
},
{
"path": "batch_whu.py",
"content": "import os\nimport subprocess\n\n# VIO\np = subprocess.Popen(\"python demo_vio_whu.py\" +\\\n \" --imagedir=/home/zhouyuxuan/data/WUH1012/cam0\" +\\\n \" --imagestamp=/home/zhouyuxuan/data/WUH1012/camstamp.txt\" +\\\n \" --imupath=/home/zhouyuxuan/data/WUH1012/imu.txt\" +\\\n \" --gtpath=/home/zhouyuxuan/data/WUH1012/IE.txt\" +\\\n \" --resultpath=results/result_whu_vio.txt\" +\\\n \" --calib=calib/1012.txt\" +\\\n \" --stride=2\" +\\\n \" --max_factors=48\" +\\\n \" --active_window=12\" +\\\n \" --frontend_window=5\" +\\\n \" --frontend_radius=2\" +\\\n \" --frontend_nms=1\" +\\\n \" --inac_range=3\" +\\\n \" --visual_only=0\" +\\\n \" --far_threshold=-1\" +\\\n \" --translation_threshold=0.25\" +\\\n \" --mask_threshold=0.0\" +\\\n \" --skip_edge=[]\" +\\\n \" --save_pkl\" +\\\n \" --use_zupt\" +\\\n \" --pklpath=results/whu.pkl\" +\\\n \" --show_plot\",\n shell=True)\np.wait()\n\n# VIO + wheel\np = subprocess.Popen(\"python demo_vio_whu.py\" +\\\n \" --imagedir=/home/zhouyuxuan/data/WUH1012/cam0\" +\\\n \" --imagestamp=/home/zhouyuxuan/data/WUH1012/camstamp.txt\" +\\\n \" --imupath=/home/zhouyuxuan/data/WUH1012/imu.txt\" +\\\n \" --gtpath=/home/zhouyuxuan/data/WUH1012/IE.txt\" +\\\n \" --resultpath=results/result_whu_viow.txt\" +\\\n \" --calib=calib/1012.txt\" +\\\n \" --stride=2\" +\\\n \" --max_factors=48\" +\\\n \" --active_window=12\" +\\\n \" --frontend_window=5\" +\\\n \" --frontend_radius=2\" +\\\n \" --frontend_nms=1\" +\\\n \" --inac_range=3\" +\\\n \" --visual_only=0\" +\\\n \" --far_threshold=-1\" +\\\n \" --translation_threshold=0.25\" +\\\n \" --mask_threshold=0.0\" +\\\n \" --skip_edge=[]\" +\\\n \" --save_pkl\" +\\\n \" --use_odo\" +\\\n \" --odopath=/home/zhouyuxuan/data/WUH1012/odo_synthesis.txt\" +\\\n \" --pklpath=results/whu.pkl\" +\\\n \" --show_plot\",\n shell=True)\np.wait()\n\n# VIO + GNSS\np = subprocess.Popen(\"python demo_vio_whu.py\" +\\\n \" --imagedir=/home/zhouyuxuan/data/WUH1012/cam0\" +\\\n \" --imagestamp=/home/zhouyuxuan/data/WUH1012/camstamp.txt\" +\\\n \" --imupath=/home/zhouyuxuan/data/WUH1012/imu.txt\" +\\\n \" --gtpath=/home/zhouyuxuan/data/WUH1012/IE.txt\" +\\\n \" --resultpath=results/result_whu_viog.txt\" +\\\n \" --calib=calib/1012.txt\" +\\\n \" --stride=2\" +\\\n \" --max_factors=48\" +\\\n \" --active_window=12\" +\\\n \" --frontend_window=5\" +\\\n \" --frontend_radius=2\" +\\\n \" --frontend_nms=1\" +\\\n \" --inac_range=3\" +\\\n \" --visual_only=0\" +\\\n \" --far_threshold=-1\" +\\\n \" --translation_threshold=0.25\" +\\\n \" --mask_threshold=0.0\" +\\\n \" --skip_edge=[]\" +\\\n \" --save_pkl\" +\\\n \" --use_gnss\" +\\\n \" --gnsspath=/home/zhouyuxuan/data/data_20221012103154/SEPT-PVT.flt\" +\\\n \" --pklpath=results/whu.pkl\" +\\\n \" --show_plot\",\n shell=True)\np.wait()\n"
},
{
"path": "calib/0412.txt",
"content": "889.32868436 889.32868436 515.73648834 202.43873596"
},
{
"path": "calib/0412new.txt",
"content": "885.839465 882.512623 505.509972 389.860117 -0.125551 0.065179 -0.000074 -0.000698\n"
},
{
"path": "calib/1012.txt",
"content": "890.21388839 889.56330572 512.88196119 381.38486858 -0.13095809 0.06640391 -0.00094794 0.0003389\n"
},
{
"path": "calib/barn.txt",
"content": "1161.545689 1161.545689 960.000000 540.000000 -0.025158 0.0 0.0 0.0"
},
{
"path": "calib/carla.txt",
"content": "886.8100 886.8100 512 256"
},
{
"path": "calib/eth.txt",
"content": "726.21081542969 726.21081542969 359.2048034668 202.47247314453"
},
{
"path": "calib/euroc.txt",
"content": "458.654 457.296 367.215 248.375 -0.28340811 0.07395907 0.00019359 1.76187114e-05"
},
{
"path": "calib/handheld.txt",
"content": "531.0895358407821 530.9183032386885 511.3708876141611 399.1276554093305 -0.3367182787437319 0.10679061024072911 0.0003055063102509499 0.0009756613499403765"
},
{
"path": "calib/kitti_360.txt",
"content": "788.629315 786.382230 687.158398 317.752196 -0.344441 0.141678 0.000414 -0.000222 -0.029608"
},
{
"path": "calib/subt.txt",
"content": "758.3153257832925 676.6492212772476 318.27111164892506 239.79816832491474 1.583106303248484 -0.059098218173967695 0.1793477408661115 0.0016819528105368057 -0.0005887999624264534"
},
{
"path": "calib/tartan.txt",
"content": "320.0 320.0 320.0 240.0"
},
{
"path": "calib/tum3.txt",
"content": "535.4 539.2 320.1 247.6"
},
{
"path": "calib/tumvi.txt",
"content": "190.97847715128717 190.9733070521226 254.93170605935475 256.8974428996504 0.0034823894022493434 0.0007150348452162257 -0.0020532361418706202 0.00020293673591811182"
},
{
"path": "dataset/euroc_to_hdf5.py",
"content": "from tqdm import tqdm\nimport numpy as np\nimport torch\nimport cv2\nimport os\nimport argparse\n\nimport h5py\nimport pickle\n\ndef show_image(image):\n image = image.permute(1, 2, 0).cpu().numpy()\n cv2.imshow('image', image / 255.0)\n cv2.waitKey(1)\n\ndef image_stream(imagedir, imagestamp, h5path, calib, stride):\n \"\"\" image generator \"\"\"\n\n calib = np.loadtxt(calib, delimiter=\" \")\n fx, fy, cx, cy = calib[:4]\n\n K = np.eye(3)\n K[0,0] = fx\n K[0,2] = cx\n K[1,1] = fy\n K[1,2] = cy\n\n Kn = np.eye(3)\n Kn[0,0] = fx \n Kn[0,2] = cx \n Kn[1,1] = fy \n Kn[1,2] = cy\n\n image_list = sorted(os.listdir(imagedir))[::stride]\n image_stamps = np.loadtxt(imagestamp,str,delimiter=',')\n image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))\n h5_f = h5py.File(h5path,'w')\n for t, imfile in enumerate(image_list):\n image = cv2.imread(os.path.join(imagedir, imfile))\n\n if len(calib) > 4:\n m1, m2 = cv2.initUndistortRectifyMap(K,calib[4:],np.eye(3),Kn,(image.shape[1],image.shape[0]),cv2.CV_32FC1)\n image = cv2.remap(image, m1, m2, interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_CONSTANT)\n\n tt = float(image_dict[imfile]) /1e9\n\n h0, w0, _ = image.shape\n h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))\n w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))\n\n image = cv2.resize(image, (w1, h1))\n image = image[:h1-h1%8, :w1-w1%8]\n image = torch.as_tensor(image).permute(2, 0, 1)\n\n intrinsics = torch.as_tensor([fx, fy, cx, cy ])\n intrinsics[0::2] *= (w1 / w0)\n intrinsics[1::2] *= (h1 / h0)\n\n h5_f.create_dataset('%.10f'%tt,data = np.fromstring(pickle.dumps((tt, image[None], intrinsics)),dtype='uint8'))\n\n yield tt, image[None], intrinsics\n h5_f.close()\n\nif __name__ == '__main__':\n\n print(torch.cuda.device_count())\n print(torch.cuda.is_available())\n print(torch.cuda.current_device())\n\n parser = argparse.ArgumentParser()\n parser.add_argument(\"--imagedir\", type=str, help=\"path to image directory\")\n parser.add_argument(\"--imagestamp\", type=str, help=\"\")\n parser.add_argument(\"--h5path\", type=str, help=\"\")\n parser.add_argument(\"--calib\", type=str, help=\"path to calibration file\")\n parser.add_argument(\"--stride\", default=4, type=int, help=\"frame stride\")\n parser.add_argument(\"--show_plot\", action=\"store_true\", help=\"\")\n\n args = parser.parse_args()\n\n for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp,\\\n args.h5path, args.calib, args.stride)):\n if args.show_plot:\n show_image(image[0])\n"
},
{
"path": "dataset/kitti360_to_hdf5.py",
"content": "from tqdm import tqdm\nimport numpy as np\nimport torch\nimport cv2\nimport os\nimport argparse\n\nimport h5py\nimport pickle\n\ndef show_image(image):\n image = image.permute(1, 2, 0).cpu().numpy()\n cv2.imshow('image', image / 255.0)\n cv2.waitKey(1)\n\ndef image_stream(imagedir, imagestamp, h5path, calib, stride):\n \"\"\" image generator \"\"\"\n\n calib = np.loadtxt(calib, delimiter=\" \")\n fx, fy, cx, cy = calib[:4]\n\n K = np.eye(3)\n K[0,0] = fx\n K[0,2] = cx\n K[1,1] = fy\n K[1,2] = cy\n\n Kn = np.eye(3)\n Kn[0,0] = fx \n Kn[0,2] = cx \n Kn[1,1] = fy \n Kn[1,2] = cy\n\n image_list = sorted(os.listdir(imagedir))[::stride]\n image_stamps = np.loadtxt(imagestamp,str)\n image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))\n h5_f = h5py.File(h5path,'w')\n for t, imfile in enumerate(image_list):\n image = cv2.imread(os.path.join(imagedir, imfile))\n\n if len(calib) > 4:\n image = cv2.undistort(image, K, calib[4:])\n tt = float(image_dict[imfile])\n\n h0, w0, _ = image.shape\n h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))\n w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))\n\n image = cv2.resize(image, (w1, h1))\n image = image[:h1-h1%8, :w1-w1%8]\n image = torch.as_tensor(image).permute(2, 0, 1)\n\n intrinsics = torch.as_tensor([fx, fy, cx, cy])\n intrinsics[0::2] *= (w1 / w0)\n intrinsics[1::2] *= (h1 / h0)\n\n h5_f.create_dataset('%.10f'%tt,data = np.fromstring(pickle.dumps((tt, image[None], intrinsics)),dtype='uint8'))\n\n yield tt, image[None], intrinsics\n h5_f.close()\n\nif __name__ == '__main__':\n\n print(torch.cuda.device_count())\n print(torch.cuda.is_available())\n print(torch.cuda.current_device())\n\n parser = argparse.ArgumentParser()\n parser.add_argument(\"--imagedir\", type=str, help=\"path to image directory\")\n parser.add_argument(\"--imagestamp\", type=str, help=\"\")\n parser.add_argument(\"--h5path\", type=str, help=\"\")\n parser.add_argument(\"--calib\", type=str, help=\"path to calibration file\")\n parser.add_argument(\"--stride\", default=2, type=int, help=\"frame stride\")\n parser.add_argument(\"--show_plot\", action=\"store_true\", help=\"\")\n\n args = parser.parse_args()\n\n for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp,\\\n args.h5path, args.calib, args.stride)):\n if args.show_plot:\n show_image(image[0])"
},
{
"path": "dataset/tumvi_to_hdf5.py",
"content": "from tqdm import tqdm\nimport numpy as np\nimport torch\nimport cv2\nimport os\nimport argparse\n\nimport h5py\nimport pickle\n\ndef show_image(image):\n image = image.permute(1, 2, 0).cpu().numpy()\n cv2.imshow('image', image / 255.0)\n cv2.waitKey(1)\n\ndef image_stream(imagedir, imagestamp, h5path, calib, stride):\n \"\"\" image generator \"\"\"\n\n calib = np.loadtxt(calib, delimiter=\" \")\n fx, fy, cx, cy = calib[:4]\n\n K = np.eye(3)\n K[0,0] = fx\n K[0,2] = cx\n K[1,1] = fy\n K[1,2] = cy\n\n Kn = np.eye(3)\n Kn[0,0] = fx \n Kn[0,2] = cx \n Kn[1,1] = fy \n Kn[1,2] = cy\n\n image_list = sorted(os.listdir(imagedir))[::stride]\n image_stamps = np.loadtxt(imagestamp,str,delimiter=',')\n image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))\n h5_f = h5py.File(h5path,'w')\n for t, imfile in enumerate(image_list):\n image = cv2.imread(os.path.join(imagedir, imfile))\n\n if len(calib) > 4:\n m1, m2 = cv2.fisheye.initUndistortRectifyMap(K,calib[4:],np.eye(3),Kn,(512,512),cv2.CV_32FC1)\n image = cv2.remap(image, m1, m2, interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_CONSTANT)\n\n tt = float(image_dict[imfile]) /1e9\n\n h0, w0, _ = image.shape\n h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))\n w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))\n\n image = cv2.resize(image, (w1, h1))\n image = image[:h1-h1%8, :w1-w1%8]\n image = torch.as_tensor(image).permute(2, 0, 1)\n\n intrinsics = torch.as_tensor([fx, fy, cx, cy ])\n intrinsics[0::2] *= (w1 / w0)\n intrinsics[1::2] *= (h1 / h0)\n\n h5_f.create_dataset('%.10f'%tt,data = np.fromstring(pickle.dumps((tt, image[None], intrinsics)),dtype='uint8'))\n\n yield tt, image[None], intrinsics\n h5_f.close()\n\nif __name__ == '__main__':\n\n print(torch.cuda.device_count())\n print(torch.cuda.is_available())\n print(torch.cuda.current_device())\n\n parser = argparse.ArgumentParser()\n parser.add_argument(\"--imagedir\", type=str, help=\"path to image directory\")\n parser.add_argument(\"--imagestamp\", type=str, help=\"\")\n parser.add_argument(\"--h5path\", type=str, help=\"\")\n parser.add_argument(\"--calib\", type=str, help=\"path to calibration file\")\n parser.add_argument(\"--stride\", default=4, type=int, help=\"frame stride\")\n parser.add_argument(\"--show_plot\", action=\"store_true\", help=\"\")\n\n args = parser.parse_args()\n\n for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp,\\\n args.h5path, args.calib, args.stride)):\n if args.show_plot:\n show_image(image[0])\n"
},
{
"path": "dbaf/covisible_graph.py",
"content": "import torch\nimport lietorch\nimport numpy as np\n\nimport matplotlib.pyplot as plt\nfrom lietorch import SE3\nfrom modules.corr import CorrBlock, AltCorrBlock\nimport geom.projective_ops as pops\nimport matplotlib.pyplot as plt\nimport cv2\nfrom depth_video import DepthVideo\nimport matplotlib.cm as cm\nimport matplotlib\n\nclass CovisibleGraph:\n def __init__(self, video: DepthVideo, update_op, device=\"cuda:0\", corr_impl=\"volume\", args = None):\n self.video = video\n self.update_op = update_op\n self.device = device\n self.max_factors = args.max_factors\n self.corr_impl = corr_impl\n self.upsample = args.upsample\n\n # operator at 1/8 resolution\n self.ht = ht = video.ht // 8\n self.wd = wd = video.wd // 8\n\n self.coords0 = pops.coords_grid(ht, wd, device=device)\n self.ii = torch.as_tensor([], dtype=torch.long, device=device)\n self.jj = torch.as_tensor([], dtype=torch.long, device=device)\n self.age = torch.as_tensor([], dtype=torch.long, device=device)\n\n self.corr, self.net, self.inp = None, None, None\n self.damping = 1e-6 * torch.ones_like(self.video.disps)\n\n self.target = torch.zeros([1, 0, ht, wd, 2], device=device, dtype=torch.float)\n self.weight = torch.zeros([1, 0, ht, wd, 2], device=device, dtype=torch.float)\n\n # inactive factors\n self.ii_inac = torch.as_tensor([], dtype=torch.long, device=device)\n self.jj_inac = torch.as_tensor([], dtype=torch.long, device=device)\n self.ii_bad = torch.as_tensor([], dtype=torch.long, device=device)\n self.jj_bad = torch.as_tensor([], dtype=torch.long, device=device)\n\n self.target_inac = torch.zeros([1, 0, ht, wd, 2], device=device, dtype=torch.float)\n self.weight_inac = torch.zeros([1, 0, ht, wd, 2], device=device, dtype=torch.float)\n\n self.far_threshold = args.far_threshold\n self.inac_range = args.inac_range\n self.mask_threshold = args.mask_threshold\n self.img_count = 0\n\n self.skip_edge = args.skip_edge\n self.frontend_window = args.frontend_window\n \n # simple online visualization\n self.show_covisible_graph = False\n self.show_oldest_disparity = False\n self.show_flow_and_weight = False\n\n def __filter_repeated_edges(self, ii, jj):\n \"\"\" remove duplicate edges \"\"\"\n\n keep = torch.zeros(ii.shape[0], dtype=torch.bool, device=ii.device)\n eset = set(\n [(i.item(), j.item()) for i, j in zip(self.ii, self.jj)] +\n [(i.item(), j.item()) for i, j in zip(self.ii_inac, self.jj_inac)])\n\n for k, (i, j) in enumerate(zip(ii, jj)):\n keep[k] = (i.item(), j.item()) not in eset\n\n return ii[keep], jj[keep]\n\n def print_edges(self):\n ii = self.ii.cpu().numpy()\n jj = self.jj.cpu().numpy()\n\n ix = np.argsort(ii)\n ii = ii[ix]\n jj = jj[ix]\n\n w = torch.mean(self.weight, dim=[0,2,3,4]).cpu().numpy()\n w = w[ix]\n for e in zip(ii, jj, w):\n print(e)\n print()\n\n def filter_edges(self):\n \"\"\" remove bad edges \"\"\"\n conf = torch.mean(self.weight, dim=[0,2,3,4])\n mask = (torch.abs(self.ii-self.jj) > 2) & (conf < 0.001)\n\n self.ii_bad = torch.cat([self.ii_bad, self.ii[mask]])\n self.jj_bad = torch.cat([self.jj_bad, self.jj[mask]])\n self.rm_factors(mask, store=False)\n\n def clear_edges(self):\n self.rm_factors(self.ii >= 0)\n self.net = None\n self.inp = None\n\n @torch.cuda.amp.autocast(enabled=True)\n def add_factors(self, ii, jj, remove=False):\n \"\"\" add edges to factor graph \"\"\"\n if not isinstance(ii, torch.Tensor):\n ii = torch.as_tensor(ii, dtype=torch.long, device=self.device)\n\n if not isinstance(jj, torch.Tensor):\n jj = torch.as_tensor(jj, dtype=torch.long, device=self.device)\n\n # remove duplicate edges\n ii, jj = self.__filter_repeated_edges(ii, jj)\n\n if ii.shape[0] == 0:\n return\n\n # place limit on number of factors\n if self.max_factors > 0 and self.ii.shape[0] + ii.shape[0] > self.max_factors \\\n and self.corr is not None and remove:\n \n ix = torch.arange(len(self.age))[torch.argsort(self.age).cpu()]\n self.rm_factors(ix >= self.max_factors - ii.shape[0], store=True)\n\n net = self.video.nets[ii].to(self.device).unsqueeze(0)\n\n # correlation volume for new edges\n if self.corr_impl == \"volume\":\n c = (ii == jj).long()\n fmap1 = self.video.fmaps[ii,0].to(self.device).unsqueeze(0)\n fmap2 = self.video.fmaps[jj,c].to(self.device).unsqueeze(0)\n corr = CorrBlock(fmap1, fmap2)\n self.corr = corr if self.corr is None else self.corr.cat(corr)\n\n inp = self.video.inps[ii].to(self.device).unsqueeze(0)\n self.inp = inp if self.inp is None else torch.cat([self.inp, inp], 1)\n\n with torch.cuda.amp.autocast(enabled=False):\n target, _ = self.video.reproject(ii, jj)\n weight = torch.zeros_like(target)\n\n self.ii = torch.cat([self.ii, ii], 0)\n self.jj = torch.cat([self.jj, jj], 0)\n self.age = torch.cat([self.age, torch.zeros_like(ii)], 0)\n\n # reprojection factors\n self.net = net if self.net is None else torch.cat([self.net, net], 1)\n\n self.target = torch.cat([self.target, target], 1)\n self.weight = torch.cat([self.weight, weight], 1)\n\n @torch.cuda.amp.autocast(enabled=True)\n def rm_factors(self, mask, store=False):\n \"\"\" drop edges from factor graph \"\"\"\n\n # store estimated factors\n if store:\n self.ii_inac = torch.cat([self.ii_inac, self.ii[mask]], 0)\n self.jj_inac = torch.cat([self.jj_inac, self.jj[mask]], 0)\n self.target_inac = torch.cat([self.target_inac, self.target[:,mask]], 1)\n self.weight_inac = torch.cat([self.weight_inac, self.weight[:,mask]], 1)\n\n self.ii = self.ii[~mask]\n self.jj = self.jj[~mask]\n self.age = self.age[~mask]\n \n if self.corr_impl == \"volume\":\n self.corr = self.corr[~mask]\n\n if self.net is not None:\n self.net = self.net[:,~mask]\n\n if self.inp is not None:\n self.inp = self.inp[:,~mask]\n\n self.target = self.target[:,~mask]\n self.weight = self.weight[:,~mask]\n\n\n @torch.cuda.amp.autocast(enabled=True)\n def rm_keyframe(self, ix):\n \"\"\" drop edges from factor graph \"\"\"\n\n\n with self.video.get_lock():\n self.video.images[ix] = self.video.images[ix+1]\n self.video.poses[ix] = self.video.poses[ix+1]\n self.video.disps[ix] = self.video.disps[ix+1]\n self.video.disps_sens[ix] = self.video.disps_sens[ix+1]\n self.video.intrinsics[ix] = self.video.intrinsics[ix+1]\n\n self.video.nets[ix] = self.video.nets[ix+1]\n self.video.inps[ix] = self.video.inps[ix+1]\n self.video.fmaps[ix] = self.video.fmaps[ix+1]\n\n self.video.tstamp[ix] = self.video.tstamp[ix+1] # BUG fix\n\n m = (self.ii_inac == ix) | (self.jj_inac == ix)\n self.ii_inac[self.ii_inac >= ix] -= 1\n self.jj_inac[self.jj_inac >= ix] -= 1\n\n if torch.any(m):\n self.ii_inac = self.ii_inac[~m]\n self.jj_inac = self.jj_inac[~m]\n self.target_inac = self.target_inac[:,~m]\n self.weight_inac = self.weight_inac[:,~m]\n\n m = (self.ii == ix) | (self.jj == ix)\n\n self.ii[self.ii >= ix] -= 1\n self.jj[self.jj >= ix] -= 1\n self.rm_factors(m, store=False)\n\n @torch.cuda.amp.autocast(enabled=True)\n def update(self, t0=None, t1=None, itrs=2, use_inactive=False, EP=1e-7, motion_only=False, marg = False):\n \"\"\" run update operator on factor graph \"\"\"\n\n self.video.logger.info('update')\n\n with torch.cuda.amp.autocast(enabled=False):\n coords1, mask = self.video.reproject(self.ii, self.jj)\n motn = torch.cat([coords1 - self.coords0, self.target - coords1], dim=-1)\n motn = motn.permute(0,1,4,2,3).clamp(-64.0, 64.0) # 1,2,4,48,96\n\n corr = self.corr(coords1) \n\n self.net, delta, weight, damping, upmask = \\\n self.update_op(self.net, self.inp, corr, motn, self.ii, self.jj, self.upsample)\n \n if t0 is None:\n t0 = max(1, self.ii.min().item()+1)\n \n self.video.logger.info('ba')\n\n with torch.cuda.amp.autocast(enabled=False):\n self.target = coords1 + delta.to(dtype=torch.float)\n self.weight = weight.to(dtype=torch.float)\n\n ht, wd = self.coords0.shape[0:2]\n if self.upsample: \n self.damping[torch.unique(self.ii)] = damping\n\n if use_inactive:\n m = (self.ii_inac >= t0 - self.inac_range) & (self.jj_inac >= t0 - self.inac_range)\n ii = torch.cat([self.ii_inac[m], self.ii], 0)\n jj = torch.cat([self.jj_inac[m], self.jj], 0)\n target = torch.cat([self.target_inac[:,m], self.target], 1)\n weight = torch.cat([self.weight_inac[:,m], self.weight], 1)\n else:\n ii, jj, target, weight = self.ii, self.jj, self.target, self.weight\n\n # Some real-time visualization for debugging\n # 1) Disparity\n if self.show_oldest_disparity:\n disp_show_front = self.video.disps[self.ii[0]].cpu().numpy()\n disp_show_front = cv2.resize(disp_show_front,[disp_show_front.shape[1]*8,disp_show_front.shape[0]*8],interpolation = cv2.INTER_NEAREST)\n disp_show_front= disp_show_front.astype(np.float32)\n \n normalizer = matplotlib.colors.Normalize(vmin=-0.2, vmax=1.0)\n mapper = cm.ScalarMappable(norm=normalizer,cmap='magma')\n colormapped_im = (mapper.to_rgba(disp_show_front)[:, :, :3] * 255).astype(np.uint8)\n colormapped_im = cv2.cvtColor(colormapped_im,cv2.COLOR_RGB2BGR)\n cv2.imshow('colormapped_im',colormapped_im)\n cv2.waitKey(1)\n\n # 2) Optical flow and weight\n if self.show_flow_and_weight:\n rgb = self.video.images[torch.max(self.ii)].cpu().numpy().transpose(1,2,0)\n new_flow_id = torch.where(torch.logical_and(self.ii==torch.max(self.ii),self.jj==torch.max(self.ii)-5))[0][0].item()\n weight_cpu = weight[0,new_flow_id].cpu().numpy().astype(np.float32)\n weight_cpu = np.linalg.norm(weight_cpu,axis=2)\n normalizer = matplotlib.colors.Normalize(vmin=-0.0, vmax=1.5)\n mapper = cm.ScalarMappable(norm=normalizer,cmap='jet')\n colormapped_im = (mapper.to_rgba(weight_cpu)[:, :, :3] * 255).astype(np.uint8)\n colormapped_im = cv2.cvtColor(colormapped_im,cv2.COLOR_RGB2BGR)\n colormapped_im = cv2.resize(colormapped_im,[rgb.shape[1],rgb.shape[0]])\n colormapped_im = cv2.addWeighted(rgb,0.5,colormapped_im,0.5,0)\n absflow = (self.target[0,new_flow_id] - self.coords0).cpu().numpy()\n for iii in range(0,absflow.shape[0],4):\n for jjj in range(0,absflow.shape[1],4):\n colormapped_im = cv2.line(colormapped_im, (jjj * 8,iii * 8),(int(round((jjj-absflow[iii,jjj,0])* 8)) ,int(round((iii-absflow[iii,jjj,1]) * 8))),(255,255,255),1,cv2.LINE_AA)\n\n cv2.imshow('weight_cpu',colormapped_im)\n self.img_count += 1\n\n # 3) Covisible graph\n if self.show_covisible_graph:\n i0 = min(ii)\n i1 = max(ii)\n ppp = SE3(self.video.poses[i0:(i1+1)]).inv().matrix()[:,0:3,3].cpu().numpy()\n # [:,:3].cpu().numpy()\n scale = max(max(ppp[:,0]) - min(ppp[:,0]),max(ppp[:,1]) - min(ppp[:,1]))\n ppp[:,0] -= np.mean(ppp[:,0])\n ppp[:,1] = -(ppp[:,1]- np.mean(ppp[:,1]))\n ppp *= max(round(1/scale * 200 / 50)*50,50)\n ppp += 500\n mmm = np.zeros([1000,1000],dtype=np.uint8)\n for iii in range(0,i1+1-i0):\n mmm = cv2.circle(mmm,(int(round(ppp[iii,0])),int(round(ppp[iii,1]))),4,255,0)\n for iii in range(self.ii_inac[m].shape[0]):\n iiii = self.ii_inac[m][iii]-i0\n jjjj = self.jj_inac[m][iii]-i0\n mmm = cv2.line(mmm,(int(round(ppp[iiii,0])),int(round(ppp[iiii,1]))),(int(round(ppp[jjjj,0])),int(round(ppp[jjjj,1]))),128,1)\n for iii in range(self.ii.shape[0]):\n iiii = self.ii[iii]-i0\n jjjj = self.jj[iii]-i0\n mmm = cv2.line(mmm,(int(round(ppp[iiii,0])),int(round(ppp[iiii,1]))),(int(round(ppp[jjjj,0])),int(round(ppp[jjjj,1]))),255,1)\n cv2.imshow('window',mmm)\n\n ## Tricks for better performance\n # 1) downweight far points\n if self.far_threshold > 0 and self.video.imu_enabled:\n disp_mask = (self.video.disps < self.far_threshold)\n mask = disp_mask[ii, :, :]\n weight[:, mask] /= 1000.0\n \n # 2) downweight far points\n if self.mask_threshold > 0 and self.video.imu_enabled:\n pose0 = SE3(self.video.poses[ii])\n pose1 = SE3(self.video.poses[jj])\n pose01 = pose0*pose1.inv()\n mask = torch.norm(pose01.translation()[:,:3],dim=1) < self.mask_threshold\n weight[:,mask,:,:,:] /= 1000.0\n \n # 3) downweight edges related to the newest frame\n downweight_newframe = True\n if downweight_newframe:\n weight[:,ii==max(ii)] /= 10.0\n weight[:,jj==max(jj)] /= 4.0\n\n damping = .2 * self.damping[torch.unique(ii)].contiguous() + EP\n\n target = target.view(-1, ht, wd, 2).permute(0,3,1,2).contiguous()\n weight = weight.view(-1, ht, wd, 2).permute(0,3,1,2).contiguous()\n\n # Dense bundle adjustment\n self.video.ba(target, weight, damping, ii, jj, t0, t1, \n itrs=itrs, lm=1e-4, ep=0.1, motion_only=motion_only)\n \n if self.upsample:\n self.video.upsample(torch.unique(self.ii), upmask)\n\n self.age += 1\n\n def add_neighborhood_factors(self, t0, t1, r=3):\n \"\"\" add edges between neighboring frames within radius r \"\"\"\n\n ii, jj = torch.meshgrid(torch.arange(t0,t1), torch.arange(t0,t1))\n ii = ii.reshape(-1).to(dtype=torch.long, device=self.device)\n jj = jj.reshape(-1).to(dtype=torch.long, device=self.device)\n\n c = 1 if self.video.stereo else 0\n\n keep = ((ii - jj).abs() > c) & ((ii - jj).abs() <= r)\n self.add_factors(ii[keep], jj[keep])\n\n \n def add_proximity_factors(self, t0=0, t1=0, rad=2, nms=2, beta=0.25, thresh=16.0, remove=False):\n \"\"\" add edges to the factor graph based on distance \"\"\"\n\n t = self.video.counter.value\n ix = torch.arange(t0, t)\n jx = torch.arange(t1, t)\n\n ii, jj = torch.meshgrid(ix, jx)\n ii = ii.reshape(-1)\n jj = jj.reshape(-1)\n \n cc = ii.shape[0]\n\n # Opportunistic \"skip\" edges in the graph\n if self.skip_edge:\n if torch.max(ii) - torch.min(ii) == self.frontend_window - 1:\n jj_add = torch.min(ii) + torch.tensor(self.skip_edge)\n jj_add = jj_add[jj_add>0]\n ii_add = torch.zeros_like(jj_add) + torch.max(ii)\n jj = torch.cat([jj,jj_add])\n ii = torch.cat([ii,ii_add])\n\n d = self.video.distance(ii, jj, beta=beta)\n d[ii - rad < jj] = np.inf\n d[d > 100] = np.inf\n\n ii1 = torch.cat([self.ii, self.ii_bad, self.ii_inac], 0)\n jj1 = torch.cat([self.jj, self.jj_bad, self.jj_inac], 0)\n for i, j in zip(ii1.cpu().numpy(), jj1.cpu().numpy()):\n for di in range(-nms, nms+1):\n for dj in range(-nms, nms+1):\n if abs(di) + abs(dj) <= max(min(abs(i-j)-2, nms), 0):\n i1 = i + di\n j1 = j + dj\n\n if (t0 <= i1 < t) and (t1 <= j1 < t):\n d[(i1-t0)*(t-t1) + (j1-t1)] = np.inf\n\n es = []\n for i in range(t0, t):\n if self.video.stereo:\n es.append((i, i))\n d[(i-t0)*(t-t1) + (i-t1)] = np.inf\n\n for j in range(max(i-rad-1,0), i):\n es.append((i,j))\n es.append((j,i))\n if (i-t0)*(t-t1) + (j-t1) >=0:\n d[(i-t0)*(t-t1) + (j-t1)] = np.inf\n\n ix = torch.argsort(d)\n for k in ix:\n if k >= cc:\n continue\n\n if d[k].item() > thresh:\n continue\n\n if len(es) > self.max_factors:\n break\n\n i = ii[k]\n j = jj[k]\n \n # bidirectional\n es.append((i, j))\n es.append((j, i))\n\n for di in range(-nms, nms+1):\n for dj in range(-nms, nms+1):\n if abs(di) + abs(dj) <= max(min(abs(i-j)-2, nms), 0):\n i1 = i + di\n j1 = j + dj\n\n if (t0 <= i1 < t) and (t1 <= j1 < t):\n d[(i1-t0)*(t-t1) + (j1-t1)] = np.inf\n \n if ii.shape[0] > cc:\n ix = torch.argsort(d[cc:ii.shape[0]])\n if d[cc + ix[0]] < thresh and d[cc + ix[0]] > 0:\n es.append((ii[cc+ix[0]],jj[cc+ix[0]]))\n es.append((jj[cc+ix[0]],ii[cc+ix[0]]))\n\n ii, jj = torch.as_tensor(es, device=self.device).unbind(dim=-1)\n self.add_factors(ii, jj, remove)\n"
},
{
"path": "dbaf/data_readers/__init__.py",
"content": "\n"
},
{
"path": "dbaf/data_readers/augmentation.py",
"content": "import torch\nimport torchvision.transforms as transforms\nimport numpy as np\nimport torch.nn.functional as F\n\n\nclass RGBDAugmentor:\n \"\"\" perform augmentation on RGB-D video \"\"\"\n\n def __init__(self, crop_size):\n self.crop_size = crop_size\n self.augcolor = transforms.Compose([\n transforms.ToPILImage(),\n transforms.ColorJitter(brightness=0.25, contrast=0.25, saturation=0.25, hue=0.4/3.14),\n transforms.RandomGrayscale(p=0.1),\n transforms.ToTensor()])\n\n self.max_scale = 0.25\n\n def spatial_transform(self, images, depths, poses, intrinsics):\n \"\"\" cropping and resizing \"\"\"\n ht, wd = images.shape[2:]\n\n max_scale = self.max_scale\n min_scale = np.log2(np.maximum(\n (self.crop_size[0] + 1) / float(ht),\n (self.crop_size[1] + 1) / float(wd)))\n\n scale = 2 ** np.random.uniform(min_scale, max_scale)\n intrinsics = scale * intrinsics\n depths = depths.unsqueeze(dim=1)\n\n images = F.interpolate(images, scale_factor=scale, mode='bilinear', \n align_corners=False, recompute_scale_factor=True)\n \n depths = F.interpolate(depths, scale_factor=scale, recompute_scale_factor=True)\n\n # always perform center crop (TODO: try non-center crops)\n y0 = (images.shape[2] - self.crop_size[0]) // 2\n x0 = (images.shape[3] - self.crop_size[1]) // 2\n\n intrinsics = intrinsics - torch.tensor([0.0, 0.0, x0, y0])\n images = images[:, :, y0:y0+self.crop_size[0], x0:x0+self.crop_size[1]]\n depths = depths[:, :, y0:y0+self.crop_size[0], x0:x0+self.crop_size[1]]\n\n depths = depths.squeeze(dim=1)\n return images, poses, depths, intrinsics\n\n def color_transform(self, images):\n \"\"\" color jittering \"\"\"\n num, ch, ht, wd = images.shape\n images = images.permute(1, 2, 3, 0).reshape(ch, ht, wd*num)\n images = 255 * self.augcolor(images[[2,1,0]] / 255.0)\n return images[[2,1,0]].reshape(ch, ht, wd, num).permute(3,0,1,2).contiguous()\n\n def __call__(self, images, poses, depths, intrinsics):\n images = self.color_transform(images)\n return self.spatial_transform(images, depths, poses, intrinsics)\n"
},
{
"path": "dbaf/data_readers/base.py",
"content": "\nimport numpy as np\nimport torch\nimport torch.utils.data as data\nimport torch.nn.functional as F\n\nimport csv\nimport os\nimport cv2\nimport math\nimport random\nimport json\nimport pickle\nimport os.path as osp\n\nfrom .augmentation import RGBDAugmentor\nfrom .rgbd_utils import *\n\nclass RGBDDataset(data.Dataset):\n def __init__(self, name, datapath, n_frames=4, crop_size=[384,512], fmin=8.0, fmax=75.0, do_aug=True):\n \"\"\" Base class for RGBD dataset \"\"\"\n self.aug = None\n self.root = datapath\n self.name = name\n\n self.n_frames = n_frames\n self.fmin = fmin # exclude very easy examples\n self.fmax = fmax # exclude very hard examples\n \n if do_aug:\n self.aug = RGBDAugmentor(crop_size=crop_size)\n\n # building dataset is expensive, cache so only needs to be performed once\n cur_path = osp.dirname(osp.abspath(__file__))\n if not os.path.isdir(osp.join(cur_path, 'cache')):\n os.mkdir(osp.join(cur_path, 'cache'))\n \n cache_path = osp.join(cur_path, 'cache', '{}.pickle'.format(self.name))\n\n if osp.isfile(cache_path):\n scene_info = pickle.load(open(cache_path, 'rb'))[0]\n else:\n scene_info = self._build_dataset()\n with open(cache_path, 'wb') as cachefile:\n pickle.dump((scene_info,), cachefile)\n\n self.scene_info = scene_info\n self._build_dataset_index()\n \n def _build_dataset_index(self):\n self.dataset_index = []\n for scene in self.scene_info:\n if not self.__class__.is_test_scene(scene):\n graph = self.scene_info[scene]['graph']\n for i in graph:\n if len(graph[i][0]) > self.n_frames:\n self.dataset_index.append((scene, i))\n else:\n print(\"Reserving {} for validation\".format(scene))\n\n @staticmethod\n def image_read(image_file):\n return cv2.imread(image_file)\n\n @staticmethod\n def depth_read(depth_file):\n return np.load(depth_file)\n\n def build_frame_graph(self, poses, depths, intrinsics, f=16, max_flow=256):\n \"\"\" compute optical flow distance between all pairs of frames \"\"\"\n def read_disp(fn):\n depth = self.__class__.depth_read(fn)[f//2::f, f//2::f]\n depth[depth < 0.01] = np.mean(depth)\n return 1.0 / depth\n\n poses = np.array(poses)\n intrinsics = np.array(intrinsics) / f\n \n disps = np.stack(list(map(read_disp, depths)), 0)\n d = f * compute_distance_matrix_flow(poses, disps, intrinsics)\n\n # uncomment for nice visualization\n # import matplotlib.pyplot as plt\n # plt.imshow(d)\n # plt.show()\n\n graph = {}\n for i in range(d.shape[0]):\n j, = np.where(d[i] < max_flow)\n graph[i] = (j, d[i,j])\n\n return graph\n\n def __getitem__(self, index):\n \"\"\" return training video \"\"\"\n\n index = index % len(self.dataset_index)\n scene_id, ix = self.dataset_index[index]\n\n frame_graph = self.scene_info[scene_id]['graph']\n images_list = self.scene_info[scene_id]['images']\n depths_list = self.scene_info[scene_id]['depths']\n poses_list = self.scene_info[scene_id]['poses']\n intrinsics_list = self.scene_info[scene_id]['intrinsics']\n\n inds = [ ix ]\n while len(inds) < self.n_frames:\n # get other frames within flow threshold\n k = (frame_graph[ix][1] > self.fmin) & (frame_graph[ix][1] < self.fmax)\n frames = frame_graph[ix][0][k]\n\n # prefer frames forward in time\n if np.count_nonzero(frames[frames > ix]):\n ix = np.random.choice(frames[frames > ix])\n \n elif np.count_nonzero(frames):\n ix = np.random.choice(frames)\n\n inds += [ ix ]\n\n images, depths, poses, intrinsics = [], [], [], []\n for i in inds:\n images.append(self.__class__.image_read(images_list[i]))\n depths.append(self.__class__.depth_read(depths_list[i]))\n poses.append(poses_list[i])\n intrinsics.append(intrinsics_list[i])\n\n images = np.stack(images).astype(np.float32)\n depths = np.stack(depths).astype(np.float32)\n poses = np.stack(poses).astype(np.float32)\n intrinsics = np.stack(intrinsics).astype(np.float32)\n\n images = torch.from_numpy(images).float()\n images = images.permute(0, 3, 1, 2)\n\n disps = torch.from_numpy(1.0 / depths)\n poses = torch.from_numpy(poses)\n intrinsics = torch.from_numpy(intrinsics)\n\n if self.aug is not None:\n images, poses, disps, intrinsics = \\\n self.aug(images, poses, disps, intrinsics)\n\n # scale scene\n if len(disps[disps>0.01]) > 0:\n s = disps[disps>0.01].mean()\n disps = disps / s\n poses[...,:3] *= s\n\n return images, poses, disps, intrinsics \n\n def __len__(self):\n return len(self.dataset_index)\n\n def __imul__(self, x):\n self.dataset_index *= x\n return self\n"
},
{
"path": "dbaf/data_readers/factory.py",
"content": "\nimport pickle\nimport os\nimport os.path as osp\n\n# RGBD-Dataset\nfrom .tartan import TartanAir\n\nfrom .stream import ImageStream\nfrom .stream import StereoStream\nfrom .stream import RGBDStream\n\n# streaming datasets for inference\nfrom .tartan import TartanAirStream\nfrom .tartan import TartanAirTestStream\n\ndef dataset_factory(dataset_list, **kwargs):\n \"\"\" create a combined dataset \"\"\"\n\n from torch.utils.data import ConcatDataset\n\n dataset_map = { 'tartan': (TartanAir, ) }\n db_list = []\n for key in dataset_list:\n # cache datasets for faster future loading\n db = dataset_map[key][0](**kwargs)\n\n print(\"Dataset {} has {} images\".format(key, len(db)))\n db_list.append(db)\n\n return ConcatDataset(db_list)\n \n\ndef create_datastream(dataset_path, **kwargs):\n \"\"\" create data_loader to stream images 1 by 1 \"\"\"\n\n from torch.utils.data import DataLoader\n\n if osp.isfile(osp.join(dataset_path, 'calibration.txt')):\n db = ETH3DStream(dataset_path, **kwargs)\n\n elif osp.isdir(osp.join(dataset_path, 'image_left')):\n db = TartanAirStream(dataset_path, **kwargs)\n\n elif osp.isfile(osp.join(dataset_path, 'rgb.txt')):\n db = TUMStream(dataset_path, **kwargs)\n\n elif osp.isdir(osp.join(dataset_path, 'mav0')):\n db = EurocStream(dataset_path, **kwargs)\n\n elif osp.isfile(osp.join(dataset_path, 'calib.txt')):\n db = KITTIStream(dataset_path, **kwargs)\n\n else:\n # db = TartanAirStream(dataset_path, **kwargs)\n db = TartanAirTestStream(dataset_path, **kwargs)\n \n stream = DataLoader(db, shuffle=False, batch_size=1, num_workers=4)\n return stream\n\n\ndef create_imagestream(dataset_path, **kwargs):\n \"\"\" create data_loader to stream images 1 by 1 \"\"\"\n from torch.utils.data import DataLoader\n\n db = ImageStream(dataset_path, **kwargs)\n return DataLoader(db, shuffle=False, batch_size=1, num_workers=4)\n\ndef create_stereostream(dataset_path, **kwargs):\n \"\"\" create data_loader to stream images 1 by 1 \"\"\"\n from torch.utils.data import DataLoader\n\n db = StereoStream(dataset_path, **kwargs)\n return DataLoader(db, shuffle=False, batch_size=1, num_workers=4)\n\ndef create_rgbdstream(dataset_path, **kwargs):\n \"\"\" create data_loader to stream images 1 by 1 \"\"\"\n from torch.utils.data import DataLoader\n\n db = RGBDStream(dataset_path, **kwargs)\n return DataLoader(db, shuffle=False, batch_size=1, num_workers=4)\n\n"
},
{
"path": "dbaf/data_readers/rgbd_utils.py",
"content": "import numpy as np\nimport os.path as osp\n\nimport torch\nfrom lietorch import SE3\n\nimport geom.projective_ops as pops\nfrom scipy.spatial.transform import Rotation\n\n\ndef parse_list(filepath, skiprows=0):\n \"\"\" read list data \"\"\"\n data = np.loadtxt(filepath, delimiter=' ', dtype=np.unicode_, skiprows=skiprows)\n return data\n\ndef associate_frames(tstamp_image, tstamp_depth, tstamp_pose, max_dt=1.0):\n \"\"\" pair images, depths, and poses \"\"\"\n associations = []\n for i, t in enumerate(tstamp_image):\n if tstamp_pose is None:\n j = np.argmin(np.abs(tstamp_depth - t))\n if (np.abs(tstamp_depth[j] - t) < max_dt):\n associations.append((i, j))\n\n else:\n j = np.argmin(np.abs(tstamp_depth - t))\n k = np.argmin(np.abs(tstamp_pose - t))\n \n if (np.abs(tstamp_depth[j] - t) < max_dt) and \\\n (np.abs(tstamp_pose[k] - t) < max_dt):\n associations.append((i, j, k))\n \n return associations\n\ndef loadtum(datapath, frame_rate=-1):\n \"\"\" read video data in tum-rgbd format \"\"\"\n if osp.isfile(osp.join(datapath, 'groundtruth.txt')):\n pose_list = osp.join(datapath, 'groundtruth.txt')\n \n elif osp.isfile(osp.join(datapath, 'pose.txt')):\n pose_list = osp.join(datapath, 'pose.txt')\n\n else:\n return None, None, None, None\n\n image_list = osp.join(datapath, 'rgb.txt')\n depth_list = osp.join(datapath, 'depth.txt')\n\n calib_path = osp.join(datapath, 'calibration.txt')\n intrinsic = None\n if osp.isfile(calib_path):\n intrinsic = np.loadtxt(calib_path, delimiter=' ')\n intrinsic = intrinsic.astype(np.float64)\n\n image_data = parse_list(image_list)\n depth_data = parse_list(depth_list)\n pose_data = parse_list(pose_list, skiprows=1)\n pose_vecs = pose_data[:,1:].astype(np.float64)\n\n tstamp_image = image_data[:,0].astype(np.float64)\n tstamp_depth = depth_data[:,0].astype(np.float64)\n tstamp_pose = pose_data[:,0].astype(np.float64)\n associations = associate_frames(tstamp_image, tstamp_depth, tstamp_pose)\n\n # print(len(tstamp_image))\n # print(len(associations))\n\n indicies = range(len(associations))[::5]\n\n # indicies = [ 0 ]\n # for i in range(1, len(associations)):\n # t0 = tstamp_image[associations[indicies[-1]][0]]\n # t1 = tstamp_image[associations[i][0]]\n # if t1 - t0 > 1.0 / frame_rate:\n # indicies += [ i ]\n\n images, poses, depths, intrinsics, tstamps = [], [], [], [], []\n for ix in indicies:\n (i, j, k) = associations[ix]\n images += [ osp.join(datapath, image_data[i,1]) ]\n depths += [ osp.join(datapath, depth_data[j,1]) ]\n poses += [ pose_vecs[k] ]\n tstamps += [ tstamp_image[i] ]\n \n if intrinsic is not None:\n intrinsics += [ intrinsic ]\n\n return images, depths, poses, intrinsics, tstamps\n\n\ndef all_pairs_distance_matrix(poses, beta=2.5):\n \"\"\" compute distance matrix between all pairs of poses \"\"\"\n poses = np.array(poses, dtype=np.float32)\n poses[:,:3] *= beta # scale to balence rot + trans\n poses = SE3(torch.from_numpy(poses))\n\n r = (poses[:,None].inv() * poses[None,:]).log()\n return r.norm(dim=-1).cpu().numpy()\n\ndef pose_matrix_to_quaternion(pose):\n \"\"\" convert 4x4 pose matrix to (t, q) \"\"\"\n q = Rotation.from_matrix(pose[:3, :3]).as_quat()\n return np.concatenate([pose[:3, 3], q], axis=0)\n\ndef compute_distance_matrix_flow(poses, disps, intrinsics):\n \"\"\" compute flow magnitude between all pairs of frames \"\"\"\n if not isinstance(poses, SE3):\n poses = torch.from_numpy(poses).float().cuda()[None]\n poses = SE3(poses).inv()\n\n disps = torch.from_numpy(disps).float().cuda()[None]\n intrinsics = torch.from_numpy(intrinsics).float().cuda()[None]\n\n N = poses.shape[1]\n \n ii, jj = torch.meshgrid(torch.arange(N), torch.arange(N))\n ii = ii.reshape(-1).cuda()\n jj = jj.reshape(-1).cuda()\n\n MAX_FLOW = 100.0\n matrix = np.zeros((N, N), dtype=np.float32)\n\n s = 2048\n for i in range(0, ii.shape[0], s):\n flow1, val1 = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s])\n flow2, val2 = pops.induced_flow(poses, disps, intrinsics, jj[i:i+s], ii[i:i+s])\n \n flow = torch.stack([flow1, flow2], dim=2)\n val = torch.stack([val1, val2], dim=2)\n \n mag = flow.norm(dim=-1).clamp(max=MAX_FLOW)\n mag = mag.view(mag.shape[1], -1)\n val = val.view(val.shape[1], -1)\n\n mag = (mag * val).mean(-1) / val.mean(-1)\n mag[val.mean(-1) < 0.7] = np.inf\n\n i1 = ii[i:i+s].cpu().numpy()\n j1 = jj[i:i+s].cpu().numpy()\n matrix[i1, j1] = mag.cpu().numpy()\n\n return matrix\n\n\ndef compute_distance_matrix_flow2(poses, disps, intrinsics, beta=0.4):\n \"\"\" compute flow magnitude between all pairs of frames \"\"\"\n # if not isinstance(poses, SE3):\n # poses = torch.from_numpy(poses).float().cuda()[None]\n # poses = SE3(poses).inv()\n\n # disps = torch.from_numpy(disps).float().cuda()[None]\n # intrinsics = torch.from_numpy(intrinsics).float().cuda()[None]\n\n N = poses.shape[1]\n \n ii, jj = torch.meshgrid(torch.arange(N), torch.arange(N))\n ii = ii.reshape(-1)\n jj = jj.reshape(-1)\n\n MAX_FLOW = 128.0\n matrix = np.zeros((N, N), dtype=np.float32)\n\n s = 2048\n for i in range(0, ii.shape[0], s):\n flow1a, val1a = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s], tonly=True)\n flow1b, val1b = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s])\n flow2a, val2a = pops.induced_flow(poses, disps, intrinsics, jj[i:i+s], ii[i:i+s], tonly=True)\n flow2b, val2b = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s])\n\n flow1 = flow1a + beta * flow1b\n val1 = val1a * val2b\n\n flow2 = flow2a + beta * flow2b\n val2 = val2a * val2b\n \n flow = torch.stack([flow1, flow2], dim=2)\n val = torch.stack([val1, val2], dim=2)\n \n mag = flow.norm(dim=-1).clamp(max=MAX_FLOW)\n mag = mag.view(mag.shape[1], -1)\n val = val.view(val.shape[1], -1)\n\n mag = (mag * val).mean(-1) / val.mean(-1)\n mag[val.mean(-1) < 0.8] = np.inf\n\n i1 = ii[i:i+s].cpu().numpy()\n j1 = jj[i:i+s].cpu().numpy()\n matrix[i1, j1] = mag.cpu().numpy()\n\n return matrix\n"
},
{
"path": "dbaf/data_readers/stream.py",
"content": "\nimport numpy as np\nimport torch\nimport torch.utils.data as data\nimport torch.nn.functional as F\n\nimport csv\nimport os\nimport cv2\nimport math\nimport random\nimport json\nimport pickle\nimport os.path as osp\n\nfrom .rgbd_utils import *\n\nclass RGBDStream(data.Dataset):\n def __init__(self, datapath, frame_rate=-1, image_size=[384,512], crop_size=[0,0]):\n self.datapath = datapath\n self.frame_rate = frame_rate\n self.image_size = image_size\n self.crop_size = crop_size\n self._build_dataset_index()\n \n @staticmethod\n def image_read(image_file):\n return cv2.imread(image_file)\n\n @staticmethod\n def depth_read(depth_file):\n return np.load(depth_file)\n\n def __len__(self):\n return len(self.images)\n\n def __getitem__(self, index):\n \"\"\" return training video \"\"\"\n image = self.__class__.image_read(self.images[index])\n image = torch.from_numpy(image).float()\n image = image.permute(2, 0, 1)\n\n try:\n tstamp = self.tstamps[index]\n except:\n tstamp = index\n\n pose = torch.from_numpy(self.poses[index]).float()\n intrinsic = torch.from_numpy(self.intrinsics[index]).float()\n\n # resize image\n sx = self.image_size[1] / image.shape[2]\n sy = self.image_size[0] / image.shape[1]\n\n image = F.interpolate(image[None], self.image_size, mode='bilinear', align_corners=False)[0]\n\n fx, fy, cx, cy = intrinsic.unbind(dim=0)\n fx, cx = sx * fx, sx * cx\n fy, cy = sy * fy, sy * cy\n \n # crop image\n if self.crop_size[0] > 0:\n cy = cy - self.crop_size[0]\n image = image[:,self.crop_size[0]:-self.crop_size[0],:]\n\n if self.crop_size[1] > 0:\n cx = cx - self.crop_size[1]\n image = image[:,:,self.crop_size[1]:-self.crop_size[1]]\n\n intrinsic = torch.stack([fx, fy, cx, cy])\n\n return tstamp, image, pose, intrinsic\n\n\nclass ImageStream(data.Dataset):\n def __init__(self, datapath, intrinsics, rate=1, image_size=[384,512]):\n rgb_list = osp.join(datapath, 'rgb.txt')\n if os.path.isfile(rgb_list):\n rgb_list = np.loadtxt(rgb_list, delimiter=' ', dtype=np.unicode_)\n self.timestamps = rgb_list[:,0].astype(np.float)\n self.images = [os.path.join(datapath, x) for x in rgb_list[:,1]]\n self.images = self.images[::rate]\n self.timestamps = self.timestamps[::rate]\n\n else:\n import glob\n self.images = sorted(glob.glob(osp.join(datapath, '*.jpg'))) + sorted(glob.glob(osp.join(datapath, '*.png')))\n self.images = self.images[::rate]\n\n self.intrinsics = intrinsics\n self.image_size = image_size\n\n def __len__(self):\n return len(self.images)\n\n @staticmethod\n def image_read(imfile):\n return cv2.imread(imfile)\n\n def __getitem__(self, index):\n \"\"\" return training video \"\"\"\n image = self.__class__.image_read(self.images[index])\n\n try:\n tstamp = self.timestamps[index]\n except:\n tstamp = index\n\n ht0, wd0 = image.shape[:2]\n ht1, wd1 = self.image_size\n\n intrinsics = torch.as_tensor(self.intrinsics)\n intrinsics[0] *= wd1 / wd0\n intrinsics[1] *= ht1 / ht0\n intrinsics[2] *= wd1 / wd0\n intrinsics[3] *= ht1 / ht0\n\n # resize image\n ikwargs = {'mode': 'bilinear', 'align_corners': True}\n image = torch.from_numpy(image).float().permute(2, 0, 1)\n image = F.interpolate(image[None], self.image_size, **ikwargs)[0]\n\n return tstamp, image, intrinsics\n\n\n\nclass StereoStream(data.Dataset):\n def __init__(self, datapath, intrinsics, rate=1, image_size=[384,512], \n map_left=None, map_right=None, left_root='image_left', right_root='image_right'):\n import glob\n self.intrinsics = intrinsics\n self.image_size = image_size\n \n imgs = sorted(glob.glob(osp.join(datapath, left_root, '*.png')))[::rate]\n self.images_l = []\n self.images_r = []\n self.tstamps = []\n\n for img_l in imgs:\n img_r = img_l.replace(left_root, right_root)\n if os.path.isfile(img_r):\n t = np.float(img_l.split('/')[-1].replace('.png', ''))\n self.tstamps.append(t)\n self.images_l += [ img_l ]\n self.images_r += [ img_r ]\n\n self.map_left = map_left\n self.map_right = map_right\n\n def __len__(self):\n return len(self.images_l)\n\n @staticmethod\n def image_read(imfile, imap=None):\n image = cv2.imread(imfile)\n if imap is not None:\n image = cv2.remap(image, imap[0], imap[1], interpolation=cv2.INTER_LINEAR)\n return image\n\n def __getitem__(self, index):\n \"\"\" return training video \"\"\"\n tstamp = self.tstamps[index]\n image_l = self.__class__.image_read(self.images_l[index], self.map_left)\n image_r = self.__class__.image_read(self.images_r[index], self.map_right)\n\n ht0, wd0 = image_l.shape[:2]\n ht1, wd1 = self.image_size\n\n intrinsics = torch.as_tensor(self.intrinsics)\n intrinsics[0] *= wd1 / wd0\n intrinsics[1] *= ht1 / ht0\n intrinsics[2] *= wd1 / wd0\n intrinsics[3] *= ht1 / ht0\n\n image_l = torch.from_numpy(image_l).float().permute(2, 0, 1)\n image_r = torch.from_numpy(image_r).float().permute(2, 0, 1)\n\n # resize image\n ikwargs = {'mode': 'bilinear', 'align_corners': True}\n image_l = F.interpolate(image_l[None], self.image_size, **ikwargs)[0]\n image_r = F.interpolate(image_r[None], self.image_size, **ikwargs)[0]\n\n return tstamp, image_l, image_r, intrinsics\n\n\n\n# class RGBDStream(data.Dataset):\n# def __init__(self, datapath, intrinsics=None, rate=1, image_size=[384,512]):\n# assoc_file = osp.join(datapath, 'associated.txt')\n# assoc_list = np.loadtxt(assoc_file, delimiter=' ', dtype=np.unicode_)\n \n# self.intrinsics = intrinsics\n# self.image_size = image_size\n \n# self.timestamps = assoc_list[:,0].astype(np.float)[::rate]\n# self.images = [os.path.join(datapath, x) for x in assoc_list[:,1]][::rate]\n# self.depths = [os.path.join(datapath, x) for x in assoc_list[:,3]][::rate]\n\n# def __len__(self):\n# return len(self.images)\n\n# @staticmethod\n# def image_read(imfile):\n# return cv2.imread(imfile)\n\n# @staticmethod\n# def depth_read(depth_file):\n# depth = cv2.imread(depth_file, cv2.IMREAD_ANYDEPTH)\n# return depth.astype(np.float32) / 5000.0\n\n# def __getitem__(self, index):\n# \"\"\" return training video \"\"\"\n# tstamp = self.timestamps[index]\n# image = self.__class__.image_read(self.images[index])\n# depth = self.__class__.depth_read(self.depths[index])\n\n# ht0, wd0 = image.shape[:2]\n# ht1, wd1 = self.image_size\n\n# intrinsics = torch.as_tensor(self.intrinsics)\n# intrinsics[0] *= wd1 / wd0\n# intrinsics[1] *= ht1 / ht0\n# intrinsics[2] *= wd1 / wd0\n# intrinsics[3] *= ht1 / ht0\n\n# # resize image\n# ikwargs = {'mode': 'bilinear', 'align_corners': True}\n# image = torch.from_numpy(image).float().permute(2, 0, 1)\n# image = F.interpolate(image[None], self.image_size, **ikwargs)[0]\n\n# depth = torch.from_numpy(depth).float()[None,None]\n# depth = F.interpolate(depth, self.image_size, mode='nearest').squeeze()\n\n# return tstamp, image, depth, intrinsics\n"
},
{
"path": "dbaf/data_readers/tartan.py",
"content": "\nimport numpy as np\nimport torch\nimport glob\nimport cv2\nimport os\nimport os.path as osp\n\nfrom lietorch import SE3\nfrom .base import RGBDDataset\nfrom .stream import RGBDStream\n\ncur_path = osp.dirname(osp.abspath(__file__))\ntest_split = osp.join(cur_path, 'tartan_test.txt')\ntest_split = open(test_split).read().split()\n\n\nclass TartanAir(RGBDDataset):\n\n # scale depths to balance rot & trans\n DEPTH_SCALE = 5.0\n\n def __init__(self, mode='training', **kwargs):\n self.mode = mode\n self.n_frames = 2\n super(TartanAir, self).__init__(name='TartanAir', **kwargs)\n\n @staticmethod \n def is_test_scene(scene):\n # print(scene, any(x in scene for x in test_split))\n return any(x in scene for x in test_split)\n\n def _build_dataset(self):\n from tqdm import tqdm\n print(\"Building TartanAir dataset\")\n\n scene_info = {}\n scenes = glob.glob(osp.join(self.root, '*/*/*/*'))\n for scene in tqdm(sorted(scenes)):\n images = sorted(glob.glob(osp.join(scene, 'image_left/*.png')))\n depths = sorted(glob.glob(osp.join(scene, 'depth_left/*.npy')))\n \n poses = np.loadtxt(osp.join(scene, 'pose_left.txt'), delimiter=' ')\n poses = poses[:, [1, 2, 0, 4, 5, 3, 6]]\n poses[:,:3] /= TartanAir.DEPTH_SCALE\n intrinsics = [TartanAir.calib_read()] * len(images)\n\n # graph of co-visible frames based on flow\n graph = self.build_frame_graph(poses, depths, intrinsics)\n\n scene = '/'.join(scene.split('/'))\n scene_info[scene] = {'images': images, 'depths': depths, \n 'poses': poses, 'intrinsics': intrinsics, 'graph': graph}\n\n return scene_info\n\n @staticmethod\n def calib_read():\n return np.array([320.0, 320.0, 320.0, 240.0])\n\n @staticmethod\n def image_read(image_file):\n return cv2.imread(image_file)\n\n @staticmethod\n def depth_read(depth_file):\n depth = np.load(depth_file) / TartanAir.DEPTH_SCALE\n depth[depth==np.nan] = 1.0\n depth[depth==np.inf] = 1.0\n return depth\n\n\nclass TartanAirStream(RGBDStream):\n def __init__(self, datapath, **kwargs):\n super(TartanAirStream, self).__init__(datapath=datapath, **kwargs)\n\n def _build_dataset_index(self):\n \"\"\" build list of images, poses, depths, and intrinsics \"\"\"\n self.root = 'datasets/TartanAir'\n\n scene = osp.join(self.root, self.datapath)\n image_glob = osp.join(scene, 'image_left/*.png')\n images = sorted(glob.glob(image_glob))\n\n poses = np.loadtxt(osp.join(scene, 'pose_left.txt'), delimiter=' ')\n poses = poses[:, [1, 2, 0, 4, 5, 3, 6]]\n\n poses = SE3(torch.as_tensor(poses))\n poses = poses[[0]].inv() * poses\n poses = poses.data.cpu().numpy()\n\n intrinsic = self.calib_read(self.datapath)\n intrinsics = np.tile(intrinsic[None], (len(images), 1))\n\n self.images = images[::int(self.frame_rate)]\n self.poses = poses[::int(self.frame_rate)]\n self.intrinsics = intrinsics[::int(self.frame_rate)]\n\n @staticmethod\n def calib_read(datapath):\n return np.array([320.0, 320.0, 320.0, 240.0])\n\n @staticmethod\n def image_read(image_file):\n return cv2.imread(image_file)\n\n\nclass TartanAirTestStream(RGBDStream):\n def __init__(self, datapath, **kwargs):\n super(TartanAirTestStream, self).__init__(datapath=datapath, **kwargs)\n\n def _build_dataset_index(self):\n \"\"\" build list of images, poses, depths, and intrinsics \"\"\"\n self.root = 'datasets/mono'\n image_glob = osp.join(self.root, self.datapath, '*.png')\n images = sorted(glob.glob(image_glob))\n\n poses = np.loadtxt(osp.join(self.root, 'mono_gt', self.datapath + '.txt'), delimiter=' ')\n poses = poses[:, [1, 2, 0, 4, 5, 3, 6]]\n\n poses = SE3(torch.as_tensor(poses))\n poses = poses[[0]].inv() * poses\n poses = poses.data.cpu().numpy()\n\n intrinsic = self.calib_read(self.datapath)\n intrinsics = np.tile(intrinsic[None], (len(images), 1))\n\n self.images = images[::int(self.frame_rate)]\n self.poses = poses[::int(self.frame_rate)]\n self.intrinsics = intrinsics[::int(self.frame_rate)]\n\n @staticmethod\n def calib_read(datapath):\n return np.array([320.0, 320.0, 320.0, 240.0])\n\n @staticmethod\n def image_read(image_file):\n return cv2.imread(image_file)"
},
{
"path": "dbaf/data_readers/tartan_test.txt",
"content": "abandonedfactory/abandonedfactory/Easy/P011\nabandonedfactory/abandonedfactory/Hard/P011\nabandonedfactory_night/abandonedfactory_night/Easy/P013\nabandonedfactory_night/abandonedfactory_night/Hard/P014\namusement/amusement/Easy/P008\namusement/amusement/Hard/P007\ncarwelding/carwelding/Easy/P007\nendofworld/endofworld/Easy/P009\ngascola/gascola/Easy/P008\ngascola/gascola/Hard/P009\nhospital/hospital/Easy/P036\nhospital/hospital/Hard/P049\njapanesealley/japanesealley/Easy/P007\njapanesealley/japanesealley/Hard/P005\nneighborhood/neighborhood/Easy/P021\nneighborhood/neighborhood/Hard/P017\nocean/ocean/Easy/P013\nocean/ocean/Hard/P009\noffice2/office2/Easy/P011\noffice2/office2/Hard/P010\noffice/office/Hard/P007\noldtown/oldtown/Easy/P007\noldtown/oldtown/Hard/P008\nseasidetown/seasidetown/Easy/P009\nseasonsforest/seasonsforest/Easy/P011\nseasonsforest/seasonsforest/Hard/P006\nseasonsforest_winter/seasonsforest_winter/Easy/P009\nseasonsforest_winter/seasonsforest_winter/Hard/P018\nsoulcity/soulcity/Easy/P012\nsoulcity/soulcity/Hard/P009\nwesterndesert/westerndesert/Easy/P013\nwesterndesert/westerndesert/Hard/P007\n"
},
{
"path": "dbaf/dbaf.py",
"content": "import torch\nimport lietorch\nimport numpy as np\nfrom droid_net import DroidNet\nfrom depth_video import DepthVideo\nfrom motion_filter import MotionFilter\nfrom dbaf_frontend import DBAFusionFrontend\nfrom collections import OrderedDict\nfrom torch.multiprocessing import Process\n\nfrom lietorch import SE3\nimport geom.projective_ops as pops\nimport droid_backends\nimport pickle\n\nclass DBAFusion:\n def __init__(self, args):\n super(DBAFusion, self).__init__()\n self.load_weights(args.weights) # load DroidNet weights\n self.args = args\n\n # store images, depth, poses, intrinsics (shared between processes)\n self.video = DepthVideo(args.image_size, args.buffer, save_pkl = args.save_pkl, stereo=args.stereo, upsample=args.upsample)\n\n # filter incoming frames so that there is enough motion\n self.filterx = MotionFilter(self.net, self.video, thresh=args.filter_thresh)\n\n # frontend process\n self.frontend = DBAFusionFrontend(self.net, self.video, self.args)\n\n self.pklpath = args.pklpath\n self.upsample = args.upsample\n\n def load_weights(self, weights):\n \"\"\" load trained model weights \"\"\"\n\n print(weights)\n self.net = DroidNet()\n state_dict = OrderedDict([\n (k.replace(\"module.\", \"\"), v) for (k, v) in torch.load(weights).items()])\n\n state_dict[\"update.weight.2.weight\"] = state_dict[\"update.weight.2.weight\"][:2]\n state_dict[\"update.weight.2.bias\"] = state_dict[\"update.weight.2.bias\"][:2]\n state_dict[\"update.delta.2.weight\"] = state_dict[\"update.delta.2.weight\"][:2]\n state_dict[\"update.delta.2.bias\"] = state_dict[\"update.delta.2.bias\"][:2]\n\n self.net.load_state_dict(state_dict)\n self.net.to(\"cuda:0\").eval()\n\n def track(self, tstamp, image, depth=None, intrinsics=None):\n \"\"\" main thread - update map \"\"\"\n\n with torch.no_grad():\n # check there is enough motion\n self.filterx.track(tstamp, image, depth, intrinsics)\n\n # local bundle adjustment\n self.frontend()\n\n def terminate(self, stream=None):\n \"\"\" terminate the visualization process, return poses [t, q] \"\"\"\n del self.frontend\n\n def save_vis_easy(self):\n mcameras = {}\n mpoints = {}\n mstamps = {}\n with torch.no_grad():\n dirty_index = torch.arange(0,self.video.count_save,device='cuda')\n\n stamps= torch.index_select(self.video.tstamp_save, 0 ,dirty_index)\n poses= torch.index_select( self.video.poses_save, 0 ,dirty_index)\n disps= torch.index_select( self.video.disps_save, 0 ,dirty_index)\n images = torch.index_select( self.video.images_save, 0 ,dirty_index)\n Ps = SE3(poses).inv().matrix().cpu().numpy()\n points = droid_backends.iproj(SE3(poses).inv().data, disps, self.video.intrinsics[0]).cpu()\n thresh = 0.4 * torch.ones_like(disps.mean(dim=[1,2])) / 4.0 * (1.0 / torch.median(disps.mean(dim=[1,2])))\n # thresh = 0.4 * torch.ones_like(disps.mean(dim=[1,2])) \n count = droid_backends.depth_filter(\n self.video.poses_save, self.video.disps_save, self.video.intrinsics[0], dirty_index, thresh)\n\n count = count.cpu()\n disps = disps.cpu()\n\n if self.upsample:\n disps_up= torch.index_select( self.video.disps_up_save, 0 ,dirty_index)\n disps_up = disps_up.cpu()\n\n masks = ((count >= 1) & (disps > .5*disps.mean(dim=[1,2], keepdim=True)))\n\n for i in range(len(dirty_index)):\n pose = Ps[i]\n ix = dirty_index[i].item()\n mcameras[ix] = pose\n mask = masks[i].reshape(-1)\n pts = points[i].reshape(-1, 3)[mask].cpu().numpy()\n clr = images[i].reshape(-1, 3)[mask].cpu().numpy()\n stamp = stamps[i].cpu()\n if self.upsample:\n mpoints[ix] = {'pts':pts,'clr':clr,'disp':disps[i].cpu().numpy(),'disps_up':disps_up[i].cpu().numpy(),'rgb':images[i].cpu().numpy()}\n else:\n mpoints[ix] = {'pts':pts,'clr':clr,'disp':disps[i].cpu().numpy(),'rgb':images[i].cpu().numpy()}\n mstamps[ix] = stamp\n ddict = {'points':mpoints,'cameras':mcameras,'stamps':mstamps}\n f_save = open(self.pklpath, 'wb')\n pickle.dump(ddict,f_save) \n\n mcameras = {}\n mpoints = {}\n mstamps = {}\n with torch.no_grad():\n dirty_index = torch.arange(0,self.video.count_save,device='cuda')\n\n stamps= torch.index_select(self.video.tstamp_save, 0 ,dirty_index)\n poses= torch.index_select( self.video.poses_save, 0 ,dirty_index)\n disps= torch.index_select( self.video.disps_save, 0 ,dirty_index)\n images = torch.index_select( self.video.images_save, 0 ,dirty_index)\n Ps = SE3(poses).inv().matrix().cpu().numpy()\n points = droid_backends.iproj(SE3(poses).inv().data, disps, self.video.intrinsics[0]).cpu()\n thresh = 0.4 * torch.ones_like(disps.mean(dim=[1,2]))\n count = droid_backends.depth_filter(\n self.video.poses_save, self.video.disps_save, self.video.intrinsics[0], dirty_index, thresh)\n\n count = count.cpu()\n disps = disps.cpu()\n masks = ((count >= 0) & (disps > .5*disps.mean(dim=[1,2], keepdim=True)))\n\n for i in range(len(dirty_index)):\n pose = Ps[i]\n ix = dirty_index[i].item()\n mcameras[ix] = pose\n mask = masks[i].reshape(-1)\n pts = points[i].reshape(-1, 3)[mask].cpu().numpy()\n clr = images[i].reshape(-1, 3)[mask].cpu().numpy()\n stamp = stamps[i].cpu()\n mpoints[ix] = {'pts':pts,'clr':clr,'disp':disps[i].cpu().numpy(),'rgb':images[i].cpu().numpy()}\n mstamps[ix] = stamp\n ddict = {'points':mpoints,'cameras':mcameras,'stamps':mstamps}\n f_save = open(self.pklpath.split('.')[0] + '_raw.pkl', 'wb')\n pickle.dump(ddict,f_save) \n"
},
{
"path": "dbaf/dbaf_frontend.py",
"content": "import torch\nimport torchvision\nimport numpy as np\n\nfrom lietorch import SE3, SO3\nfrom covisible_graph import CovisibleGraph\nimport matplotlib.pyplot as plt\n\nimport gtsam\nimport math\nimport bisect\nfrom math import atan2, cos, sin\nimport geoFunc.trans as trans\nfrom scipy.spatial.transform import Rotation\n\nclass DBAFusionFrontend:\n def __init__(self, net, video, args):\n self.video = video\n self.update_op = net.update\n self.graph = CovisibleGraph(video, net.update, args=args)\n\n # local optimization window\n self.t0 = 0\n self.t1 = 0\n\n # frontend variables\n self.is_initialized = False\n self.count = 0\n\n self.warmup = args.warmup\n self.vi_warmup = 12\n if 'vi_warmup' in args: self.vi_warmup = args.vi_warmup\n self.beta = args.beta\n self.frontend_nms = args.frontend_nms\n self.keyframe_thresh = args.keyframe_thresh\n self.frontend_window = args.frontend_window\n self.frontend_thresh = args.frontend_thresh\n self.frontend_radius = args.frontend_radius\n\n ### DBAFusion\n self.all_imu = None\n self.cur_imu_ii = 0\n self.is_init = False\n self.all_gnss = None\n self.all_odo = None\n self.all_gt = None\n self.all_gt_keys = None\n self.all_stamp = None\n self.cur_stamp_ii = 0\n self.visual_only = args.visual_only\n self.visual_only_init = False\n self.translation_threshold = 0.0\n self.active_window = args.active_window\n self.high_freq_output = True\n self.zupt = ('use_zupt' in args and args.use_zupt)\n\n if not self.visual_only:\n self.max_age = 25\n self.iters1 = 2\n self.iters2 = 1\n else:\n self.max_age = 25\n self.iters1 = 4\n self.iters2 = 2\n\n # visualization/output\n self.show_plot = args.show_plot\n self.result_file = open(args.resultpath,'wt')\n self.plt_pos = [[],[]] # X, Y\n self.plt_pos_ref = [[],[]] # X, Y\n self.plt_att = [[],[],[]] # pitch, roll, yaw\n self.plt_bg = [[],[],[]] # X, Y, Z\n self.plt_t = []\n self.refTw = np.eye(4,4)\n\n if self.show_plot:\n plt.figure('monitor',figsize=[13,4])\n plt.subplot(1,3,1); plt.gca().set_title('Trajectory')\n plt.gca().set_aspect(1)\n plt.subplot(1,3,2); plt.gca().set_title('Attitude Error/Attitude')\n plt.subplot(1,3,3); plt.gca().set_title('Gyroscope Bias')\n plt.ion()\n plt.pause(0.1)\n\n def get_pose_ref(self, tt:float):\n tt_found = self.all_gt_keys[bisect.bisect(self.all_gt_keys,tt)]\n return tt_found, self.all_gt[tt_found]\n \n def __rollup(self, roll):\n \"\"\" roll up window states to save memory \"\"\"\n self.t1 -= roll\n self.count -= roll\n self.video.counter.value -= roll\n self.video.tstamp = torch.roll(self.video.tstamp ,-roll,0) \n self.video.images = torch.roll(self.video.images ,-roll,0) \n self.video.dirty = torch.roll(self.video.dirty ,-roll,0) \n self.video.red = torch.roll(self.video.red ,-roll,0) \n self.video.poses = torch.roll(self.video.poses ,-roll,0) \n self.video.disps = torch.roll(self.video.disps ,-roll,0) \n self.video.disps_sens = torch.roll(self.video.disps_sens,-roll,0) \n self.video.disps_up = torch.roll(self.video.disps_up ,-roll,0) \n self.video.intrinsics = torch.roll(self.video.intrinsics,-roll,0) \n self.video.fmaps = torch.roll(self.video.fmaps ,-roll,0) \n self.video.nets = torch.roll(self.video.nets ,-roll,0) \n self.video.inps = torch.roll(self.video.inps ,-roll,0) \n self.graph.ii -= roll\n self.graph.jj -= roll\n self.graph.ii_inac -= roll\n self.graph.jj_inac -= roll\n rm_inac_index = torch.logical_and(torch.greater_equal(self.graph.ii_inac,0),torch.greater_equal(self.graph.jj_inac,0))\n self.graph.ii_inac = self.graph.ii_inac[rm_inac_index]\n self.graph.jj_inac = self.graph.jj_inac[rm_inac_index]\n self.graph.target_inac = self.graph.target_inac[:,rm_inac_index,:,:,:]\n self.graph.weight_inac = self.graph.weight_inac[:,rm_inac_index,:,:,:] # need test\n\n self.graph.ii_bad -= roll\n self.graph.jj_bad -= roll\n\n self.video.last_t0 -= roll\n self.video.last_t1 -= roll\n self.video.cur_ii -= roll\n self.video.cur_jj -= roll\n if self.video.imu_enabled:\n graph_temp = gtsam.NonlinearFactorGraph()\n for i in range(self.video.cur_graph.size()):\n f = self.video.cur_graph.at(i)\n graph_temp.push_back(f.rekey((np.array(f.keys())-roll).tolist()))\n self.video.cur_graph = graph_temp\n result_temp = gtsam.Values()\n for i in self.video.cur_result.keys():\n if gtsam.Symbol(i).chr() == ord('b'):\n result_temp.insert(i-roll,self.video.cur_result.atConstantBias(i))\n elif gtsam.Symbol(i).chr() == ord('v'):\n result_temp.insert(i-roll,self.video.cur_result.atVector(i))\n elif gtsam.Symbol(i).chr() == ord('x'):\n result_temp.insert(i-roll,self.video.cur_result.atPose3(i))\n else:\n raise Exception()\n self.video.cur_result = result_temp\n self.video.marg_factor = self.video.marg_factor.rekey((np.array(self.video.marg_factor.keys())-roll).tolist())\n\n self.video.state.timestamps = self.video.state.timestamps [roll:]\n self.video.state.wTbs = self.video.state.wTbs [roll:]\n self.video.state.vs = self.video.state.vs [roll:]\n self.video.state.bs = self.video.state.bs [roll:]\n self.video.state.preintegrations = self.video.state.preintegrations [roll:]\n self.video.state.preintegrations_meas = self.video.state.preintegrations_meas [roll:]\n self.video.state.gnss_valid = self.video.state.gnss_valid [roll:]\n self.video.state.gnss_position = self.video.state.gnss_position [roll:]\n self.video.state.odo_valid = self.video.state.odo_valid [roll:]\n self.video.state.odo_vel = self.video.state.odo_vel [roll:]\n\n def __update(self):\n \"\"\" add edges, perform update \"\"\"\n self.count += 1\n self.t1 += 1\n\n if self.video.imu_enabled and (self.video.tstamp[self.t1-1] - self.video.vi_init_time > 5.0):\n self.video.reinit = True\n self.video.vi_init_time = 1e9\n\n ## new frame comes, append IMU\n cur_t = float(self.video.tstamp[self.t1-1].detach().cpu())\n self.video.logger.info('predict %f' %cur_t)\n\n while self.all_imu[self.cur_imu_ii][0] < cur_t:\n ## high-frequency output\n # predict the pose of skipped frames through IMU preintegration\n if self.high_freq_output and self.video.imu_enabled: \n if self.all_imu[self.cur_imu_ii][0] > float(self.all_stamp[self.cur_stamp_ii][0]):\n self.video.state.append_imu_temp(float(self.all_stamp[self.cur_stamp_ii][0]),\\\n self.all_imu[self.cur_imu_ii][4:7],\\\n self.all_imu[self.cur_imu_ii][1:4]/180*math.pi,True)\n if float(self.all_stamp[self.cur_stamp_ii][0]) > self.video.state.timestamps[-1] and\\\n math.fabs(cur_t - float(self.all_stamp[self.cur_stamp_ii][0]))>1e-3:\n pose_temp = self.video.state.pose_temp\n ppp = pose_temp.pose().translation()\n qqq = Rotation.from_matrix(pose_temp.pose().rotation().matrix()).as_quat()\n line = '%.6f %.6f %.6f %.6f %.6f %.6f %.6f %.6f'%(float(self.all_stamp[self.cur_stamp_ii][0]),ppp[0],ppp[1],ppp[2]\\\n ,qqq[0],qqq[1],qqq[2],qqq[3])\n if self.video.gnss_init_t1>0:\n p = self.video.ten0 + np.matmul(trans.Cen(self.video.ten0), ppp)\n line += ' %.6f %.6f %.6f'% (p[0],p[1],p[2]) \n self.result_file.writelines(line+'\\n')\n # self.result_file.flush()\n self.cur_stamp_ii += 1\n self.video.state.append_imu_temp(self.all_imu[self.cur_imu_ii][0],\\\n self.all_imu[self.cur_imu_ii][4:7],\\\n self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)\n \n self.video.state.append_imu(self.all_imu[self.cur_imu_ii][0],\\\n self.all_imu[self.cur_imu_ii][4:7],\\\n self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)\n self.cur_imu_ii += 1\n self.video.state.append_imu(cur_t,\\\n self.all_imu[self.cur_imu_ii][4:7],\\\n self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)\n self.video.state.append_img(cur_t)\n \n ## append GNSS\n if len(self.all_gnss) > 0: gnss_found = bisect.bisect(self.all_gnss[:,0],cur_t - 1e-6)\n else: gnss_found = -1 \n if gnss_found > 0 and self.all_gnss[gnss_found,0] - cur_t < 0.01 :\n self.video.state.append_gnss(cur_t,self.all_gnss[gnss_found,1:4])\n\n ## append ZUPT\n if self.zupt and self.video.state.preintegrations[self.t1-3].deltaTij() > 3.0:\n if np.linalg.norm(self.video.state.vs[self.t1-2]) < 0.025:\n self.video.state.append_odo(cur_t,np.array([.0,.0,.0]))\n\n ## append ODO\n if len(self.all_odo) > 0: odo_found = bisect.bisect(self.all_odo[:,0],cur_t - 1e-6)\n else: odo_found = -1 \n if odo_found > 0 and self.all_odo[odo_found,0] - cur_t < 0.01 :\n self.video.state.append_odo(cur_t,self.all_odo[odo_found,1:4])\n\n self.video.state.append_imu(self.all_imu[self.cur_imu_ii][0],\\\n self.all_imu[self.cur_imu_ii][4:7],\\\n self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)\n self.cur_imu_ii += 1\n\n ## predict pose (<5 ms)\n if self.video.imu_enabled:\n Twc = (self.video.state.wTbs[-1] * self.video.Tbc).matrix()\n TTT = torch.tensor(np.linalg.inv(Twc))\n q = torch.tensor(Rotation.from_matrix(TTT[:3, :3]).as_quat())\n t = TTT[:3,3]\n self.video.poses[self.t1-1] = torch.cat([t,q])\n\n self.video.logger.info('manage edges')\n\n ## manage edges (60 ms)\n if self.graph.corr is not None:\n if self.visual_only:\n self.graph.rm_factors(torch.logical_and(self.graph.age > self.max_age,\\\n torch.logical_or(self.graph.ii < self.t1-self.active_window,self.graph.jj < self.t1-self.active_window)), store=True)\n else:\n self.graph.rm_factors(torch.logical_or(self.graph.age > self.max_age,\\\n torch.logical_or(self.graph.ii < self.t1-self.active_window,self.graph.jj < self.t1-self.active_window)), store=True)\n\n self.graph.add_proximity_factors(self.t1-5, max(self.t1-self.frontend_window, 0), \n rad=self.frontend_radius, nms=self.frontend_nms, thresh=self.frontend_thresh, beta=self.beta, remove=True)\n\n self.video.logger.info('non-keyframes %d' % self.graph.ii.shape[0])\n\n ## non-keyframe update\n self.video.disps[self.t1-1] = torch.where(self.video.disps_sens[self.t1-1] > 0, \n self.video.disps_sens[self.t1-1], self.video.disps[self.t1-1])\n\n for itr in range(self.iters1):\n self.graph.update(None, None, use_inactive=True)\n\n self.rollup = False\n if self.t1 > 65:\n self.__rollup(30)\n print('rollup ',self.graph.ii)\n self.rollup = True\n\n self.video.logger.info('output')\n\n ## visualization/output\n poses = SE3(self.video.poses)\n d = self.video.distance([self.t1-3], [self.t1-2], beta=self.beta, bidirectional=True)\n TTT = np.matmul(poses[self.t1-1].cpu().inv().matrix(),np.linalg.inv(self.video.Ti1c))\n if self.video.imu_enabled or (self.visual_only and self.visual_only_init):\n ppp = TTT[0:3,3]\n qqq = Rotation.from_matrix(TTT[:3, :3]).as_quat()\n line = '%.6f %.6f %.6f %.6f %.6f %.6f %.6f %.6f'%(cur_t,ppp[0],ppp[1],ppp[2]\\\n ,qqq[0],qqq[1],qqq[2],qqq[3])\n if self.video.gnss_init_t1>0:\n p = self.video.ten0 + np.matmul(trans.Cen(self.video.ten0), ppp.numpy())\n line += ' %.6f %.6f %.6f'% (p[0],p[1],p[2]) \n self.result_file.writelines(line+'\\n')\n self.result_file.flush()\n\n TTTref = np.matmul(self.refTw,TTT)\n ppp = TTTref[0:3,3]\n if self.show_plot:\n # if math.fabs(tt_found - cur_t) < 0.1: # for kitti and whu\n self.plt_pos[0].append(ppp[0])\n self.plt_pos[1].append(ppp[1])\n a1 = np.array(trans.m2att(TTTref[0:3,0:3]) )* 57.3\n if self.all_gt is not None:\n tt_found,dd = self.get_pose_ref(cur_t -1e-3)\n self.plt_pos_ref[0].append(dd['T'][0,3])\n self.plt_pos_ref[1].append(dd['T'][1,3]) \n a2 = np.array(trans.m2att(dd['T'][0:3,0:3]) )* 57.3 \n a1 -= a2\n self.plt_att[0].append(a1[0])\n self.plt_att[1].append(a1[1])\n self.plt_att[2].append(a1[2])\n bg = self.video.state.bs[self.t1-1].gyroscope()\n self.plt_bg[0].append(bg[0])\n self.plt_bg[1].append(bg[1])\n self.plt_bg[2].append(bg[2])\n self.plt_t.append(cur_t)\n \n if self.rollup:\n plt.subplot(1,3,1)\n plt.cla(); plt.gca().set_title('Trajectory')\n plt.plot(self.plt_pos[0],self.plt_pos[1],marker='^')\n plt.plot(self.plt_pos_ref[0],self.plt_pos_ref[1],marker='^')\n plt.subplot(1,3,2)\n plt.cla(); plt.gca().set_title('Attitude Error/Attitude')\n plt.plot(self.plt_t,self.plt_att[0],c='r')\n plt.plot(self.plt_t,self.plt_att[1],c='g')\n plt.plot(self.plt_t,self.plt_att[2],c='b')\n plt.ylim([-10,10])\n plt.subplot(1,3,3)\n plt.cla(); plt.gca().set_title('Gyroscope Bias')\n plt.plot(self.plt_t,self.plt_bg[0],c='r')\n plt.plot(self.plt_t,self.plt_bg[1],c='g')\n plt.plot(self.plt_t,self.plt_bg[2],c='b')\n plt.pause(0.1)\n\n\n ## keyframe update\n self.video.logger.info('keyframes %d' % self.graph.ii.shape[0])\n if self.t1 > 10:\n cam_translation = torch.norm((poses[(self.t1-10):(self.t1-3)] * poses[self.t1-2].inv()[None]).translation()[:,0:3],dim=1)\n else:\n cam_translation = torch.norm((poses[(self.t1-6):(self.t1-3)] * poses[self.t1-2].inv()[None]).translation()[:,0:3],dim=1)\n\n if (d.item() < self.keyframe_thresh or (self.video.imu_enabled and torch.sum(cam_translation < self.translation_threshold)>0)): # gnss\n self.video.logger.info('remove new frame!!!!!!!!!!!!1')\n self.graph.rm_keyframe(self.t1 - 2)\n\n # merge preintegration[self.t1-2] and preintegration[self.t1-3]\n for iii in range(len(self.video.state.preintegrations_meas[self.t1-2])):\n dd = self.video.state.preintegrations_meas[self.t1-2][iii]\n if dd[2] > 0:\n self.video.state.preintegrations[self.t1-3].integrateMeasurement(dd[0],\\\n dd[1],\\\n dd[2])\n self.video.state.preintegrations_meas[self.t1-3].append(dd)\n \n self.video.state.preintegrations[self.t1-2] = self.video.state.preintegrations[self.t1-1]\n self.video.state.preintegrations_meas[self.t1-2] = self.video.state.preintegrations_meas[self.t1-1]\n self.video.state.preintegrations.pop()\n self.video.state.preintegrations_meas.pop()\n\n self.video.rm_new_gnss(self.t1-2)\n self.video.state.wTbs[self.t1-2] = self.video.state.wTbs[self.t1-1]; self.video.state.wTbs.pop()\n self.video.state.bs [self.t1-2] = self.video.state.bs [self.t1-1]; self.video.state.bs.pop()\n self.video.state.vs [self.t1-2] = self.video.state.vs [self.t1-1]; self.video.state.vs .pop()\n self.video.state.gnss_valid [self.t1-2] = self.video.state.gnss_valid [self.t1-1]; self.video.state.gnss_valid .pop()\n self.video.state.gnss_position [self.t1-2] = self.video.state.gnss_position [self.t1-1]; self.video.state.gnss_position .pop()\n self.video.state.odo_valid [self.t1-2] = self.video.state.odo_valid [self.t1-1]; self.video.state.odo_valid .pop()\n self.video.state.odo_vel [self.t1-2] = self.video.state.odo_vel [self.t1-1]; self.video.state.odo_vel .pop() \n\n with self.video.get_lock():\n self.video.counter.value -= 1\n self.t1 -= 1\n else:\n for itr in range(self.iters2):\n # print('b%d' % itr)\n self.graph.update(None, None, use_inactive=True)\n\n ## try initializing VI/GNSS\n if self.t1 > self.vi_warmup and self.video.vi_init_t1 < 0:\n self.init_VI()\n if not self.visual_only:\n for i in range(len(self.all_stamp)): # skip to next image\n if float(self.all_stamp[i][0]) < cur_t + 1e-6: continue\n else:\n self.cur_stamp_ii = i\n break\n if self.video.imu_enabled and self.video.gnss_init_time <= 0.0 and len(self.all_gnss)>0:\n self.init_GNSS()\n\n ## set pose for next itration\n self.video.poses[self.t1] = self.video.poses[self.t1-1]\n self.video.disps[self.t1] = self.video.disps[self.t1-1].mean() * 1.0\n\n self.video.dirty[self.graph.ii.min():self.t1] = True\n\n def init_IMU(self):\n \"\"\" initialize IMU states \"\"\"\n cur_t = float(self.video.tstamp[self.t0].detach().cpu())\n for i in range(len(self.all_imu)):\n if self.all_imu[i][0] < cur_t - 1e-6: continue\n else:\n self.cur_imu_ii = i\n break\n\n for i in range(self.t0,self.t1):\n tt = self.video.tstamp[i]\n if i == self.t0:\n self.video.state.init_first_state(cur_t,np.zeros(3),\\\n np.eye(3),\\\n np.zeros(3))\n self.video.state.append_imu(self.all_imu[self.cur_imu_ii][0],\\\n self.all_imu[self.cur_imu_ii][4:7],\\\n self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)\n self.cur_imu_ii += 1\n self.is_init = True\n else:\n cur_t = float(self.video.tstamp[i].detach().cpu())\n while self.all_imu[self.cur_imu_ii][0] < cur_t:\n self.video.state.append_imu(self.all_imu[self.cur_imu_ii][0],\\\n self.all_imu[self.cur_imu_ii][4:7],\\\n self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)\n self.cur_imu_ii += 1\n self.video.state.append_imu(cur_t,\\\n self.all_imu[self.cur_imu_ii][4:7],\\\n self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)\n self.video.state.append_img(cur_t)\n \n if len(self.all_gnss) > 0: gnss_found = bisect.bisect(self.all_gnss[:,0],cur_t - 1e-6)\n else: gnss_found = -1\n if gnss_found > 0 and self.all_gnss[gnss_found,0] - cur_t < 0.01:\n self.video.state.append_gnss(cur_t,self.all_gnss[gnss_found,1:4])\n\n if len(self.all_odo) > 0: odo_found = bisect.bisect(self.all_odo[:,0],cur_t - 1e-6)\n else: odo_found = -1 \n if odo_found > 0 and self.all_odo[odo_found,0] - cur_t < 0.01 :\n self.video.state.append_odo(cur_t,self.all_odo[odo_found,1:4])\n\n self.video.state.append_imu(self.all_imu[self.cur_imu_ii][0],\\\n self.all_imu[self.cur_imu_ii][4:7],\\\n self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)\n \n self.cur_imu_ii += 1\n Twc = np.matmul(np.array([[1,0,0,0],\\\n [0,1,0,0],\\\n [0,0,1,0.02*i],\\\n [0,0,0,1]]),self.video.Ti1c) # perturb the camera poses, which benefits the robustness of initial BA\n TTT = torch.tensor(np.linalg.inv(Twc))\n q = torch.tensor(Rotation.from_matrix(TTT[:3, :3]).as_quat())\n t = TTT[:3,3]\n if not self.video.imu_enabled:\n self.video.poses[i] = torch.cat([t,q])\n\n def init_VI(self):\n \"\"\" initialize the V-I system, referring to VIN-Fusion \"\"\"\n sum_g = np.zeros(3,dtype = np.float64)\n ccount = 0\n for i in range(self.t1 - 8 ,self.t1-1):\n dt = self.video.state.preintegrations[i].deltaTij()\n tmp_g = self.video.state.preintegrations[i].deltaVij()/dt\n sum_g += tmp_g\n ccount += 1\n aver_g = sum_g * 1.0 / ccount\n var_g = 0.0\n for i in range(self.t1 - 8 ,self.t1-1):\n dt = self.video.state.preintegrations[i].deltaTij()\n tmp_g = self.video.state.preintegrations[i].deltaVij()/dt\n var_g += np.linalg.norm(tmp_g - aver_g)**2\n var_g =math.sqrt(var_g/ccount)\n if var_g < 0.25:\n print(\"IMU excitation not enough!\",var_g)\n else:\n poses = SE3(self.video.poses)\n self.plt_pos = [[],[]]\n self.plt_pos_ref = [[],[]]\n for i in range(0,self.t1):\n ppp = np.matmul(poses[i].cpu().inv().matrix(),np.linalg.inv(self.video.Ti1c))[0:3,3]\n self.plt_pos[0].append(ppp[0])\n self.plt_pos[1].append(ppp[1])\n if self.all_gt is not None:\n tt_found,dd = self.get_pose_ref(self.video.tstamp[i]-1e-3)\n self.plt_pos_ref[0].append(dd['T'][0,3])\n self.plt_pos_ref[1].append(dd['T'][1,3]) \n\n if self.show_plot:\n plt.subplot(1,3,1) \n plt.cla(); plt.gca().set_title('Trajectory')\n plt.plot(self.plt_pos[0],self.plt_pos[1],marker='^')\n plt.plot(self.plt_pos_ref[0],self.plt_pos_ref[1],marker='^')\n plt.pause(0.1)\n\n if not self.visual_only:\n self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= True)\n self.graph.update(None, None, use_inactive=True)\n self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= False)\n self.graph.update(None, None, use_inactive=True)\n self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= False)\n self.video.imu_enabled = True\n else:\n self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= True)\n self.graph.update(None, None, use_inactive=True)\n self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= False)\n self.graph.update(None, None, use_inactive=True)\n self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= False)\n self.visual_only_init = True\n\n self.video.set_prior(self.video.last_t0,self.t1)\n\n self.plt_pos = [[],[]]\n self.plt_pos_ref = [[],[]]\n for i in range(0,self.t1):\n TTT = self.video.state.wTbs[i].matrix()\n ppp = TTT[0:3,3]\n qqq = Rotation.from_matrix(TTT[:3, :3]).as_quat()\n self.result_file.writelines('%.6f %.6f %.6f %.6f %.6f %.6f %.6f %.6f\\n'%(self.video.tstamp[i],ppp[0],ppp[1],ppp[2]\\\n ,qqq[0],qqq[1],qqq[2],qqq[3]))\n \n TTTref = np.matmul(self.refTw,TTT) # for visualization\n ppp = TTTref[0:3,3]\n qqq = Rotation.from_matrix(TTTref[:3, :3]).as_quat()\n self.plt_pos[0].append(ppp[0])\n self.plt_pos[1].append(ppp[1])\n if self.all_gt is not None:\n tt_found,dd = self.get_pose_ref(self.video.tstamp[i]-1e-3)\n self.plt_pos_ref[0].append(dd['T'][0,3])\n self.plt_pos_ref[1].append(dd['T'][1,3])\n if self.show_plot:\n plt.subplot(1,3,1)\n plt.cla(); plt.gca().set_title('Trajectory')\n plt.plot(self.plt_pos[0],self.plt_pos[1],marker='^')\n plt.plot(self.plt_pos_ref[0],self.plt_pos_ref[1],marker='^')\n plt.pause(0.1)\n\n for itr in range(1):\n self.graph.update(None, None, use_inactive=True)\n\n def init_GNSS(self):\n \"\"\" initialize the GNSS for geo-referencing fusion \"\"\"\n ten0 = np.array([self.all_gt[self.all_gt_keys[0]]['X0'],\\\n self.all_gt[self.all_gt_keys[0]]['Y0'],\\\n self.all_gt[self.all_gt_keys[0]]['Z0']])\n self.video.ten0 = ten0\n tn0 = []; tw =[]\n for i in range(len(self.video.state.wTbs) - 10,len(self.video.state.wTbs)):\n if self.video.state.gnss_valid[i]:\n # if not is_ref_set:\n # ten0 = self.video.sgraph.gnss_position[i]\n # is_ref_set = True\n teg = self.video.state.gnss_position[i]\n print(self.video.ten0)\n print(self.video.state.gnss_position[i])\n tn0g = np.matmul(trans.Cen(self.video.ten0).T,(self.video.state.gnss_position[i] - self.video.ten0))\n twb = self.video.state.wTbs[i].translation()\n tn0.append(tn0g)\n tw.append(twb)\n if len(tn0) > 1:\n tn0 = np.array(tn0)\n tw = np.array(tw)\n bl = np.linalg.norm(tn0[-1] - tn0[0])\n print('GNSS Alignment Baseline: %.5f' % bl)\n if bl < 10.0:\n print('Baseline too short!!')\n return\n heading_w = math.atan2(tw[-1,1]-tw[0,1],tw[-1,0]-tw[0,0])\n heading_n0 = math.atan2(tn0[-1,1]-tn0[0,1],tn0[-1,0]-tn0[0,0])\n s_w = np.linalg.norm(tw[-1] - tw[0])\n s_n0 = np.linalg.norm(tn0[-1] - tn0[0])\n\n s = s_n0 / s_w\n Rn0w = trans.att2m(np.array([.0,.0,-heading_w + heading_n0]))\n tn0w = tn0 - np.matmul(Rn0w,tw.T * s).T\n\n poses = SE3(self.video.poses)\n wTcs = poses.inv().matrix().cpu().numpy()\n wTbs = np.matmul(wTcs,self.video.Tbc.inverse().matrix())\n wTbs[:,0:3,3] = np.matmul(Rn0w,(wTbs[:,0:3,3]*s).T).T + tn0w[0]\n wTbs[:,0:3,0:3] = np.matmul(Rn0w, (wTbs[:,0:3,0:3]).T).T\n \n self.refTw = np.eye(4,4)\n \n for i in range(0,self.t1):\n self.video.state.wTbs[i] = gtsam.Pose3(wTbs[i])\n self.video.state.vs[i] *= s\n wTcs = np.matmul(wTbs,self.video.Tbc.matrix())\n for i in range(0,self.t1):\n TTT = np.linalg.inv(wTcs[i])\n q = torch.tensor(Rotation.from_matrix(TTT[:3, :3]).as_quat())\n t = torch.tensor(TTT[:3,3])\n self.video.poses[i] = torch.cat([t,q])\n self.video.disps[i] /= s\n\n self.video.gnss_init_t1 = self.t1\n self.video.gnss_init_time = self.video.tstamp[self.t1-1]\n \n self.video.set_prior(self.video.last_t0,self.t1)\n\n self.plt_pos = [[],[]]\n self.plt_pos_ref = [[],[]]\n for i in range(0,self.t1):\n TTT = self.video.state.wTbs[i].matrix()\n ppp = TTT[0:3,3]\n qqq = Rotation.from_matrix(TTT[:3, :3]).as_quat()\n self.result_file.writelines('%.6f %.6f %.6f %.6f %.6f %.6f %.6f %.6f\\n'%(self.video.tstamp[i],ppp[0],ppp[1],ppp[2]\\\n ,qqq[0],qqq[1],qqq[2],qqq[3]))\n \n TTTref = np.matmul(self.refTw,TTT) # for visualization\n ppp = TTTref[0:3,3]\n qqq = Rotation.from_matrix(TTTref[:3, :3]).as_quat()\n self.plt_pos[0].append(ppp[0])\n self.plt_pos[1].append(ppp[1])\n if self.all_gt is not None:\n tt_found,dd = self.get_pose_ref(self.video.tstamp[i]-1e-3)\n self.plt_pos_ref[0].append(dd['T'][0,3])\n self.plt_pos_ref[1].append(dd['T'][1,3])\n if self.show_plot:\n plt.subplot(1,3,1)\n plt.cla(); plt.gca().set_title('Trajectory')\n plt.plot(self.plt_pos[0],self.plt_pos[1],marker='^')\n plt.plot(self.plt_pos_ref[0],self.plt_pos_ref[1],marker='^')\n plt.pause(0.1)\n\n for itr in range(1):\n self.graph.update(None, None, use_inactive=True)\n print('GNSS initialized!!!!')\n\n def VisualIMUAlignment(self, t0, t1, ignore_lever, disable_scale = False):\n poses = SE3(self.video.poses)\n wTcs = poses.inv().matrix().cpu().numpy()\n\n if not ignore_lever:\n wTbs = np.matmul(wTcs,self.video.Tbc.inverse().matrix())\n else:\n T_tmp = self.video.Tbc.inverse().matrix()\n T_tmp[0:3,3] = 0.0\n wTbs = np.matmul(wTcs,T_tmp)\n cost = 0.0\n\n # solveGyroscopeBias\n A = np.zeros([3,3])\n b = np.zeros(3)\n H1 =np.zeros([15,6], order='F', dtype=np.float64)\n H2 =np.zeros([15,3], order='F', dtype=np.float64)\n H3 =np.zeros([15,6], order='F', dtype=np.float64)\n H4 =np.zeros([15,3], order='F', dtype=np.float64)\n H5 =np.zeros([15,6], order='F', dtype=np.float64) # navstate wrt. bias\n H6 =np.zeros([15,6], order='F', dtype=np.float64)\n for i in range(t0,t1-1):\n pose_i = gtsam.Pose3(wTbs[i])\n pose_j = gtsam.Pose3(wTbs[i+1])\n Rij = np.matmul(pose_i.rotation().matrix().T,pose_j.rotation().matrix())\n imu_factor = gtsam.gtsam.CombinedImuFactor(0,1,2,3,4,5,self.video.state.preintegrations[i])\n err = imu_factor.evaluateErrorCustom(pose_i,self.video.state.vs[i],\\\n pose_j,self.video.state.vs[i+1],\\\n self.video.state.bs[i],self.video.state.bs[i+1],\\\n H1,H2,H3,H4,H5,H6)\n tmp_A = H5[0:3,3:6]\n tmp_b = err[0:3]\n cost += np.dot(tmp_b,tmp_b)\n A += np.matmul(tmp_A.T,tmp_A)\n b += np.matmul(tmp_A.T,tmp_b)\n bg = -np.matmul(np.linalg.inv(A),b)\n\n for i in range(0,t1-1):\n pim = gtsam.PreintegratedCombinedMeasurements(self.video.state.params,\\\n gtsam.imuBias.ConstantBias(np.array([.0,.0,.0]),bg))\n for iii in range(len(self.video.state.preintegrations_meas[i])):\n dd = self.video.state.preintegrations_meas[i][iii]\n if dd[2] > 0: pim.integrateMeasurement(dd[0],dd[1],dd[2])\n self.video.state.preintegrations[i] = pim\n self.video.state.bs[i] = gtsam.imuBias.ConstantBias(np.array([.0,.0,.0]),bg)\n print('bg: ',bg)\n \n # linearAlignment\n all_frame_count = t1 - t0\n n_state = all_frame_count * 3 + 3 + 1\n A = np.zeros([n_state,n_state])\n b = np.zeros(n_state)\n i_count = 0\n for i in range(t0,t1-1):\n pose_i = gtsam.Pose3(wTbs[i])\n pose_j = gtsam.Pose3(wTbs[i+1])\n R_i = pose_i.rotation().matrix()\n t_i = pose_i.translation()\n R_j = pose_j.rotation().matrix()\n t_j = pose_j.translation()\n pim = self.video.state.preintegrations[i]\n tic = self.video.Tbc.translation()\n\n tmp_A = np.zeros([6,10])\n tmp_b = np.zeros(6)\n dt = pim.deltaTij()\n tmp_A[0:3,0:3] = -dt * np.eye(3,3)\n tmp_A[0:3,6:9] = R_i.T * dt * dt / 2\n tmp_A[0:3,9] = np.matmul(R_i.T, t_j-t_i) / 100.0\n tmp_b[0:3] = pim.deltaPij()\n tmp_A[3:6,0:3] = -np.eye(3,3)\n tmp_A[3:6,3:6] = np.matmul(R_i.T, R_j)\n tmp_A[3:6,6:9] = R_i.T * dt\n tmp_b[3:6] = pim.deltaVij()\n\n r_A = np.matmul(tmp_A.T,tmp_A)\n r_b = np.matmul(tmp_A.T,tmp_b)\n\n A[i_count*3:i_count*3+6,i_count*3:i_count*3+6] += r_A[0:6,0:6]\n b[i_count*3:i_count*3+6] += r_b[0:6]\n A[-4:,-4:] += r_A[-4:,-4:]\n b[-4:] += r_b[-4:]\n \n A[i_count*3:i_count*3+6,n_state-4:] += r_A[0:6,-4:]\n A[n_state-4:,i_count*3:i_count*3+6] += r_A[-4:,0:6]\n i_count += 1\n \n A = A * 1000.0\n b = b * 1000.0\n x = np.matmul(np.linalg.inv(A),b)\n s = x[n_state-1] / 100.0\n\n g = x[-4:-1]\n\n # RefineGravity\n g0 = g / np.linalg.norm(g) * 9.81\n lx = np.zeros(3)\n ly = np.zeros(3)\n n_state = all_frame_count * 3 + 2 + 1\n A = np.zeros([n_state,n_state])\n b = np.zeros(n_state)\n\n for k in range(4):\n aa = g / np.linalg.norm(g)\n tmp = np.array([.0,.0,1.0])\n\n bb = (tmp - np.dot(aa,tmp) * aa)\n bb /= np.linalg.norm(bb)\n cc = np.cross(aa,bb)\n bc = np.zeros([3,2])\n bc[0:3,0] = bb\n bc[0:3,1] = cc\n lxly = bc\n \n i_count = 0\n for i in range(t0,t1-1):\n pose_i = gtsam.Pose3(wTbs[i])\n pose_j = gtsam.Pose3(wTbs[i+1])\n R_i = pose_i.rotation().matrix()\n t_i = pose_i.translation()\n R_j = pose_j.rotation().matrix()\n t_j = pose_j.translation()\n tmp_A = np.zeros([6,9])\n tmp_b = np.zeros(6)\n pim = self.video.state.preintegrations[i]\n dt = pim.deltaTij()\n\n tmp_A[0:3,0:3] = -dt *np.eye(3,3)\n tmp_A[0:3,6:8] = np.matmul(R_i.T,lxly) * dt * dt /2 \n tmp_A[0:3,8] = np.matmul(R_i.T,t_j - t_i) / 100.0\n tmp_b[0:3] = pim.deltaPij() - np.matmul(R_i.T,g0) * dt * dt / 2\n\n tmp_A[3:6,0:3] = -np.eye(3)\n tmp_A[3:6,3:6] = np.matmul(R_i.T,R_j)\n tmp_A[3:6,6:8] = np.matmul(R_i.T,lxly) * dt\n tmp_b[3:6] = pim.deltaVij() - np.matmul(R_i.T,g0) * dt\n\n r_A = np.matmul(tmp_A.T,tmp_A)\n r_b = np.matmul(tmp_A.T,tmp_b)\n\n A[i_count*3:i_count*3+6,i_count*3:i_count*3+6] += r_A[0:6,0:6]\n b[i_count*3:i_count*3+6] += r_b[0:6]\n A[-3:,-3:] += r_A[-3:,-3:]\n b[-3:] += r_b[-3:]\n\n A[i_count*3:i_count*3+6,n_state-3:] += r_A[0:6,-3:]\n A[n_state-3:,i_count*3:i_count*3+6] += r_A[-3:,0:6]\n i_count += 1\n \n A = A * 1000.0\n b = b * 1000.0\n x = np.matmul(np.linalg.inv(A),b)\n dg = x[-3:-1]\n g0 = g0 + np.matmul(lxly,dg)\n g0 = g0 / np.linalg.norm(g0) * 9.81\n s = x[-1] / 100.0\n print(s,g0,x)\n\n if disable_scale:\n s = 1.0\n \n print('g,s:',g,s)\n if math.fabs(np.linalg.norm(g) - 9.81) < 0.5 and s > 0:\n print('V-I successfully initialized!')\n \n # visualInitialAlign\n wTbs[:,0:3,3] *= s # !!!!!!!!!!!!!!!!!!!!!!!!\n for i in range(0, t1-t0):\n self.video.state.vs[i+t0] = np.matmul(wTbs[i+t0,0:3,0:3],x[i*3:i*3+3])\n \n # g2R\n ng1 = g0/ np.linalg.norm(g0)\n ng2 = np.array([0,0,1.0])\n R0 = trans.FromTwoVectors(ng1,ng2)\n yaw = trans.R2ypr(R0)[0]\n R0 = np.matmul(trans.ypr2R(np.array([-yaw,0,0])),R0)\n\n # align for visualization\n ppp = np.matmul(R0,wTbs[t1-1,0:3,3])\n RRR = np.matmul(R0,wTbs[t1-1,0:3,0:3])\n\n if self.all_gt is not None: # align the initial poses for visualization\n tt_found,dd = self.get_pose_ref(self.video.tstamp[t1-1]-1e-3)\n self.refTw = np.matmul(dd['T'],np.linalg.inv(wTbs[t1-1]))\n self.refTw[0:3,0:3] = trans.att2m([0,0,trans.m2att(self.refTw[0:3,0:3])[2]])\n\n g = np.matmul(R0,g0)\n for i in range(0,t1):\n wTbs[i,0:3,3] = np.matmul(R0,wTbs[i,0:3,3])\n wTbs[i,0:3,0:3] = np.matmul(R0,wTbs[i,0:3,0:3])\n self.video.state.vs[i] = np.matmul(R0, self.video.state.vs[i])\n self.video.state.wTbs[i] = gtsam.Pose3(wTbs[i])\n\n self.video.vi_init_t1 = t1\n self.video.vi_init_time = self.video.tstamp[t1-1]\n\n if not ignore_lever:\n wTcs = np.matmul(wTbs,self.video.Tbc.matrix())\n else:\n T_tmp = self.video.Tbc.matrix()\n T_tmp[0:3,3] = 0.0\n wTcs = np.matmul(wTbs,T_tmp)\n \n for i in range(0,t1):\n TTT = np.linalg.inv(wTcs[i])\n q = torch.tensor(Rotation.from_matrix(TTT[:3, :3]).as_quat())\n t = torch.tensor(TTT[:3,3])\n self.video.poses[i] = torch.cat([t,q])\n self.video.disps[i] /= s\n\n def __initialize(self):\n \"\"\" initialize the SLAM system \"\"\"\n\n self.t0 = 0\n self.t1 = self.video.counter.value\n\n self.graph.add_neighborhood_factors(self.t0, self.t1, r=3)\n\n self.init_IMU()\n\n self.graph.video.imu_enabled = False\n for itr in range(8):\n self.graph.update(1, use_inactive=True)\n\n self.graph.add_proximity_factors(0, 0, rad=2, nms=2, thresh=self.frontend_thresh, remove=False)\n\n for itr in range(8):\n self.graph.update(1, use_inactive=True)\n\n self.graph.video.imu_enabled = False\n for itr in range(8):\n self.graph.update(1, use_inactive=True)\n \n # torch.concat([self.graph.ii[None],self.graph.jj[None]]).T\n # self.video.normalize()\n self.video.poses[self.t1] = self.video.poses[self.t1-1].clone()\n self.video.disps[self.t1] = self.video.disps[self.t1-4:self.t1].mean()\n \n # initialization complete\n self.is_initialized = True\n\n with self.video.get_lock():\n self.video.ready.value = 1\n self.video.dirty[:self.t1] = True\n\n self.graph.rm_factors(self.graph.ii < self.warmup-4, store=True)\n\n def __call__(self):\n \"\"\" main update \"\"\"\n\n # do initialization\n if not self.is_initialized and self.video.counter.value == self.warmup:\n self.__initialize()\n # do update\n elif self.is_initialized and self.t1 < self.video.counter.value:\n self.__update()\n\n \n"
},
{
"path": "dbaf/depth_video.py",
"content": "import numpy as np\nimport torch\nimport lietorch\nimport droid_backends\n\nfrom torch.multiprocessing import Process, Queue, Lock, Value\n\nfrom droid_net import cvx_upsample\nimport geom.projective_ops as pops\n\nfrom multi_sensor import MultiSensorState\nimport gtsam\nfrom gtsam.symbol_shorthand import B, V, X\nfrom scipy.spatial.transform import Rotation\nimport copy\nimport logging\nimport geoFunc.trans as trans\nfrom lietorch import SE3\n\ndef BA2GTSAM(H: np.ndarray, v: np.ndarray, Tbc: gtsam.Pose3):\n A = -Tbc.inverse().AdjointMap()\n # A = -np.eye(6,6)\n A = np.concatenate([A[3:6,:],A[0:3,:]],axis=0)\n ss = H.shape[0]//6\n J = np.zeros_like(H)\n for i in range(ss):\n J[(i*6):(i*6+6),(i*6):(i*6+6)] = A\n JT = J.T\n return np.matmul(np.matmul(JT,H),J),np.matmul(JT,v)\n\ndef CustomHessianFactor(values: gtsam.Values, H: np.ndarray, v: np.ndarray):\n info_expand = np.zeros([H.shape[0]+1,H.shape[1]+1])\n info_expand[0:-1,0:-1] = H\n info_expand[0:-1,-1] = v\n info_expand[-1,-1] = 0.0 # This is meaningless.\n h_f = gtsam.HessianFactor(values.keys(),[6]*len(values.keys()),info_expand)\n l_c = gtsam.LinearContainerFactor(h_f,values)\n return l_c\n\nclass DepthVideo:\n def __init__(self, image_size=[480, 640], buffer=1024, save_pkl = False, stereo=False, upsample = False, device=\"cuda:0\"):\n \n # current keyframe count\n self.counter = Value('i', 0)\n self.ready = Value('i', 0)\n self.ht = ht = image_size[0]\n self.wd = wd = image_size[1]\n\n ### state attributes ###\n self.tstamp = torch.zeros(buffer, device=\"cuda\", dtype=torch.float64).share_memory_()\n self.images = torch.zeros(buffer, 3, ht, wd, device=\"cuda\", dtype=torch.uint8)\n self.dirty = torch.zeros(buffer, device=\"cuda\", dtype=torch.bool).share_memory_()\n self.red = torch.zeros(buffer, device=\"cuda\", dtype=torch.bool).share_memory_()\n self.poses = torch.zeros(buffer, 7, device=\"cuda\", dtype=torch.float).share_memory_()\n self.disps = torch.ones(buffer, ht//8, wd//8, device=\"cuda\", dtype=torch.float).share_memory_()\n self.disps_sens = torch.zeros(buffer, ht//8, wd//8, device=\"cuda\", dtype=torch.float).share_memory_()\n self.disps_up = torch.zeros(buffer, ht, wd, device=\"cuda\", dtype=torch.float).share_memory_()\n self.intrinsics = torch.zeros(buffer, 4, device=\"cuda\", dtype=torch.float).share_memory_()\n\n self.stereo = stereo\n c = 1 if not self.stereo else 2\n\n ### feature attributes ###\n self.fmaps = torch.zeros(buffer, c, 128, ht//8, wd//8, dtype=torch.half, device=\"cuda\").share_memory_()\n self.nets = torch.zeros(buffer, 128, ht//8, wd//8, dtype=torch.half, device=\"cuda\").share_memory_()\n self.inps = torch.zeros(buffer, 128, ht//8, wd//8, dtype=torch.half, device=\"cuda\").share_memory_()\n\n # initialize poses to identity transformation\n self.poses[:] = torch.as_tensor([0, 0, 0, 0, 0, 0, 1], dtype=torch.float, device=\"cuda\")\n \n ### DBAFusion\n # for .pkl saving\n self.disps_save = torch.ones(5000, ht//8, wd//8, device=\"cuda\", dtype=torch.float)\n self.poses_save = torch.ones(5000, 7, device=\"cuda\", dtype=torch.float)\n self.tstamp_save = torch.zeros(5000, device=\"cuda\", dtype=torch.float64)\n self.images_save = torch.zeros(5000, ht//8, wd//8, 3, device=\"cuda\", dtype=torch.float)\n if upsample:\n self.disps_up_save = torch.zeros(5000, ht, wd, device=\"cuda\", dtype=torch.float).share_memory_()\n self.count_save = 0\n self.save_pkl = save_pkl\n self.upsample_flag = upsample\n\n self.state = MultiSensorState()\n self.last_t0 = 0\n self.last_t1 = 0\n self.cur_graph = None\n self.cur_result = None\n self.marg_factor = None\n self.prior_factor = []\n self.prior_factor_map = {}\n self.cur_ii = None\n self.cur_jj = None\n self.cur_target = None\n self.cur_weight = None\n self.cur_eta = None\n\n self.imu_enabled = False\n self.ignore_imu = False\n\n self.xyz_ref = []\n \n # extrinsics, need to be set in the main .py\n self.Ti1c = None # shape = (4,4)\n self.Tbc = None # gtsam.Pose3\n self.tbg = None # shape = (3)\n\n self.reinit = False\n self.vi_init_t1 = -1\n self.vi_init_time = 0.0\n self.gnss_init_t1 = -1\n self.gnss_init_time = 0.0\n self.ten0 = None\n \n self.init_pose_sigma =np.array([0.1, 0.1, 0.0001, 0.0001,0.0001,0.0001])\n self.init_bias_sigma =np.array([1.0,1.0,1.0, 0.1, 0.1, 0.1])\n\n self.logger = logging.getLogger('dba_fusion')\n self.logger.setLevel(logging.DEBUG)\n fh = logging.FileHandler('dba_fusion.log')\n formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')\n fh.setFormatter(formatter)\n # add the handlers to the logger\n self.logger.addHandler(fh)\n self.logger.info('Start logging!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!')\n \n def get_lock(self):\n return self.counter.get_lock()\n\n def __item_setter(self, index, item):\n if isinstance(index, int) and index >= self.counter.value:\n self.counter.value = index + 1\n \n elif isinstance(index, torch.Tensor) and index.max().item() > self.counter.value:\n self.counter.value = index.max().item() + 1\n\n self.tstamp[index] = item[0]\n self.images[index] = item[1]\n\n if item[2] is not None:\n self.poses[index] = item[2]\n\n if item[3] is not None:\n self.disps[index] = item[3]\n\n if item[4] is not None:\n depth = item[4][3::8,3::8]\n self.disps_sens[index] = torch.where(depth>0, 1.0/depth, depth)\n\n if item[5] is not None:\n self.intrinsics[index] = item[5]\n\n if len(item) > 6:\n self.fmaps[index] = item[6]\n\n if len(item) > 7:\n self.nets[index] = item[7]\n\n if len(item) > 8:\n self.inps[index] = item[8]\n\n def __setitem__(self, index, item):\n with self.get_lock():\n self.__item_setter(index, item)\n\n def __getitem__(self, index):\n \"\"\" index the depth video \"\"\"\n\n with self.get_lock():\n # support negative indexing\n if isinstance(index, int) and index < 0:\n index = self.counter.value + index\n\n item = (\n self.poses[index],\n self.disps[index],\n self.intrinsics[index],\n self.fmaps[index],\n self.nets[index],\n self.inps[index])\n\n return item\n\n def append(self, *item):\n with self.get_lock():\n self.__item_setter(self.counter.value, item)\n\n\n ### geometric operations ###\n\n @staticmethod\n def format_indicies(ii, jj):\n \"\"\" to device, long, {-1} \"\"\"\n\n if not isinstance(ii, torch.Tensor):\n ii = torch.as_tensor(ii)\n\n if not isinstance(jj, torch.Tensor):\n jj = torch.as_tensor(jj)\n\n ii = ii.to(device=\"cuda\", dtype=torch.long).reshape(-1)\n jj = jj.to(device=\"cuda\", dtype=torch.long).reshape(-1)\n\n return ii, jj\n\n def upsample(self, ix, mask):\n \"\"\" upsample disparity \"\"\"\n\n disps_up = cvx_upsample(self.disps[ix].unsqueeze(-1), mask)\n self.disps_up[ix] = disps_up.squeeze()\n\n def normalize(self):\n \"\"\" normalize depth and poses \"\"\"\n\n with self.get_lock():\n s = self.disps[:self.counter.value].mean()\n self.disps[:self.counter.value] /= s\n self.poses[:self.counter.value,:3] *= s\n self.dirty[:self.counter.value] = True\n\n\n def reproject(self, ii, jj):\n \"\"\" project points from ii -> jj \"\"\"\n ii, jj = DepthVideo.format_indicies(ii, jj)\n Gs = lietorch.SE3(self.poses[None])\n\n coords, valid_mask = \\\n pops.projective_transform(Gs, self.disps[None], self.intrinsics[None], ii, jj)\n\n return coords, valid_mask\n \n def reproject_comp(self, ii, jj, xyz_comp):\n ii, jj = DepthVideo.format_indicies(ii,jj)\n Gs = lietorch.SE3(self.poses[None])\n\n coords, valid_mask = \\\n pops.projective_transform_comp(Gs, self.disps[None], self.intrinsics[None], ii, jj, xyz_comp)\n\n return coords, valid_mask\n \n def distance(self, ii=None, jj=None, beta=0.3, bidirectional=True):\n \"\"\" frame distance metric \"\"\"\n\n return_matrix = False\n if ii is None:\n return_matrix = True\n N = self.counter.value\n ii, jj = torch.meshgrid(torch.arange(N), torch.arange(N))\n \n ii, jj = DepthVideo.format_indicies(ii, jj)\n\n if bidirectional:\n\n poses = self.poses[:self.counter.value].clone()\n\n d1 = droid_backends.frame_distance(\n poses, self.disps, self.intrinsics[0], ii, jj, beta)\n\n d2 = droid_backends.frame_distance(\n poses, self.disps, self.intrinsics[0], jj, ii, beta)\n\n d = .5 * (d1 + d2)\n\n else:\n d = droid_backends.frame_distance(\n self.poses, self.disps, self.intrinsics[0], ii, jj, beta)\n\n if return_matrix:\n return d.reshape(N, N)\n\n return d\n\n def rm_new_gnss(self, t1):\n if (self.gnss_init_t1> 0 and self.state.gnss_valid[t1]) or self.state.odo_valid[t1]:\n graph_temp = gtsam.NonlinearFactorGraph()\n linear_point = self.marg_factor.linearizationPoint()\n graph_temp.push_back(self.marg_factor)\n\n if self.state.gnss_valid[t1]:\n T1 = self.state.wTbs[t1]\n T0 = self.state.wTbs[t1-1]\n p = np.matmul(trans.Cen(self.ten0).T, self.state.gnss_position[t1] - self.ten0)\n n0pbg = self.state.wTbs[t1].rotation().rotate(self.tbg)\n p = p - n0pbg\n p = p - T1.translation() + T0.translation()\n if not linear_point.exists(X(t1-1)):\n linear_point.insert(X(t1-1), self.cur_result.atPose3(X(t1-1)))\n gnss_factor = gtsam.GPSFactor(X(t1-1), p,\\\n gtsam.noiseModel.Robust.Create(\\\n gtsam.noiseModel.mEstimator.Cauchy(0.08),\\\n gtsam.noiseModel.Diagonal.Sigmas(np.array([1.0,1.0,5.0]))))\n graph_temp.push_back(gnss_factor)\n if self.state.odo_valid[t1]:\n v1 = np.matmul(self.state.wTbs[t1].rotation().matrix().T, self.state.vs[t1])\n v0 = np.matmul(self.state.wTbs[t1-1].rotation().matrix().T, self.state.vs[t1-1])\n v = self.state.odo_vel[t1] - v1 + v0\n if not linear_point.exists(X(t1-1)):\n linear_point.insert(X(t1-1), self.cur_result.atPose3(X(t1-1)))\n if not linear_point.exists(V(t1-1)):\n linear_point.insert(V(t1-1), self.cur_result.atVector(V(t1-1)))\n odo_factor = gtsam.VelFactor(X(t1-1),V(t1-1),v,gtsam.noiseModel.Diagonal.Sigmas(np.array([2.0,2.0,2.0])))\n graph_temp.push_back(odo_factor) \n \n h_factor = graph_temp.linearizeToHessianFactor(linear_point)\n self.marg_factor = gtsam.LinearContainerFactor(h_factor,linear_point)\n \n \n def set_prior(self, t0, t1):\n for i in range(t0,t0+2):\n self.prior_factor_map[i] = []\n init_pose_sigma = self.init_pose_sigma\n if len(self.init_pose_sigma.shape) > 1:\n init_pose_sigma = self.init_pose_sigma[i-t0]\n self.prior_factor_map[i].append(gtsam.PriorFactorPose3(X(i),\\\n self.state.wTbs[i], \\\n gtsam.noiseModel.Diagonal.Sigmas(init_pose_sigma)))\n if not self.ignore_imu:\n self.prior_factor_map[i].append(gtsam.PriorFactorConstantBias(B(i),\\\n self.state.bs[i], \\\n gtsam.noiseModel.Diagonal.Sigmas(self.init_bias_sigma)))\n self.last_t0 = t0\n self.last_t1 = t1\n\n def ba(self, target, weight, eta, ii, jj, t0=1, t1=None, itrs=2, lm=1e-4, ep=0.1, motion_only=False):\n \"\"\" dense bundle adjustment (DBA) \"\"\"\n with self.get_lock():\n if t1 is None:\n t1 = max(ii.max().item(), jj.max().item()) + 1\n\n # 1) visual-only BA\n # 2) multi-sensor BA\n if not self.imu_enabled: \n droid_backends.ba(self.poses, self.disps, self.intrinsics[0], self.disps_sens,\n target, weight, eta, ii, jj, t0, t1, itrs, lm, ep, motion_only)\n for i in range(self.last_t0, min(ii.min().item(), jj.min().item())):\n if self.save_pkl:\n # save marginalized results\n self.tstamp_save[self.count_save] = self.tstamp[i].clone()\n self.disps_save[self.count_save] = self.disps[i].clone()\n self.poses_save[self.count_save] = self.poses[i].clone()\n if self.upsample_flag:\n self.disps_up_save[self.count_save] = self.disps_up[i].clone()\n self.images_save[self.count_save] = self.images[i,[2,1,0],::8,::8].permute(1,2,0) / 255.0 # might be \"3::8, 3::8\"?\n self.count_save += 1\n\n self.last_t0 = min(ii.min().item(), jj.min().item())\n self.last_t1 = t1\n else:\n t0 = min(ii.min().item(), jj.min().item())\n\n \"\"\" marginalization \"\"\"\n if self.last_t1!=t1 or self.last_t0 != t0:\n if self.last_t0 > t0:\n t0 = self.last_t0\n elif self.last_t0 == t0:\n t0 = self.last_t0\n else:\n marg_paras = []\n # Construct a temporary factor graph (related to the old states) to obtain the marginalization information\n graph = gtsam.NonlinearFactorGraph()\n marg_idx = torch.logical_and(torch.greater_equal(self.cur_ii,self.last_t0),\\\n torch.less(self.cur_ii,t0))\n marg_idx2 = torch.logical_and(torch.less(self.cur_ii,self.last_t1-2),\\\n torch.less(self.cur_jj,self.last_t1-2))\n marg_idx = torch.logical_and(marg_idx,marg_idx2)\n\n marg_ii = self.cur_ii[marg_idx]\n marg_jj = self.cur_jj[marg_idx]\n marg_t0 = self.last_t0 \n marg_t1 = t0 + 1\n if len(marg_ii) > 0:\n marg_t0 = self.last_t0 \n marg_t1 = torch.max(marg_jj).item()+1\n marg_result = gtsam.Values()\n for i in range(self.last_t0,marg_t1):\n if i < t0:\n marg_paras.append(X(i))\n if self.save_pkl:\n # save marginalized results\n self.tstamp_save[self.count_save] = self.tstamp[i].clone()\n self.disps_save[self.count_save] = self.disps[i].clone()\n self.poses_save[self.count_save] = self.poses[i].clone()\n if self.upsample_flag:\n self.disps_up_save[self.count_save] = self.disps_up[i].clone()\n self.images_save[self.count_save] = self.images[i,[2,1,0],::8,::8].permute(1,2,0) / 255.0 # might be \"3::8, 3::8\"?\n self.count_save += 1\n marg_result.insert(X(i), self.cur_result.atPose3(X(i)))\n \n marg_target = self.cur_target[marg_idx]\n marg_weight = self.cur_weight[marg_idx]\n marg_eta = self.cur_eta[0:marg_t1-marg_t0]\n \n bacore = droid_backends.BACore()\n bacore.init(self.poses, self.disps, self.intrinsics[0], torch.zeros_like(self.disps_sens),\n marg_target, marg_weight, marg_eta, marg_ii, marg_jj, marg_t0, marg_t1, itrs, lm, ep, motion_only)\n H = torch.zeros([(marg_t1-marg_t0)*6,(marg_t1-marg_t0)*6],dtype=torch.float64,device='cpu')\n v = torch.zeros([(marg_t1-marg_t0)*6],dtype=torch.float64,device='cpu')\n bacore.hessian(H,v)\n \n for i in range(6): H[i,i] += 0.00025 # for stability\n\n # Hg,vg = BA2GTSAM(H,v,self.Tbc)\n Hgg = gtsam.BA2GTSAM(H,v,self.Tbc)\n Hg = Hgg[0:(marg_t1-marg_t0)*6,0:(marg_t1-marg_t0)*6]\n vg = Hgg[0:(marg_t1-marg_t0)*6, (marg_t1-marg_t0)*6]\n vis_factor = CustomHessianFactor(marg_result,Hg,vg)\n \n graph.push_back(vis_factor)\n\n for i in range(self.last_t0,marg_t1):\n if i < t0:\n if X(i) not in marg_paras:\n marg_paras.append(X(i))\n if not self.ignore_imu:\n marg_paras.append(V(i))\n marg_paras.append(B(i))\n graph.push_back(gtsam.gtsam.CombinedImuFactor(\\\n X(i),V(i),X(i+1),V(i+1),B(i),B(i+1),\\\n self.state.preintegrations[i]))\n if self.gnss_init_t1 > 0:\n if self.state.gnss_valid[i]:\n p = np.matmul(trans.Cen(self.ten0).T, self.state.gnss_position[i] - self.ten0)\n n0pbg = self.state.wTbs[i].rotation().rotate(self.tbg)\n p = p - n0pbg\n gnss_factor = gtsam.GPSFactor(X(i), p,\\\n gtsam.noiseModel.Robust.Create(\\\n gtsam.noiseModel.mEstimator.Cauchy(0.08),\\\n gtsam.noiseModel.Diagonal.Sigmas(np.array([1.0,1.0,5.0]))))\n graph.push_back(gnss_factor)\n if self.state.odo_valid[i]:\n vb = self.state.odo_vel[i]\n odo_factor = gtsam.VelFactor(X(i),V(i),vb,gtsam.noiseModel.Diagonal.Sigmas(np.array([2.0,2.0,2.0])))\n graph.push_back(odo_factor)\n \n keys = self.prior_factor_map.keys()\n for i in sorted(keys):\n if i < t0:\n for iii in range(len(self.prior_factor_map[i])):\n graph.push_back(self.prior_factor_map[i][iii])\n del self.prior_factor_map[i]\n if not self.marg_factor == None:\n graph.push_back(self.marg_factor)\n\n self.marg_factor = gtsam.marginalizeOut(graph,self.cur_result,marg_paras)\n\n # covariance inflation of IMU biases\n if self.reinit == True:\n all_keys = self.marg_factor.keys()\n for i in range(len(all_keys)):\n if all_keys[i] == B(t0):\n all_keys[i] = B(0)\n graph = gtsam.NonlinearFactorGraph()\n graph.push_back(self.marg_factor.rekey(all_keys))\n b_l = gtsam.BetweenFactorConstantBias(B(0),B(t0),gtsam.imuBias.ConstantBias(np.array([.0,.0,.0]),np.array([.0,.0,.0])),\\\n gtsam.noiseModel.Diagonal.Sigmas(self.init_bias_sigma))\n graph.push_back(b_l)\n result_tmp = self.marg_factor.linearizationPoint()\n result_tmp.insert(B(0),result_tmp.atConstantBias(B(t0)))\n self.marg_factor = gtsam.marginalizeOut(graph,result_tmp,[B(0)])\n self.reinit = False\n\n self.last_t0 = t0\n self.last_t1 = t1\n\n \"\"\" optimization \"\"\"\n H = torch.zeros([(t1-t0)*6,(t1-t0)*6],dtype=torch.float64,device='cpu')\n v = torch.zeros([(t1-t0)*6],dtype=torch.float64,device='cpu')\n dx = torch.zeros([(t1-t0)*6],dtype=torch.float64,device='cpu') \n\n bacore = droid_backends.BACore()\n active_index = torch.logical_and(ii>=t0,jj>=t0)\n self.cur_ii = ii[active_index]\n self.cur_jj = jj[active_index]\n self.cur_target = target[active_index]\n self.cur_weight = weight[active_index]\n self.cur_eta = eta[(t0-ii.min().item()):]\n\n bacore.init(self.poses, self.disps, self.intrinsics[0], self.disps_sens,\n self.cur_target, self.cur_weight, self.cur_eta, self.cur_ii, self.cur_jj, t0, t1, itrs, lm, ep, motion_only)\n\n self.cur_graph = gtsam.NonlinearFactorGraph()\n params = gtsam.LevenbergMarquardtParams()#;params.setMaxIterations(1)\n\n # imu factor\n if not self.ignore_imu:\n for i in range(t0,t1):\n if i > t0:\n imu_factor = gtsam.gtsam.CombinedImuFactor(\\\n X(i-1),V(i-1),X(i),V(i),B(i-1),B(i),\\\n self.state.preintegrations[i-1])\n self.cur_graph.add(imu_factor)\n\n # prior factor\n keys = self.prior_factor_map.keys()\n for i in sorted(keys):\n if i >= t0 and i < t1:\n for iii in range(len(self.prior_factor_map[i])):\n self.cur_graph.push_back(self.prior_factor_map[i][iii])\n \n # marginalization factor\n if self.marg_factor is not None:\n self.cur_graph.push_back(self.marg_factor)\n\n # GNSS factor\n if self.gnss_init_t1 > 0:\n for i in range(t0,t1):\n if self.state.gnss_valid[i]:\n p = np.matmul(trans.Cen(self.ten0).T, self.state.gnss_position[i] - self.ten0)\n n0pbg = self.state.wTbs[i].rotation().rotate(self.tbg)\n p = p - n0pbg\n gnss_factor = gtsam.GPSFactor(X(i), p,\\\n gtsam.noiseModel.Robust.Create(\\\n gtsam.noiseModel.mEstimator.Cauchy(0.08),\\\n gtsam.noiseModel.Diagonal.Sigmas(np.array([1.0,1.0,5.0]))))\n self.cur_graph.push_back(gnss_factor)\n \n # Odo factor\n for i in range(t0,t1):\n if self.state.odo_valid[i]:\n vb = self.state.odo_vel[i]\n odo_factor = gtsam.VelFactor(X(i),V(i),vb,gtsam.noiseModel.Diagonal.Sigmas(np.array([2.0,2.0,2.0])))\n self.cur_graph.push_back(odo_factor)\n\n \"\"\" multi-sensor DBA iterations \"\"\"\n for iter in range(2):\n if iter > 0:\n self.cur_graph.resize(self.cur_graph.size()-1)\n bacore.hessian(H,v) # camera frame\n Hgg = gtsam.BA2GTSAM(H,v,self.Tbc)\n Hg = Hgg[0:(t1-t0)*6,0:(t1-t0)*6]\n vg = Hgg[0:(t1-t0)*6,(t1-t0)*6]\n\n initial = gtsam.Values()\n for i in range(t0,t1):\n initial.insert(X(i), self.state.wTbs[i]) # the indice need to be handled\n initial_vis = copy.deepcopy(initial)\n vis_factor = CustomHessianFactor(initial_vis,Hg,vg)\n self.cur_graph.push_back(vis_factor)\n \n if not self.ignore_imu:\n for i in range(t0,t1):\n initial.insert(B(i),self.state.bs[i])\n initial.insert(V(i),self.state.vs[i])\n\n optimizer = gtsam.LevenbergMarquardtOptimizer(self.cur_graph, initial, params)\n self.cur_result = optimizer.optimize()\n\n # retraction and depth update\n for i in range(t0,t1):\n p0 = initial.atPose3(X(i))\n p1 = self.cur_result.atPose3(X(i))\n xi = gtsam.Pose3.Logmap(p0.inverse()*p1)\n dx[(i-t0)*6:(i-t0)*6+6] = torch.tensor(xi)\n if not self.ignore_imu:\n self.state.bs[i] = self.cur_result.atConstantBias(B(i))\n self.state.vs[i] = self.cur_result.atVector(V(i))\n self.state.wTbs[i] = self.cur_result.atPose3(X(i))\n dx = torch.tensor(gtsam.GTSAM2BA(dx,self.Tbc))\n dx_dz = bacore.retract(dx)\n del bacore\n self.disps.clamp_(min=0.001)\n"
},
{
"path": "dbaf/droid_net.py",
"content": "import numpy as np\nimport torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nfrom collections import OrderedDict\nfrom modules.extractor import BasicEncoder\nfrom modules.corr import CorrBlock\nfrom modules.gru import ConvGRU\nfrom modules.clipping import GradientClip\nfrom lietorch import SE3\nfrom geom.ba import BA\nimport geom.projective_ops as pops\nfrom geom.graph_utils import graph_to_edge_list, keyframe_indicies\nfrom torch_scatter import scatter_mean\nimport time\n\ndef cvx_upsample(data, mask):\n \"\"\" upsample pixel-wise transformation field \"\"\"\n batch, ht, wd, dim = data.shape\n data = data.permute(0, 3, 1, 2)\n mask = mask.view(batch, 1, 9, 8, 8, ht, wd)\n mask = torch.softmax(mask, dim=2)\n\n up_data = F.unfold(data, [3,3], padding=1)\n up_data = up_data.view(batch, dim, 9, 1, 1, ht, wd)\n\n up_data = torch.sum(mask * up_data, dim=2)\n up_data = up_data.permute(0, 4, 2, 5, 3, 1)\n up_data = up_data.reshape(batch, 8*ht, 8*wd, dim)\n\n return up_data\n\ndef upsample_disp(disp, mask):\n batch, num, ht, wd = disp.shape\n disp = disp.view(batch*num, ht, wd, 1)\n mask = mask.view(batch*num, -1, ht, wd)\n return cvx_upsample(disp, mask).view(batch, num, 8*ht, 8*wd)\n\n\nclass GraphAgg(nn.Module):\n def __init__(self):\n super(GraphAgg, self).__init__()\n self.conv1 = nn.Conv2d(128, 128, 3, padding=1)\n self.conv2 = nn.Conv2d(128, 128, 3, padding=1)\n self.relu = nn.ReLU(inplace=True)\n\n self.eta = nn.Sequential(\n nn.Conv2d(128, 1, 3, padding=1),\n GradientClip(),\n nn.Softplus())\n\n self.upmask = nn.Sequential(\n nn.Conv2d(128, 8*8*9, 1, padding=0))\n\n def forward(self, net, ii):\n batch, num, ch, ht, wd = net.shape\n net = net.view(batch*num, ch, ht, wd)\n\n _, ix = torch.unique(ii, return_inverse=True)\n net = self.relu(self.conv1(net))\n\n net = net.view(batch, num, 128, ht, wd)\n net = scatter_mean(net, ix, dim=1)\n net = net.view(-1, 128, ht, wd)\n\n net = self.relu(self.conv2(net))\n\n eta = self.eta(net).view(batch, -1, ht, wd)\n upmask = self.upmask(net).view(batch, -1, 8*8*9, ht, wd)\n\n return .01 * eta, upmask\n\n\nclass UpdateModule(nn.Module):\n def __init__(self):\n super(UpdateModule, self).__init__()\n cor_planes = 4 * (2*3 + 1)**2\n\n self.corr_encoder = nn.Sequential(\n nn.Conv2d(cor_planes, 128, 1, padding=0),\n nn.ReLU(inplace=True),\n nn.Conv2d(128, 128, 3, padding=1),\n nn.ReLU(inplace=True))\n\n self.flow_encoder = nn.Sequential(\n nn.Conv2d(4, 128, 7, padding=3),\n nn.ReLU(inplace=True),\n nn.Conv2d(128, 64, 3, padding=1),\n nn.ReLU(inplace=True))\n\n self.weight = nn.Sequential(\n nn.Conv2d(128, 128, 3, padding=1),\n nn.ReLU(inplace=True),\n nn.Conv2d(128, 2, 3, padding=1),\n GradientClip(),\n nn.Sigmoid())\n\n self.delta = nn.Sequential(\n nn.Conv2d(128, 128, 3, padding=1),\n nn.ReLU(inplace=True),\n nn.Conv2d(128, 2, 3, padding=1),\n GradientClip())\n\n self.gru = ConvGRU(128, 128+128+64)\n self.agg = GraphAgg()\n\n def forward(self, net, inp, corr, flow=None, ii=None, jj=None, upsample = False):\n \"\"\" RaftSLAM update operator \"\"\"\n batch, num, ch, ht, wd = net.shape\n if flow is None:\n flow = torch.zeros(batch, num, 4, ht, wd, device=net.device)\n\n output_dim = (batch, num, -1, ht, wd)\n net = net.view(batch*num, -1, ht, wd)\n inp = inp.view(batch*num, -1, ht, wd)\n corr = corr.view(batch*num, -1, ht, wd) \n flow = flow.view(batch*num, -1, ht, wd) \n\n corr = self.corr_encoder(corr)\n flow = self.flow_encoder(flow)\n net = self.gru(net, inp, corr, flow)\n\n ### update variables ###\n delta = self.delta(net).view(*output_dim)\n weight = self.weight(net).view(*output_dim)\n\n delta = delta.permute(0,1,3,4,2)[...,:2].contiguous()\n weight = weight.permute(0,1,3,4,2)[...,:2].contiguous()\n\n net = net.view(*output_dim)\n\n if ii is not None:\n ### ATTENTION!!!! ###\n # We found this useless for VIO performance, thus disable it to save computation.\n # Feel free to re-enable it.\n if upsample:\n eta, upmask = self.agg(net, ii.to(net.device))\n return net, delta, weight, eta, upmask\n else:\n return net, delta, weight, None, None\n else:\n return net, delta, weight\n\n\nclass DroidNet(nn.Module):\n def __init__(self):\n super(DroidNet, self).__init__()\n self.fnet = BasicEncoder(output_dim=128, norm_fn='instance')\n self.cnet = BasicEncoder(output_dim=256, norm_fn='none')\n self.update = UpdateModule()\n\n\n def extract_features(self, images):\n \"\"\" run feeature extraction networks \"\"\"\n\n # normalize images\n images = images[:, :, [2,1,0]] / 255.0\n mean = torch.as_tensor([0.485, 0.456, 0.406], device=images.device)\n std = torch.as_tensor([0.229, 0.224, 0.225], device=images.device)\n images = images.sub_(mean[:, None, None]).div_(std[:, None, None])\n\n fmaps = self.fnet(images)\n net = self.cnet(images)\n \n net, inp = net.split([128,128], dim=2)\n net = torch.tanh(net)\n inp = torch.relu(inp)\n return fmaps, net, inp\n\n\n def forward(self, Gs, images, disps, intrinsics, graph=None, num_steps=12, fixedp=2):\n \"\"\" Estimates SE3 or Sim3 between pair of frames \"\"\"\n\n u = keyframe_indicies(graph)\n ii, jj, kk = graph_to_edge_list(graph)\n\n ii = ii.to(device=images.device, dtype=torch.long)\n jj = jj.to(device=images.device, dtype=torch.long)\n\n fmaps, net, inp = self.extract_features(images)\n net, inp = net[:,ii], inp[:,ii]\n corr_fn = CorrBlock(fmaps[:,ii], fmaps[:,jj], num_levels=4, radius=3)\n\n ht, wd = images.shape[-2:]\n coords0 = pops.coords_grid(ht//8, wd//8, device=images.device)\n \n coords1, _ = pops.projective_transform(Gs, disps, intrinsics, ii, jj)\n target = coords1.clone()\n\n Gs_list, disp_list, residual_list = [], [], []\n for step in range(num_steps):\n Gs = Gs.detach()\n disps = disps.detach()\n coords1 = coords1.detach()\n target = target.detach()\n\n # extract motion features\n corr = corr_fn(coords1)\n resd = target - coords1\n flow = coords1 - coords0\n\n motion = torch.cat([flow, resd], dim=-1)\n motion = motion.permute(0,1,4,2,3).clamp(-64.0, 64.0)\n\n net, delta, weight, eta, upmask = \\\n self.update(net, inp, corr, motion, ii, jj)\n\n target = coords1 + delta\n\n for i in range(2):\n Gs, disps = BA(target, weight, eta, Gs, disps, intrinsics, ii, jj, fixedp=2)\n\n coords1, valid_mask = pops.projective_transform(Gs, disps, intrinsics, ii, jj)\n residual = (target - coords1)\n\n Gs_list.append(Gs)\n disp_list.append(upsample_disp(disps, upmask))\n residual_list.append(valid_mask * residual)\n\n\n return Gs_list, disp_list, residual_list\n"
},
{
"path": "dbaf/geoFunc/__init__.py",
"content": ""
},
{
"path": "dbaf/geoFunc/const_value.py",
"content": "import math\r\n\r\npi=math.pi\r\na = 6378137\r\nfinv = 298.257223563"
},
{
"path": "dbaf/geoFunc/trans.py",
"content": "import math\r\nfrom math import atan2, sin, cos\r\nfrom . import const_value\r\nimport numpy as np\r\nfrom scipy.spatial.transform import Rotation\r\n\r\ndef cart2geod(Xinput):\r\n X=Xinput[0]\r\n Y=Xinput[1]\r\n Z=Xinput[2]\r\n\r\n tolsq = 1e-10\r\n maxit = 10\r\n\r\n rtd = 180/const_value.pi\r\n\r\n esq = (2-1/const_value.finv) / const_value.finv\r\n \r\n oneesq = 1-esq\r\n P=math.sqrt(X*X+Y*Y)\r\n\r\n if P > 1e-20:\r\n dlambda = math.atan2(Y,X) *rtd\r\n else:\r\n dlambda = 0\r\n\r\n if dlambda <0:\r\n dlambda = dlambda +360\r\n \r\n r=math.sqrt(P*P+Z*Z)\r\n\r\n if r>1e-20:\r\n sinphi=Z/r\r\n else :\r\n sinphi=0\r\n \r\n dphi = math.asin(sinphi)\r\n\r\n if r<1e-20:\r\n h=0\r\n return\r\n \r\n h = r-const_value.a*(1-sinphi*sinphi/const_value.finv)\r\n\r\n for i in range(maxit):\r\n sinphi = math.sin(dphi)\r\n cosphi = math.cos(dphi)\r\n\r\n N_phi = const_value.a/math.sqrt(1-esq*sinphi*sinphi)\r\n\r\n dP =P -(N_phi+h)*cosphi\r\n dZ=Z-(N_phi*oneesq+h)*sinphi\r\n\r\n h=h+(sinphi*dZ+cosphi*dP)\r\n dphi =dphi+(cosphi*dZ - sinphi*dP)/(N_phi+h)\r\n\r\n if dP*dP + dZ*dZ= 0) & (jj >= 0) & (ii < n) & (jj < m)\n return scatter_sum(A[:,v], ii[v]*m + jj[v], dim=1, dim_size=n*m)\n\ndef safe_scatter_add_vec(b, ii, n):\n v = (ii >= 0) & (ii < n)\n return scatter_sum(b[:,v], ii[v], dim=1, dim_size=n)\n\n# apply retraction operator to inv-depth maps\ndef disp_retr(disps, dz, ii):\n ii = ii.to(device=dz.device)\n return disps + scatter_sum(dz, ii, dim=1, dim_size=disps.shape[1])\n\n# apply retraction operator to poses\ndef pose_retr(poses, dx, ii):\n ii = ii.to(device=dx.device)\n return poses.retr(scatter_sum(dx, ii, dim=1, dim_size=poses.shape[1]))\n\n\ndef BA(target, weight, eta, poses, disps, intrinsics, ii, jj, fixedp=1, rig=1):\n \"\"\" Full Bundle Adjustment \"\"\"\n\n B, P, ht, wd = disps.shape\n N = ii.shape[0]\n D = poses.manifold_dim\n\n ### 1: commpute jacobians and residuals ###\n coords, valid, (Ji, Jj, Jz) = pops.projective_transform(\n poses, disps, intrinsics, ii, jj, jacobian=True)\n\n r = (target - coords).view(B, N, -1, 1)\n w = .001 * (valid * weight).view(B, N, -1, 1)\n\n ### 2: construct linear system ###\n Ji = Ji.reshape(B, N, -1, D)\n Jj = Jj.reshape(B, N, -1, D)\n wJiT = (w * Ji).transpose(2,3)\n wJjT = (w * Jj).transpose(2,3)\n\n Jz = Jz.reshape(B, N, ht*wd, -1)\n\n Hii = torch.matmul(wJiT, Ji) # [B,num,6,6]\n Hij = torch.matmul(wJiT, Jj)\n Hji = torch.matmul(wJjT, Ji)\n Hjj = torch.matmul(wJjT, Jj)\n\n vi = torch.matmul(wJiT, r).squeeze(-1) # [B,num,6]\n vj = torch.matmul(wJjT, r).squeeze(-1)\n\n Ei = (wJiT.view(B,N,D,ht*wd,-1) * Jz[:,:,None]).sum(dim=-1) # [B,num,6,ht*wd]\n Ej = (wJjT.view(B,N,D,ht*wd,-1) * Jz[:,:,None]).sum(dim=-1) # [B,num,6,ht*wd]\n\n w = w.view(B, N, ht*wd, -1)\n r = r.view(B, N, ht*wd, -1)\n wk = torch.sum(w*r*Jz, dim=-1)\n Ck = torch.sum(w*Jz*Jz, dim=-1) # [B,num,ht*wd]\n\n kx, kk = torch.unique(ii, return_inverse=True)\n M = kx.shape[0]\n\n # only optimize keyframe poses\n P = P // rig - fixedp\n ii = ii // rig - fixedp\n jj = jj // rig - fixedp\n\n H = safe_scatter_add_mat(Hii, ii, ii, P, P) + \\\n safe_scatter_add_mat(Hij, ii, jj, P, P) + \\\n safe_scatter_add_mat(Hji, jj, ii, P, P) + \\\n safe_scatter_add_mat(Hjj, jj, jj, P, P)\n\n E = safe_scatter_add_mat(Ei, ii, kk, P, M) + \\\n safe_scatter_add_mat(Ej, jj, kk, P, M)\n\n v = safe_scatter_add_vec(vi, ii, P) + \\\n safe_scatter_add_vec(vj, jj, P)\n\n C = safe_scatter_add_vec(Ck, kk, M)\n w = safe_scatter_add_vec(wk, kk, M)\n\n C = C + eta.view(*C.shape) + 1e-7\n\n H = H.view(B, P, P, D, D)\n E = E.view(B, P, M, D, ht*wd)\n\n ### 3: solve the system ###\n dx, dz = schur_solve(H, E, C, v, w)\n \n ### 4: apply retraction ###\n poses = pose_retr(poses, dx, torch.arange(P) + fixedp)\n disps = disp_retr(disps, dz.view(B,-1,ht,wd), kx)\n\n disps = torch.where(disps > 10, torch.zeros_like(disps), disps)\n disps = disps.clamp(min=0.0)\n\n return poses, disps\n\n\ndef MoBA(target, weight, eta, poses, disps, intrinsics, ii, jj, fixedp=1, rig=1):\n \"\"\" Motion only bundle adjustment \"\"\"\n\n B, P, ht, wd = disps.shape\n N = ii.shape[0]\n D = poses.manifold_dim\n\n ### 1: commpute jacobians and residuals ###\n coords, valid, (Ji, Jj, Jz) = pops.projective_transform(\n poses, disps, intrinsics, ii, jj, jacobian=True)\n\n r = (target - coords).view(B, N, -1, 1)\n w = .001 * (valid * weight).view(B, N, -1, 1)\n\n ### 2: construct linear system ###\n Ji = Ji.reshape(B, N, -1, D)\n Jj = Jj.reshape(B, N, -1, D)\n wJiT = (w * Ji).transpose(2,3)\n wJjT = (w * Jj).transpose(2,3)\n\n Hii = torch.matmul(wJiT, Ji)\n Hij = torch.matmul(wJiT, Jj)\n Hji = torch.matmul(wJjT, Ji)\n Hjj = torch.matmul(wJjT, Jj)\n\n vi = torch.matmul(wJiT, r).squeeze(-1)\n vj = torch.matmul(wJjT, r).squeeze(-1)\n\n # only optimize keyframe poses\n P = P // rig - fixedp\n ii = ii // rig - fixedp\n jj = jj // rig - fixedp\n\n H = safe_scatter_add_mat(Hii, ii, ii, P, P) + \\\n safe_scatter_add_mat(Hij, ii, jj, P, P) + \\\n safe_scatter_add_mat(Hji, jj, ii, P, P) + \\\n safe_scatter_add_mat(Hjj, jj, jj, P, P)\n\n v = safe_scatter_add_vec(vi, ii, P) + \\\n safe_scatter_add_vec(vj, jj, P)\n \n H = H.view(B, P, P, D, D)\n\n ### 3: solve the system ###\n dx = block_solve(H, v)\n\n ### 4: apply retraction ###\n poses = pose_retr(poses, dx, torch.arange(P) + fixedp)\n return poses\n\n"
},
{
"path": "dbaf/geom/chol.py",
"content": "import torch\nimport torch.nn.functional as F\nimport geom.projective_ops as pops\n\nclass CholeskySolver(torch.autograd.Function):\n @staticmethod\n def forward(ctx, H, b):\n # don't crash training if cholesky decomp fails\n try:\n U = torch.linalg.cholesky(H)\n xs = torch.cholesky_solve(b, U)\n ctx.save_for_backward(U, xs)\n ctx.failed = False\n except Exception as e:\n print(e)\n ctx.failed = True\n xs = torch.zeros_like(b)\n\n return xs\n\n @staticmethod\n def backward(ctx, grad_x):\n if ctx.failed:\n return None, None\n\n U, xs = ctx.saved_tensors\n dz = torch.cholesky_solve(grad_x, U)\n dH = -torch.matmul(xs, dz.transpose(-1,-2))\n\n return dH, dz\n\ndef block_solve(H, b, ep=0.1, lm=0.0001):\n \"\"\" solve normal equations \"\"\"\n B, N, _, D, _ = H.shape\n I = torch.eye(D).to(H.device)\n H = H + (ep + lm*H) * I\n\n H = H.permute(0,1,3,2,4)\n H = H.reshape(B, N*D, N*D)\n b = b.reshape(B, N*D, 1)\n\n x = CholeskySolver.apply(H,b)\n return x.reshape(B, N, D)\n\n\ndef schur_solve(H, E, C, v, w, ep=0.1, lm=0.0001, sless=False):\n \"\"\" solve using shur complement \"\"\"\n \n B, P, M, D, HW = E.shape\n H = H.permute(0,1,3,2,4).reshape(B, P*D, P*D)\n E = E.permute(0,1,3,2,4).reshape(B, P*D, M*HW)\n Q = (1.0 / C).view(B, M*HW, 1)\n\n # damping\n I = torch.eye(P*D).to(H.device)\n H = H + (ep + lm*H) * I\n \n v = v.reshape(B, P*D, 1)\n w = w.reshape(B, M*HW, 1)\n\n Et = E.transpose(1,2)\n S = H - torch.matmul(E, Q*Et)\n v = v - torch.matmul(E, Q*w)\n\n dx = CholeskySolver.apply(S, v)\n if sless:\n return dx.reshape(B, P, D)\n\n dz = Q * (w - Et @ dx) \n dx = dx.reshape(B, P, D)\n dz = dz.reshape(B, M, HW)\n\n return dx, dz"
},
{
"path": "dbaf/geom/graph_utils.py",
"content": "\nimport torch\nimport numpy as np\nfrom collections import OrderedDict\n\nimport lietorch\nfrom data_readers.rgbd_utils import compute_distance_matrix_flow, compute_distance_matrix_flow2\n\n\ndef graph_to_edge_list(graph):\n ii, jj, kk = [], [], []\n for s, u in enumerate(graph):\n for v in graph[u]:\n ii.append(u)\n jj.append(v)\n kk.append(s)\n\n ii = torch.as_tensor(ii)\n jj = torch.as_tensor(jj)\n kk = torch.as_tensor(kk)\n return ii, jj, kk\n\ndef keyframe_indicies(graph):\n return torch.as_tensor([u for u in graph])\n\ndef meshgrid(m, n, device='cuda'):\n ii, jj = torch.meshgrid(torch.arange(m), torch.arange(n))\n return ii.reshape(-1).to(device), jj.reshape(-1).to(device)\n\ndef neighbourhood_graph(n, r):\n ii, jj = meshgrid(n, n)\n d = (ii - jj).abs()\n keep = (d >= 1) & (d <= r)\n return ii[keep], jj[keep]\n\n\ndef build_frame_graph(poses, disps, intrinsics, num=16, thresh=24.0, r=2):\n \"\"\" construct a frame graph between co-visible frames \"\"\"\n N = poses.shape[1]\n poses = poses[0].cpu().numpy()\n disps = disps[0][:,3::8,3::8].cpu().numpy()\n intrinsics = intrinsics[0].cpu().numpy() / 8.0\n d = compute_distance_matrix_flow(poses, disps, intrinsics)\n\n count = 0\n graph = OrderedDict()\n \n for i in range(N):\n graph[i] = []\n d[i,i] = np.inf\n for j in range(i-r, i+r+1):\n if 0 <= j < N and i != j:\n graph[i].append(j)\n d[i,j] = np.inf\n count += 1\n\n while count < num:\n ix = np.argmin(d)\n i, j = ix // N, ix % N\n\n if d[i,j] < thresh:\n graph[i].append(j)\n d[i,j] = np.inf\n count += 1\n else:\n break\n \n return graph\n\n\n\ndef build_frame_graph_v2(poses, disps, intrinsics, num=16, thresh=24.0, r=2):\n \"\"\" construct a frame graph between co-visible frames \"\"\"\n N = poses.shape[1]\n # poses = poses[0].cpu().numpy()\n # disps = disps[0].cpu().numpy()\n # intrinsics = intrinsics[0].cpu().numpy()\n d = compute_distance_matrix_flow2(poses, disps, intrinsics)\n\n # import matplotlib.pyplot as plt\n # plt.imshow(d)\n # plt.show()\n\n count = 0\n graph = OrderedDict()\n \n for i in range(N):\n graph[i] = []\n d[i,i] = np.inf\n for j in range(i-r, i+r+1):\n if 0 <= j < N and i != j:\n graph[i].append(j)\n d[i,j] = np.inf\n count += 1\n\n while 1:\n ix = np.argmin(d)\n i, j = ix // N, ix % N\n\n if d[i,j] < thresh:\n graph[i].append(j)\n\n for i1 in range(i-1, i+2):\n for j1 in range(j-1, j+2):\n if 0 <= i1 < N and 0 <= j1 < N:\n d[i1, j1] = np.inf\n\n count += 1\n else:\n break\n \n return graph\n\n"
},
{
"path": "dbaf/geom/losses.py",
"content": "from collections import OrderedDict\nimport numpy as np\nimport torch\nfrom lietorch import SO3, SE3, Sim3\nfrom .graph_utils import graph_to_edge_list\nfrom .projective_ops import projective_transform\n\n\ndef pose_metrics(dE):\n \"\"\" Translation/Rotation/Scaling metrics from Sim3 \"\"\"\n t, q, s = dE.data.split([3, 4, 1], -1)\n ang = SO3(q).log().norm(dim=-1)\n\n # convert radians to degrees\n r_err = (180 / np.pi) * ang\n t_err = t.norm(dim=-1)\n s_err = (s - 1.0).abs()\n return r_err, t_err, s_err\n\n\ndef fit_scale(Ps, Gs):\n b = Ps.shape[0]\n t1 = Ps.data[...,:3].detach().reshape(b, -1)\n t2 = Gs.data[...,:3].detach().reshape(b, -1)\n\n s = (t1*t2).sum(-1) / ((t2*t2).sum(-1) + 1e-8)\n return s\n\n\ndef geodesic_loss(Ps, Gs, graph, gamma=0.9, do_scale=True):\n \"\"\" Loss function for training network \"\"\"\n\n # relative pose\n ii, jj, kk = graph_to_edge_list(graph)\n dP = Ps[:,jj] * Ps[:,ii].inv()\n\n n = len(Gs)\n geodesic_loss = 0.0\n\n for i in range(n):\n w = gamma ** (n - i - 1)\n dG = Gs[i][:,jj] * Gs[i][:,ii].inv()\n\n if do_scale:\n s = fit_scale(dP, dG)\n dG = dG.scale(s[:,None])\n \n # pose error\n d = (dG * dP.inv()).log()\n\n if isinstance(dG, SE3):\n tau, phi = d.split([3,3], dim=-1)\n geodesic_loss += w * (\n tau.norm(dim=-1).mean() + \n phi.norm(dim=-1).mean())\n\n elif isinstance(dG, Sim3):\n tau, phi, sig = d.split([3,3,1], dim=-1)\n geodesic_loss += w * (\n tau.norm(dim=-1).mean() + \n phi.norm(dim=-1).mean() + \n 0.05 * sig.norm(dim=-1).mean())\n \n dE = Sim3(dG * dP.inv()).detach()\n r_err, t_err, s_err = pose_metrics(dE)\n\n metrics = {\n 'rot_error': r_err.mean().item(),\n 'tr_error': t_err.mean().item(),\n 'bad_rot': (r_err < .1).float().mean().item(),\n 'bad_tr': (t_err < .01).float().mean().item(),\n }\n\n return geodesic_loss, metrics\n\n\ndef residual_loss(residuals, gamma=0.9):\n \"\"\" loss on system residuals \"\"\"\n residual_loss = 0.0\n n = len(residuals)\n\n for i in range(n):\n w = gamma ** (n - i - 1)\n residual_loss += w * residuals[i].abs().mean()\n\n return residual_loss, {'residual': residual_loss.item()}\n\n\ndef flow_loss(Ps, disps, poses_est, disps_est, intrinsics, graph, gamma=0.9):\n \"\"\" optical flow loss \"\"\"\n\n N = Ps.shape[1]\n graph = OrderedDict()\n for i in range(N):\n graph[i] = [j for j in range(N) if abs(i-j)==1]\n\n ii, jj, kk = graph_to_edge_list(graph)\n coords0, val0 = projective_transform(Ps, disps, intrinsics, ii, jj)\n val0 = val0 * (disps[:,ii] > 0).float().unsqueeze(dim=-1)\n\n n = len(poses_est)\n flow_loss = 0.0\n\n for i in range(n):\n w = gamma ** (n - i - 1)\n coords1, val1 = projective_transform(poses_est[i], disps_est[i], intrinsics, ii, jj)\n\n v = (val0 * val1).squeeze(dim=-1)\n epe = v * (coords1 - coords0).norm(dim=-1)\n flow_loss += w * epe.mean()\n\n epe = epe.reshape(-1)[v.reshape(-1) > 0.5]\n metrics = {\n 'f_error': epe.mean().item(),\n '1px': (epe<1.0).float().mean().item(),\n }\n\n return flow_loss, metrics\n"
},
{
"path": "dbaf/geom/projective_ops.py",
"content": "import torch\nimport torch.nn.functional as F\n\nfrom lietorch import SE3, Sim3\n\nMIN_DEPTH = 0.2\n\ndef extract_intrinsics(intrinsics):\n return intrinsics[...,None,None,:].unbind(dim=-1)\n\ndef coords_grid(ht, wd, **kwargs):\n y, x = torch.meshgrid(\n torch.arange(ht).to(**kwargs).float(),\n torch.arange(wd).to(**kwargs).float())\n\n return torch.stack([x, y], dim=-1)\n\ndef iproj(disps, intrinsics, jacobian=False):\n \"\"\" pinhole camera inverse projection \"\"\"\n ht, wd = disps.shape[2:]\n fx, fy, cx, cy = extract_intrinsics(intrinsics)\n \n y, x = torch.meshgrid(\n torch.arange(ht).to(disps.device).float(),\n torch.arange(wd).to(disps.device).float())\n\n i = torch.ones_like(disps)\n X = (x - cx) / fx\n Y = (y - cy) / fy\n pts = torch.stack([X, Y, i, disps], dim=-1)\n\n if jacobian:\n J = torch.zeros_like(pts)\n J[...,-1] = 1.0\n return pts, J\n\n return pts, None\n\ndef proj(Xs, intrinsics, jacobian=False, return_depth=False):\n \"\"\" pinhole camera projection \"\"\"\n fx, fy, cx, cy = extract_intrinsics(intrinsics)\n X, Y, Z, D = Xs.unbind(dim=-1)\n\n Z = torch.where(Z < 0.5*MIN_DEPTH, torch.ones_like(Z), Z)\n d = 1.0 / Z\n\n x = fx * (X * d) + cx\n y = fy * (Y * d) + cy\n if return_depth:\n coords = torch.stack([x, y, D*d], dim=-1)\n else:\n coords = torch.stack([x, y], dim=-1)\n\n if jacobian:\n B, N, H, W = d.shape\n o = torch.zeros_like(d)\n proj_jac = torch.stack([\n fx*d, o, -fx*X*d*d, o,\n o, fy*d, -fy*Y*d*d, o,\n # o, o, -D*d*d, d,\n ], dim=-1).view(B, N, H, W, 2, 4)\n\n return coords, proj_jac\n\n return coords, None\n\ndef actp(Gij, X0, jacobian=False):\n \"\"\" action on point cloud \"\"\"\n X1 = Gij[:,:,None,None] * X0\n \n if jacobian:\n X, Y, Z, d = X1.unbind(dim=-1)\n o = torch.zeros_like(d)\n B, N, H, W = d.shape\n\n if isinstance(Gij, SE3):\n Ja = torch.stack([\n d, o, o, o, Z, -Y,\n o, d, o, -Z, o, X, \n o, o, d, Y, -X, o,\n o, o, o, o, o, o,\n ], dim=-1).view(B, N, H, W, 4, 6)\n\n elif isinstance(Gij, Sim3):\n Ja = torch.stack([\n d, o, o, o, Z, -Y, X,\n o, d, o, -Z, o, X, Y,\n o, o, d, Y, -X, o, Z,\n o, o, o, o, o, o, o\n ], dim=-1).view(B, N, H, W, 4, 7)\n\n return X1, Ja\n\n return X1, None\n\ndef projective_transform(poses, depths, intrinsics, ii, jj, jacobian=False, return_depth=False):\n \"\"\" map points from ii->jj \"\"\"\n\n # inverse project (pinhole)\n X0, Jz = iproj(depths[:,ii], intrinsics[:,ii], jacobian=jacobian)\n \n # transform\n Gij = poses[:,jj] * poses[:,ii].inv()\n\n Gij.data[:,ii==jj] = torch.as_tensor([-0.1, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], device=\"cuda\")\n X1, Ja = actp(Gij, X0, jacobian=jacobian) # 4*6\n \n # project (pinhole)\n x1, Jp = proj(X1, intrinsics[:,jj], jacobian=jacobian, return_depth=return_depth) # 2*4\n\n # exclude points too close to camera\n valid = ((X1[...,2] > MIN_DEPTH) & (X0[...,2] > MIN_DEPTH)).float()\n valid = valid.unsqueeze(-1)\n\n if jacobian:\n # Ji transforms according to dual adjoint\n Jj = torch.matmul(Jp, Ja) # 2*6\n Ji = -Gij[:,:,None,None,None].adjT(Jj) # 2*6 * 6*6\n\n Jz = Gij[:,:,None,None] * Jz\n Jz = torch.matmul(Jp, Jz.unsqueeze(-1))\n\n return x1, valid, (Ji, Jj, Jz)\n\n return x1, valid\n\ndef projective_transform_comp(poses, depths, intrinsics, ii, jj, xyz_comp, jacobian=False, return_depth=False):\n \"\"\" map points from ii->jj \"\"\"\n\n # inverse project (pinhole)\n X0, Jz = iproj(depths[:,ii], intrinsics[:,ii], jacobian=jacobian)\n \n # transform\n Gij = poses[:,jj] * poses[:,ii].inv()\n\n Gij.data[:,ii==jj] = torch.as_tensor([-0.1, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], device=\"cuda\")\n X1, Ja = actp(Gij, X0, jacobian=jacobian)\n\n X1 = X1 + xyz_comp # compensate the object motion\n \n # project (pinhole)\n x1, Jp = proj(X1, intrinsics[:,jj], jacobian=jacobian, return_depth=return_depth)\n\n # exclude points too close to camera\n valid = ((X1[...,2] > MIN_DEPTH) & (X0[...,2] > MIN_DEPTH)).float()\n valid = valid.unsqueeze(-1)\n\n if jacobian:\n # Ji transforms according to dual adjoint\n Jj = torch.matmul(Jp, Ja)\n Ji = -Gij[:,:,None,None,None].adjT(Jj)\n\n Jz = Gij[:,:,None,None] * Jz\n Jz = torch.matmul(Jp, Jz.unsqueeze(-1))\n\n return x1, valid, (Ji, Jj, Jz)\n\n return x1, valid\n\ndef induced_flow(poses, disps, intrinsics, ii, jj):\n \"\"\" optical flow induced by camera motion \"\"\"\n\n ht, wd = disps.shape[2:]\n y, x = torch.meshgrid(\n torch.arange(ht).to(disps.device).float(),\n torch.arange(wd).to(disps.device).float())\n\n coords0 = torch.stack([x, y], dim=-1)\n coords1, valid = projective_transform(poses, disps, intrinsics, ii, jj, False)\n\n return coords1[...,:2] - coords0, valid\n\n"
},
{
"path": "dbaf/modules/__init__.py",
"content": ""
},
{
"path": "dbaf/modules/clipping.py",
"content": "import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nGRAD_CLIP = .01\n\nclass GradClip(torch.autograd.Function):\n @staticmethod\n def forward(ctx, x):\n return x\n\n @staticmethod\n def backward(ctx, grad_x):\n o = torch.zeros_like(grad_x)\n grad_x = torch.where(grad_x.abs()>GRAD_CLIP, o, grad_x)\n grad_x = torch.where(torch.isnan(grad_x), o, grad_x)\n return grad_x\n\nclass GradientClip(nn.Module):\n def __init__(self):\n super(GradientClip, self).__init__()\n\n def forward(self, x):\n return GradClip.apply(x)"
},
{
"path": "dbaf/modules/corr.py",
"content": "import torch\nimport torch.nn.functional as F\n\nimport droid_backends\n\nclass CorrSampler(torch.autograd.Function):\n\n @staticmethod\n def forward(ctx, volume, coords, radius):\n ctx.save_for_backward(volume,coords)\n ctx.radius = radius\n corr, = droid_backends.corr_index_forward(volume, coords, radius)\n return corr\n\n @staticmethod\n def backward(ctx, grad_output):\n volume, coords = ctx.saved_tensors\n grad_output = grad_output.contiguous()\n grad_volume, = droid_backends.corr_index_backward(volume, coords, grad_output, ctx.radius)\n return grad_volume, None, None\n\n\nclass CorrBlock:\n def __init__(self, fmap1, fmap2, num_levels=4, radius=3):\n self.num_levels = num_levels\n self.radius = radius\n self.corr_pyramid = []\n\n # all pairs correlation\n corr = CorrBlock.corr(fmap1, fmap2)\n\n batch, num, h1, w1, h2, w2 = corr.shape\n corr = corr.reshape(batch*num*h1*w1, 1, h2, w2)\n \n for i in range(self.num_levels):\n self.corr_pyramid.append(\n corr.view(batch*num, h1, w1, h2//2**i, w2//2**i))\n corr = F.avg_pool2d(corr, 2, stride=2)\n \n def __call__(self, coords):\n out_pyramid = []\n batch, num, ht, wd, _ = coords.shape\n coords = coords.permute(0,1,4,2,3)\n coords = coords.contiguous().view(batch*num, 2, ht, wd)\n \n for i in range(self.num_levels):\n corr = CorrSampler.apply(self.corr_pyramid[i], coords/2**i, self.radius) # 我的理解，每一个点与其他每个点都有一个(2*3+1)^2=49长度的特征\n out_pyramid.append(corr.view(batch, num, -1, ht, wd))\n\n return torch.cat(out_pyramid, dim=2)\n\n def cat(self, other):\n for i in range(self.num_levels):\n self.corr_pyramid[i] = torch.cat([self.corr_pyramid[i], other.corr_pyramid[i]], 0)\n return self\n\n def __getitem__(self, index):\n for i in range(self.num_levels):\n self.corr_pyramid[i] = self.corr_pyramid[i][index]\n return self\n\n\n @staticmethod\n def corr(fmap1, fmap2):\n \"\"\" all-pairs correlation \"\"\"\n batch, num, dim, ht, wd = fmap1.shape\n fmap1 = fmap1.reshape(batch*num, dim, ht*wd) / 4.0\n fmap2 = fmap2.reshape(batch*num, dim, ht*wd) / 4.0\n \n corr = torch.matmul(fmap1.transpose(1,2), fmap2)\n return corr.view(batch, num, ht, wd, ht, wd)\n\n\nclass CorrLayer(torch.autograd.Function):\n @staticmethod\n def forward(ctx, fmap1, fmap2, coords, r):\n ctx.r = r\n ctx.save_for_backward(fmap1, fmap2, coords)\n corr, = droid_backends.altcorr_forward(fmap1, fmap2, coords, ctx.r)\n return corr\n\n @staticmethod\n def backward(ctx, grad_corr):\n fmap1, fmap2, coords = ctx.saved_tensors\n grad_corr = grad_corr.contiguous()\n fmap1_grad, fmap2_grad, coords_grad = \\\n droid_backends.altcorr_backward(fmap1, fmap2, coords, grad_corr, ctx.r)\n return fmap1_grad, fmap2_grad, coords_grad, None\n\n\nclass AltCorrBlock:\n def __init__(self, fmaps, num_levels=4, radius=3):\n self.num_levels = num_levels\n self.radius = radius\n\n B, N, C, H, W = fmaps.shape\n fmaps = fmaps.view(B*N, C, H, W) / 4.0\n \n self.pyramid = []\n for i in range(self.num_levels):\n sz = (B, N, H//2**i, W//2**i, C)\n fmap_lvl = fmaps.permute(0, 2, 3, 1).contiguous()\n self.pyramid.append(fmap_lvl.view(*sz))\n fmaps = F.avg_pool2d(fmaps, 2, stride=2)\n \n def corr_fn(self, coords, ii, jj):\n B, N, H, W, S, _ = coords.shape\n coords = coords.permute(0, 1, 4, 2, 3, 5)\n\n corr_list = []\n for i in range(self.num_levels):\n r = self.radius\n fmap1_i = self.pyramid[0][:, ii]\n fmap2_i = self.pyramid[i][:, jj]\n\n coords_i = (coords / 2**i).reshape(B*N, S, H, W, 2).contiguous()\n fmap1_i = fmap1_i.reshape((B*N,) + fmap1_i.shape[2:])\n fmap2_i = fmap2_i.reshape((B*N,) + fmap2_i.shape[2:])\n\n corr = CorrLayer.apply(fmap1_i.float(), fmap2_i.float(), coords_i, self.radius)\n corr = corr.view(B, N, S, -1, H, W).permute(0, 1, 3, 4, 5, 2)\n corr_list.append(corr)\n\n corr = torch.cat(corr_list, dim=2)\n return corr\n\n\n def __call__(self, coords, ii, jj):\n squeeze_output = False\n if len(coords.shape) == 5:\n coords = coords.unsqueeze(dim=-2)\n squeeze_output = True\n\n corr = self.corr_fn(coords, ii, jj)\n \n if squeeze_output:\n corr = corr.squeeze(dim=-1)\n\n return corr.contiguous()\n\n"
},
{
"path": "dbaf/modules/extractor.py",
"content": "import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\n\nclass ResidualBlock(nn.Module):\n def __init__(self, in_planes, planes, norm_fn='group', stride=1):\n super(ResidualBlock, self).__init__()\n \n self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, padding=1, stride=stride)\n self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1)\n self.relu = nn.ReLU(inplace=True)\n\n num_groups = planes // 8\n\n if norm_fn == 'group':\n self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)\n self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)\n if not stride == 1:\n self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)\n \n elif norm_fn == 'batch':\n self.norm1 = nn.BatchNorm2d(planes)\n self.norm2 = nn.BatchNorm2d(planes)\n if not stride == 1:\n self.norm3 = nn.BatchNorm2d(planes)\n \n elif norm_fn == 'instance':\n self.norm1 = nn.InstanceNorm2d(planes)\n self.norm2 = nn.InstanceNorm2d(planes)\n if not stride == 1:\n self.norm3 = nn.InstanceNorm2d(planes)\n\n elif norm_fn == 'none':\n self.norm1 = nn.Sequential()\n self.norm2 = nn.Sequential()\n if not stride == 1:\n self.norm3 = nn.Sequential()\n\n if stride == 1:\n self.downsample = None\n \n else: \n self.downsample = nn.Sequential(\n nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm3)\n\n def forward(self, x):\n y = x\n y = self.relu(self.norm1(self.conv1(y)))\n y = self.relu(self.norm2(self.conv2(y)))\n\n if self.downsample is not None:\n x = self.downsample(x)\n\n return self.relu(x+y)\n\n\nclass BottleneckBlock(nn.Module):\n def __init__(self, in_planes, planes, norm_fn='group', stride=1):\n super(BottleneckBlock, self).__init__()\n \n self.conv1 = nn.Conv2d(in_planes, planes//4, kernel_size=1, padding=0)\n self.conv2 = nn.Conv2d(planes//4, planes//4, kernel_size=3, padding=1, stride=stride)\n self.conv3 = nn.Conv2d(planes//4, planes, kernel_size=1, padding=0)\n self.relu = nn.ReLU(inplace=True)\n\n num_groups = planes // 8\n\n if norm_fn == 'group':\n self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes//4)\n self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes//4)\n self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)\n if not stride == 1:\n self.norm4 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)\n \n elif norm_fn == 'batch':\n self.norm1 = nn.BatchNorm2d(planes//4)\n self.norm2 = nn.BatchNorm2d(planes//4)\n self.norm3 = nn.BatchNorm2d(planes)\n if not stride == 1:\n self.norm4 = nn.BatchNorm2d(planes)\n \n elif norm_fn == 'instance':\n self.norm1 = nn.InstanceNorm2d(planes//4)\n self.norm2 = nn.InstanceNorm2d(planes//4)\n self.norm3 = nn.InstanceNorm2d(planes)\n if not stride == 1:\n self.norm4 = nn.InstanceNorm2d(planes)\n\n elif norm_fn == 'none':\n self.norm1 = nn.Sequential()\n self.norm2 = nn.Sequential()\n self.norm3 = nn.Sequential()\n if not stride == 1:\n self.norm4 = nn.Sequential()\n\n if stride == 1:\n self.downsample = None\n \n else: \n self.downsample = nn.Sequential(\n nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm4)\n\n def forward(self, x):\n y = x\n y = self.relu(self.norm1(self.conv1(y)))\n y = self.relu(self.norm2(self.conv2(y)))\n y = self.relu(self.norm3(self.conv3(y)))\n\n if self.downsample is not None:\n x = self.downsample(x)\n\n return self.relu(x+y)\n\n\nDIM=32\n\nclass BasicEncoder(nn.Module):\n def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0, multidim=False):\n super(BasicEncoder, self).__init__()\n self.norm_fn = norm_fn\n self.multidim = multidim\n\n if self.norm_fn == 'group':\n self.norm1 = nn.GroupNorm(num_groups=8, num_channels=DIM)\n \n elif self.norm_fn == 'batch':\n self.norm1 = nn.BatchNorm2d(DIM)\n\n elif self.norm_fn == 'instance':\n self.norm1 = nn.InstanceNorm2d(DIM)\n\n elif self.norm_fn == 'none':\n self.norm1 = nn.Sequential()\n\n self.conv1 = nn.Conv2d(3, DIM, kernel_size=7, stride=2, padding=3)\n self.relu1 = nn.ReLU(inplace=True)\n\n self.in_planes = DIM\n self.layer1 = self._make_layer(DIM, stride=1)\n self.layer2 = self._make_layer(2*DIM, stride=2)\n self.layer3 = self._make_layer(4*DIM, stride=2)\n\n # output convolution\n self.conv2 = nn.Conv2d(4*DIM, output_dim, kernel_size=1)\n\n if self.multidim:\n self.layer4 = self._make_layer(256, stride=2)\n self.layer5 = self._make_layer(512, stride=2)\n\n self.in_planes = 256\n self.layer6 = self._make_layer(256, stride=1)\n\n self.in_planes = 128\n self.layer7 = self._make_layer(128, stride=1)\n\n self.up1 = nn.Conv2d(512, 256, 1)\n self.up2 = nn.Conv2d(256, 128, 1)\n self.conv3 = nn.Conv2d(128, output_dim, kernel_size=1)\n\n if dropout > 0:\n self.dropout = nn.Dropout2d(p=dropout)\n else:\n self.dropout = None\n\n for m in self.modules():\n if isinstance(m, nn.Conv2d):\n nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')\n elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)):\n if m.weight is not None:\n nn.init.constant_(m.weight, 1)\n if m.bias is not None:\n nn.init.constant_(m.bias, 0)\n\n def _make_layer(self, dim, stride=1):\n layer1 = ResidualBlock(self.in_planes, dim, self.norm_fn, stride=stride)\n layer2 = ResidualBlock(dim, dim, self.norm_fn, stride=1)\n layers = (layer1, layer2)\n \n self.in_planes = dim\n return nn.Sequential(*layers)\n\n def forward(self, x):\n b, n, c1, h1, w1 = x.shape\n x = x.view(b*n, c1, h1, w1)\n\n x = self.conv1(x)\n x = self.norm1(x)\n x = self.relu1(x)\n\n x = self.layer1(x)\n x = self.layer2(x)\n x = self.layer3(x)\n\n x = self.conv2(x)\n\n _, c2, h2, w2 = x.shape\n return x.view(b, n, c2, h2, w2)\n"
},
{
"path": "dbaf/modules/gru.py",
"content": "import torch\nimport torch.nn as nn\n\n\nclass ConvGRU(nn.Module):\n def __init__(self, h_planes=128, i_planes=128):\n super(ConvGRU, self).__init__()\n self.do_checkpoint = False\n self.convz = nn.Conv2d(h_planes+i_planes, h_planes, 3, padding=1)\n self.convr = nn.Conv2d(h_planes+i_planes, h_planes, 3, padding=1)\n self.convq = nn.Conv2d(h_planes+i_planes, h_planes, 3, padding=1)\n\n self.w = nn.Conv2d(h_planes, h_planes, 1, padding=0)\n\n self.convz_glo = nn.Conv2d(h_planes, h_planes, 1, padding=0)\n self.convr_glo = nn.Conv2d(h_planes, h_planes, 1, padding=0)\n self.convq_glo = nn.Conv2d(h_planes, h_planes, 1, padding=0)\n\n def forward(self, net, *inputs):\n inp = torch.cat(inputs, dim=1)\n net_inp = torch.cat([net, inp], dim=1) #[1,128+128+128+64,H//8,W//8]\n\n b, c, h, w = net.shape\n glo = torch.sigmoid(self.w(net)) * net\n glo = glo.view(b, c, h*w).mean(-1).view(b, c, 1, 1) # global context\n\n z = torch.sigmoid(self.convz(net_inp) + self.convz_glo(glo))\n r = torch.sigmoid(self.convr(net_inp) + self.convr_glo(glo))\n q = torch.tanh(self.convq(torch.cat([r*net, inp], dim=1)) + self.convq_glo(glo))\n\n net = (1-z) * net + z * q\n return net\n\n\n"
},
{
"path": "dbaf/motion_filter.py",
"content": "import cv2\nimport torch\nimport lietorch\n\nfrom collections import OrderedDict\nfrom droid_net import DroidNet\n\nimport geom.projective_ops as pops\nfrom modules.corr import CorrBlock\nimport numpy as np\n\nclass MotionFilter:\n \"\"\" This class is used to filter incoming frames and extract features \"\"\"\n\n def __init__(self, net, video, thresh=2.5, device=\"cuda:0\"):\n \n # split net modules\n self.cnet = net.cnet\n self.fnet = net.fnet\n self.update = net.update\n\n self.video = video\n self.thresh = thresh\n self.device = device\n\n self.count = 0\n\n # mean, std for image normalization\n self.MEAN = torch.as_tensor([0.485, 0.456, 0.406], device=self.device)[:, None, None]\n self.STDV = torch.as_tensor([0.229, 0.224, 0.225], device=self.device)[:, None, None]\n \n @torch.cuda.amp.autocast(enabled=True)\n def __context_encoder(self, image):\n \"\"\" context features \"\"\"\n net, inp = self.cnet(image).split([128,128], dim=2)\n return net.tanh().squeeze(0), inp.relu().squeeze(0)\n \n @torch.cuda.amp.autocast(enabled=True)\n def context_encoder(self, image):\n \"\"\" context features \"\"\"\n net, inp = self.cnet(image).split([128,128], dim=2)\n return net.tanh().squeeze(0), inp.relu().squeeze(0)\n \n @torch.cuda.amp.autocast(enabled=True)\n def __feature_encoder(self, image):\n \"\"\" features for correlation volume \"\"\"\n return self.fnet(image).squeeze(0)\n\n @torch.cuda.amp.autocast(enabled=True)\n def feature_encoder(self, image):\n \"\"\" features for correlation volume \"\"\"\n return self.fnet(image).squeeze(0)\n \n @torch.cuda.amp.autocast(enabled=True)\n @torch.no_grad()\n def track(self, tstamp, image, depth=None, intrinsics=None):\n \"\"\" main update operation - run on every frame in video \"\"\"\n\n Id = lietorch.SE3.Identity(1,).data.squeeze()\n ht = image.shape[-2] // 8\n wd = image.shape[-1] // 8\n\n # normalize images\n inputs = image[None, :, [2,1,0]].to(self.device) / 255.0\n inputs = inputs.sub_(self.MEAN).div_(self.STDV)\n\n # extract features\n gmap = self.__feature_encoder(inputs) #当前帧的特征, fnet\n\n ### always add first frame to the depth video ###\n if self.video.counter.value == 0:\n net, inp = self.__context_encoder(inputs[:,[0]])\n self.net, self.inp, self.fmap = net, inp, gmap # [1,128,H//8,W//8], [1,128,H//8,W//8], [1,128,H//8,W//8]\n self.video.append(tstamp, image[0], Id, 1.0, depth, intrinsics / 8.0, gmap, net[0,0], inp[0,0])\n\n ### only add new frame if there is enough motion ###\n else: \n # index correlation volume\n coords0 = pops.coords_grid(ht, wd, device=self.device)[None,None]\n\n corr = CorrBlock(self.fmap[None,[0]], gmap[None,[0]])(coords0) #关键帧和当前帧之间的相关运算 [None,[0]]即保留第一行之后进行unsqueeze(0)，\n\n # approximate flow magnitude using 1 update iteration\n _, delta, weight = self.update(self.net[None], self.inp[None], corr)\n\n # check motion magnitue / add new frame to video\n if delta.norm(dim=-1).mean().item() > self.thresh:\n self.count = 0\n net, inp = self.__context_encoder(inputs[:,[0]]) \n self.net, self.inp, self.fmap = net, inp, gmap \n self.video.append(tstamp, image[0], None, None, depth, intrinsics / 8.0, gmap, net[0], inp[0])\n else:\n self.count += 1\n"
},
{
"path": "dbaf/multi_sensor.py",
"content": "import numpy as np\nimport gtsam\nimport math\n\nGRAVITY = 9.807\n\nclass MultiSensorState:\n def __init__(self):\n self.cur_t = 0.0\n\n \"\"\" IMU-centered states \"\"\"\n self.timestamps = [] # timestamps (len == N)\n\n self.wTbs = [] # poses (len == N)\n self.vs = [] # vels (len == N)\n self.bs = [] # biases (len == N)\n\n self.preintegrations = [] # preintegrations (len == N)\n self.preintegrations_meas = [] # raw IMU data (len == N)\n self.preintegration_temp = None # used for high-frequency prediction\n self.pose_temp = None # used for high-frequency prediction\n\n self.gnss_valid = [] # GNSS flags (len == N)\n self.gnss_position = [] # GNSS pos (len == N)\n\n self.odo_valid = [] # Odo flags (len == N)\n self.odo_vel = [] # Odo vel (len == N)\n\n self.marg_factor = None\n self.set_imu_params()\n \n def set_imu_params(self, noise = None):\n\n # default\n accel_noise_sigma = 0.0\n gyro_noise_sigma = 0.0\n accel_bias_rw_sigma = 0.0\n gyro_bias_rw_sigma = 0.0\n\n if noise != None:\n accel_noise_sigma = noise[0]\n gyro_noise_sigma = noise[1]\n accel_bias_rw_sigma = noise[2]\n gyro_bias_rw_sigma = noise[3]\n\n measured_acc_cov = np.eye(3,3) * math.pow(accel_noise_sigma,2)\n measured_omega_cov = np.eye(3,3) * math.pow(gyro_noise_sigma,2)\n integration_error_cov = np.eye(3,3) * 0e-8\n bias_acc_cov = np.eye(3,3) * math.pow(accel_bias_rw_sigma,2)\n bias_omega_cov = np.eye(3,3) * math.pow(gyro_bias_rw_sigma,2)\n bias_acc_omega_init = np.eye(6,6) * 0e-5\n\n params = gtsam.PreintegrationCombinedParams.MakeSharedU(GRAVITY)\n params.setAccelerometerCovariance(measured_acc_cov)\n params.setIntegrationCovariance(integration_error_cov)\n params.setGyroscopeCovariance(measured_omega_cov)\n params.setBiasAccCovariance(bias_acc_cov)\n params.setBiasOmegaCovariance(bias_omega_cov)\n params.setBiasAccOmegaInit(bias_acc_omega_init)\n self.params = params\n\n params_loose = gtsam.PreintegrationCombinedParams.MakeSharedU(GRAVITY)\n params_loose.setAccelerometerCovariance(measured_acc_cov* 100)\n params_loose.setIntegrationCovariance(integration_error_cov)\n params_loose.setGyroscopeCovariance(measured_omega_cov * 100)\n params_loose.setBiasAccCovariance(bias_acc_cov)\n params_loose.setBiasOmegaCovariance(bias_omega_cov)\n params_loose.setBiasAccOmegaInit(bias_acc_omega_init)\n self.params_loose = params_loose\n\n def init_first_state(self,t,pos,R,vel):\n self.timestamps.append(t)\n self.wTbs.append(gtsam.Pose3(gtsam.Rot3(R), gtsam.Point3(pos)))\n self.vs.append(vel)\n self.bs.append(gtsam.imuBias.ConstantBias(np.array([.0,.0,.0]),np.array([.0,.0,.0])))\n self.preintegrations.append(gtsam.PreintegratedCombinedMeasurements(self.params,self.bs[-1]))\n self.preintegrations_meas.append([])\n self.preintegration_temp = gtsam.PreintegratedCombinedMeasurements(self.params,self.bs[-1])\n self.gnss_valid.append(False)\n self.gnss_position.append(np.array([.0,.0,.0]))\n self.odo_valid.append(False)\n self.odo_vel.append(np.array([.0,.0,.0]))\n\n self.cur_t = t\n\n def append_imu(self, t, measuredAcc, measuredOmega):\n if t - self.cur_t > 0:\n if t-self.cur_t > 0.025: # IMU gap found, loose the IMU factor\n new_preintegration = gtsam.PreintegratedCombinedMeasurements(self.params_loose,self.bs[-1])\n for iii in range(len(self.preintegrations_meas[-1])):\n dd = self.preintegrations_meas[-1][iii]\n if dd[2] > 0:\n new_preintegration.integrateMeasurement(dd[0],dd[1],dd[2])\n self.preintegrations[-1] = new_preintegration\n self.preintegrations[-1].integrateMeasurement(\\\n measuredAcc, measuredOmega, t - self.cur_t)\n if t - self.cur_t < 0:\n raise Exception(\"may not happen\")\n self.preintegrations_meas[-1].append([measuredAcc, measuredOmega, t - self.cur_t, t])\n # print('append_imu: ',measuredAcc,measuredOmega,t - self.cur_t,t)\n self.last_measuredAcc = measuredAcc\n self.last_measuredOmega = measuredOmega\n self.cur_t = t\n \n def append_imu_temp(self, t, measuredAcc, measuredOmega, predict_pose = False):\n if t - self.cur_t > 0:\n self.preintegration_temp.integrateMeasurement(\\\n measuredAcc, measuredOmega, t - self.cur_t)\n if predict_pose:\n prev_state = gtsam.gtsam.NavState(self.wTbs[-1],self.vs[-1])\n prev_bias = self.bs[-1]\n self.pose_temp = self.preintegration_temp.predict(prev_state, prev_bias)\n\n def append_img(self, t):\n self.cur_t = t\n prev_state = gtsam.gtsam.NavState(self.wTbs[-1],self.vs[-1])\n prev_bias = self.bs[-1]\n prop_state = self.preintegrations[-1].predict(prev_state, prev_bias)\n if self.preintegrations[-1].deltaTij()>1.0:\n prop_state = gtsam.gtsam.NavState(self.wTbs[-1],self.vs[-1])\n \n self.timestamps.append(t)\n self.wTbs.append(prop_state.pose()) \n self.vs.append(prop_state.velocity())\n self.bs.append(prev_bias)\n self.gnss_valid.append(False)\n self.gnss_position.append(np.array([.0,.0,.0]))\n self.odo_valid.append(False)\n self.odo_vel.append(np.array([.0,.0,.0]))\n\n self.preintegrations.append(\\\n gtsam.PreintegratedCombinedMeasurements(self.params,self.bs[-1]))\n self.preintegrations_meas.append([])\n self.preintegration_temp = gtsam.PreintegratedCombinedMeasurements(self.params,self.bs[-1])\n \n # ugly implementation\n # this should be called after append_img()\n def append_gnss(self,t,pos):\n if math.fabs(self.cur_t - t) > 0.01:\n print('Skip GNSS data due to unsynchronization!!')\n else:\n self.gnss_valid[-1] = True\n self.gnss_position[-1] = pos\n\n def append_odo(self,t,vel):\n if math.fabs(self.cur_t - t) > 0.01:\n print('Skip ODO data due to unsynchronization!!')\n else:\n self.odo_valid[-1] = True\n self.odo_vel[-1] = vel\n\n def predict(self):\n prev_state = gtsam.gtsam.NavState(self.wTbs[-1],self.vs[-1])\n prev_bias = self.bs[-1]\n self.preintegrations[-1].predict(prev_state,prev_bias)"
},
{
"path": "demo_vio_kitti360.py",
"content": "import sys\nsys.path.append('dbaf')\nsys.path.append('dbaf/geoFunc')\n\nfrom tqdm import tqdm\nimport numpy as np\nimport torch\nimport cv2\nimport os\nimport argparse\nfrom dbaf import DBAFusion\n\nimport h5py\nimport pickle\nimport re\nimport math\nimport gtsam\nimport quaternion\n\ndef show_image(image):\n image = image.permute(1, 2, 0).cpu().numpy()\n cv2.imshow('image', image / 255.0)\n cv2.waitKey(1)\n\ndef image_stream(imagedir, imagestamp, enable_h5, h5path, calib, stride):\n \"\"\" image generator \"\"\"\n\n calib = np.loadtxt(calib, delimiter=\" \")\n fx, fy, cx, cy = calib[:4]\n\n K = np.eye(3)\n K[0,0] = fx\n K[0,2] = cx\n K[1,1] = fy\n K[1,2] = cy\n\n if not enable_h5:\n image_list = sorted(os.listdir(imagedir))[::stride]\n image_stamps = np.loadtxt(imagestamp,str)\n image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))\n for t, imfile in enumerate(image_list):\n image = cv2.imread(os.path.join(imagedir, imfile))\n\n if len(calib) > 4:\n image = cv2.undistort(image, K, calib[4:])\n tt = float(image_dict[imfile])\n\n h0, w0, _ = image.shape\n h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))\n w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))\n\n image = cv2.resize(image, (w1, h1))\n image = image[:h1-h1%8, :w1-w1%8]\n image = torch.as_tensor(image).permute(2, 0, 1)\n\n intrinsics = torch.as_tensor([fx, fy, cx, cy])\n intrinsics[0::2] *= (w1 / w0)\n intrinsics[1::2] *= (h1 / h0)\n\n yield tt, image[None], intrinsics\n else:\n ccount = 0\n h5_f = h5py.File(h5path,'r')\n all_keys = sorted(list(h5_f.keys()))\n for key in all_keys:\n ccount += 1\n yield pickle.loads(np.array(h5_f[key]))\n\nif __name__ == '__main__':\n\n print(torch.cuda.device_count())\n print(torch.cuda.is_available())\n print(torch.cuda.current_device())\n\n parser = argparse.ArgumentParser()\n parser.add_argument(\"--imagedir\", type=str, help=\"path to image directory\")\n parser.add_argument(\"--imagestamp\", type=str, help=\"\")\n parser.add_argument(\"--imupath\", type=str, help=\"\")\n parser.add_argument(\"--gtpath\", type=str, help=\"\")\n parser.add_argument(\"--enable_h5\", action=\"store_true\", help=\"\")\n parser.add_argument(\"--h5path\", type=str, help=\"\")\n parser.add_argument(\"--resultpath\", type=str, default=\"result.txt\", help=\"\")\n\n parser.add_argument(\"--calib\", type=str, help=\"path to calibration file\")\n parser.add_argument(\"--t0\", default=0, type=int, help=\"starting frame\")\n parser.add_argument(\"--stride\", default=3, type=int, help=\"frame stride\")\n\n parser.add_argument(\"--weights\", default=\"droid.pth\")\n parser.add_argument(\"--buffer\", type=int, default=80)\n parser.add_argument(\"--image_size\", default=[240, 320])\n\n parser.add_argument(\"--max_factors\", type=int, default=48, help=\"maximum active edges (which determines the GPU memory usage)\")\n parser.add_argument(\"--beta\", type=float, default=0.3, help=\"weight for translation / rotation components of flow\")\n parser.add_argument(\"--filter_thresh\", type=float, default=2.4, help=\"how much motion before considering new keyframe\")\n parser.add_argument(\"--warmup\", type=int, default=8, help=\"number of warmup frames\")\n parser.add_argument(\"--keyframe_thresh\", type=float, default=4.0, help=\"threshold to create a new keyframe\")\n parser.add_argument(\"--frontend_thresh\", type=float, default=16.0, help=\"add edges between frames whithin this distance\")\n parser.add_argument(\"--frontend_window\", type=int, default=25, help=\"frontend optimization window\")\n parser.add_argument(\"--active_window\", type=int, default=8, help=\"maximum frames involved in DBA\")\n parser.add_argument(\"--inac_range\", type=int, default=3, help=\"maximum inactive frames (whose flow wouldn't be updated) involved in DBA\")\n parser.add_argument(\"--frontend_radius\", type=int, default=2, help=\"force edges between frames within radius\")\n parser.add_argument(\"--frontend_nms\", type=int, default=1, help=\"non-maximal supression of edges\")\n parser.add_argument(\"--backend_thresh\", type=float, default=22.0)\n parser.add_argument(\"--backend_radius\", type=int, default=2)\n parser.add_argument(\"--backend_nms\", type=int, default=3)\n parser.add_argument(\"--upsample\", action=\"store_true\")\n parser.add_argument(\"--visual_only\", type=int,default=0, help=\"wheter to disbale the IMU\")\n parser.add_argument(\"--far_threshold\", type=float, default=0.02, help=\"far pixels would be downweighted (unit: m^-1)\")\n parser.add_argument(\"--translation_threshold\", type=float, default=0.2, help=\"avoid the insertion of too close keyframes (unit: m)\")\n parser.add_argument(\"--mask_threshold\", type=float, default=-1, help=\"downweight too close edges (unit: m)\")\n parser.add_argument(\"--skip_edge\", type = str, default =\"[]\", help=\"whether to add 'skip' edges in the graph (for example, [-4,-5,-6] relative to the oldest active frame)\")\n parser.add_argument(\"--save_pkl\", action=\"store_true\")\n parser.add_argument(\"--pklpath\", default=\"result.pkl\", help=\"path to saved reconstruction\")\n parser.add_argument(\"--show_plot\", action=\"store_true\", help=\"plot the image/trajectory during running\")\n \n args = parser.parse_args()\n args.skip_edge = eval(args.skip_edge)\n\n args.stereo = False\n dbaf = None\n torch.multiprocessing.set_start_method('spawn')\n\n \"\"\" Load reference trajectory (for visualization) \"\"\"\n all_gt ={}\n try:\n fp = open(args.gtpath,'rt')\n while True:\n line = fp.readline().strip()\n if line == '':break\n if line[0] == '#' : continue\n line = re.sub('\\s\\s+',' ',line)\n elem = line.split(' ')\n sod = float(elem[0])\n if sod not in all_gt.keys():\n all_gt[sod] ={}\n R = quaternion.as_rotation_matrix(quaternion.from_float_array([float(elem[7]),\\\n float(elem[4]),\\\n float(elem[5]),\\\n float(elem[6])]))\n TTT = np.eye(4,4)\n TTT[0:3,0:3] = R\n TTT[0:3,3] = np.array([ float(elem[1]), float(elem[2]), float(elem[3])])\n all_gt[sod]['T'] = TTT\n all_gt_keys =sorted(all_gt.keys())\n fp.close()\n except:\n pass\n\n \"\"\" Load IMU data \"\"\"\n all_imu = np.loadtxt(args.imupath)\n all_odo = []\n all_gnss = []\n tstamps = []\n \n \"\"\" Load images \"\"\"\n try:\n for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp, args.enable_h5,\\\n args.h5path, args.calib, args.stride)):\n if args.show_plot:\n show_image(image[0])\n if dbaf is None:\n args.image_size = [image.shape[2], image.shape[3]]\n dbaf = DBAFusion(args)\n all_imu[:,0] -= 0.04 # IMU-camera time offset\n dbaf.frontend.all_imu = all_imu\n dbaf.frontend.all_gnss = all_gnss\n dbaf.frontend.all_odo = all_odo\n dbaf.frontend.all_stamp = np.loadtxt(args.imagestamp,str)\n\n dbaf.frontend.all_stamp = dbaf.frontend.all_stamp[:,0].astype(np.float64)[None].transpose(1,0)\n if len(all_gt) > 0:\n dbaf.frontend.all_gt = all_gt\n dbaf.frontend.all_gt_keys = all_gt_keys\n \n # IMU-Camera Extrinsics\n dbaf.video.Ti1c = np.array(\n [0.99944133,-0.00228419,-0.03334389,-0.03734697,\n 0.03268308,-0.14183394,0.98935078,1.75837780,\n -0.00698916,-0.98988784,-0.14168005,0.59911765,\n 0.00000000,0.00000000,0.00000000,1.00000000]).reshape([4,4])\n dbaf.video.Tbc = gtsam.Pose3(dbaf.video.Ti1c)\n\n # IMU parameters\n dbaf.video.state.set_imu_params([ 0.0003924 * 25,0.000205689024915 * 25, 0.004905 * 10, 0.000001454441043 * 500])\n dbaf.video.init_pose_sigma = np.array([1.0, 1.0, 0.0001, 1.0, 1.0, 1.0])\n dbaf.video.init_bias_sigma = np.array([0.1, 0.1, 0.1, 0.1, 0.1, 0.1])\n dbaf.frontend.translation_threshold = args.translation_threshold\n dbaf.frontend.graph.mask_threshold = args.mask_threshold\n\n dbaf.track(t, image, intrinsics=intrinsics)\n dbaf.save_vis_easy()\n except Exception as err:\n print(err)\n dbaf.save_vis_easy()\n dbaf.terminate()\n"
},
{
"path": "demo_vio_subt.py",
"content": "import sys\nsys.path.append('dbaf')\n\nfrom tqdm import tqdm\nimport numpy as np\nimport torch\nimport cv2\nimport os\nimport argparse\nfrom dbaf import DBAFusion\n\nimport h5py\nimport pickle\nimport re\nimport math\nimport quaternion\nimport gtsam\n\ndef show_image(image):\n image = image.permute(1, 2, 0).cpu().numpy()\n cv2.imshow('image', image / 255.0)\n cv2.waitKey(1)\n\ndef image_stream(imagedir, imagestamp, enable_h5, h5path, calib, stride):\n \"\"\" image generator \"\"\"\n\n calib = np.loadtxt(calib, delimiter=\" \")\n fx, fy, cx, cy = calib[:4]\n\n K = np.eye(3)\n K[0,0] = fx\n K[0,2] = cx\n K[1,1] = fy\n K[1,2] = cy\n\n Kn = np.eye(3)\n Kn[0,0] = fx*0.35 \n Kn[0,2] = cx \n Kn[1,1] = fy*0.35\n Kn[1,2] = cy\n D = calib[5:]\n xi = np.array([calib[4]])\n\n if not enable_h5:\n image_list = [f for f in os.listdir(imagedir) if f.endswith('.png')]\n image_list = sorted(image_list,key = lambda x: int(x.split('.')[0]))[2000::stride]\n image_stamps = np.loadtxt(imagestamp,str,delimiter=',')[2000::stride]\n for ii in range(len(image_list)):\n image = cv2.imread(os.path.join(imagedir, image_list[ii]))\n\n if len(calib) > 4:\n m1, m2 = cv2.omnidir.initUndistortRectifyMap(K,D,xi,np.eye(3),Kn,(image.shape[1],image.shape[0]),cv2.CV_32FC1, cv2.omnidir.RECTIFY_PERSPECTIVE)\n image = cv2.remap(image, m1, m2, interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_CONSTANT)\n\n tt = float(image_stamps[ii]) /1e9\n\n h0, w0, _ = image.shape\n h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))\n w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))\n\n image = cv2.resize(image, (w1, h1))\n image = image[:h1-h1%8, :w1-w1%8]\n image = torch.as_tensor(image).permute(2, 0, 1)\n\n intrinsics = torch.as_tensor([fx*0.35, fy*0.35, cx, cy ])\n intrinsics[0::2] *= (w1 / w0)\n intrinsics[1::2] *= (h1 / h0)\n\n yield tt, image[None], intrinsics\n else:\n ccount = 0\n h5_f = h5py.File(h5path,'r')\n all_keys = sorted(list(h5_f.keys()))\n for key in all_keys:\n ccount += 1\n yield pickle.loads(np.array(h5_f[key]))\n\nif __name__ == '__main__':\n\n print(torch.cuda.device_count())\n print(torch.cuda.is_available())\n print(torch.cuda.current_device())\n\n parser = argparse.ArgumentParser()\n parser.add_argument(\"--imagedir\", type=str, help=\"path to image directory\")\n parser.add_argument(\"--imagestamp\", type=str, help=\"\")\n parser.add_argument(\"--imupath\", type=str, help=\"\")\n parser.add_argument(\"--gtpath\", type=str, help=\"\")\n parser.add_argument(\"--enable_h5\", action=\"store_true\", help=\"\")\n parser.add_argument(\"--h5path\", type=str, help=\"\")\n parser.add_argument(\"--resultpath\", type=str, default=\"result.txt\", help=\"\")\n\n parser.add_argument(\"--calib\", type=str, help=\"path to calibration file\")\n parser.add_argument(\"--t0\", default=0, type=int, help=\"starting frame\")\n parser.add_argument(\"--stride\", default=3, type=int, help=\"frame stride\")\n\n parser.add_argument(\"--weights\", default=\"droid.pth\")\n parser.add_argument(\"--buffer\", type=int, default=80)\n parser.add_argument(\"--image_size\", default=[240, 320])\n\n parser.add_argument(\"--max_factors\", type=int, default=48, help=\"maximum active edges (which determines the GPU memory usage)\")\n parser.add_argument(\"--beta\", type=float, default=0.3, help=\"weight for translation / rotation components of flow\")\n parser.add_argument(\"--filter_thresh\", type=float, default=2.4, help=\"how much motion before considering new keyframe\")\n parser.add_argument(\"--warmup\", type=int, default=8, help=\"number of warmup frames\")\n parser.add_argument(\"--keyframe_thresh\", type=float, default=4.0, help=\"threshold to create a new keyframe\")\n parser.add_argument(\"--frontend_thresh\", type=float, default=16.0, help=\"add edges between frames whithin this distance\")\n parser.add_argument(\"--frontend_window\", type=int, default=25, help=\"frontend optimization window\")\n parser.add_argument(\"--active_window\", type=int, default=8, help=\"maximum frames involved in DBA\")\n parser.add_argument(\"--inac_range\", type=int, default=3, help=\"maximum inactive frames (whose flow wouldn't be updated) involved in DBA\")\n parser.add_argument(\"--frontend_radius\", type=int, default=2, help=\"force edges between frames within radius\")\n parser.add_argument(\"--frontend_nms\", type=int, default=1, help=\"non-maximal supression of edges\")\n parser.add_argument(\"--backend_thresh\", type=float, default=22.0)\n parser.add_argument(\"--backend_radius\", type=int, default=2)\n parser.add_argument(\"--backend_nms\", type=int, default=3)\n parser.add_argument(\"--upsample\", action=\"store_true\")\n parser.add_argument(\"--visual_only\", type=int,default=0, help=\"wheter to disbale the IMU\")\n parser.add_argument(\"--far_threshold\", type=float, default=0.02, help=\"far pixels would be downweighted (unit: m^-1)\")\n parser.add_argument(\"--translation_threshold\", type=float, default=0.2, help=\"avoid the insertion of too close keyframes (unit: m)\")\n parser.add_argument(\"--mask_threshold\", type=float, default=-1, help=\"downweight too close edges (unit: m)\")\n parser.add_argument(\"--skip_edge\", type = str, default =\"[]\", help=\"whether to add 'skip' edges in the graph (for example, [-4,-5,-6] relative to the oldest active frame)\")\n parser.add_argument(\"--save_pkl\", action=\"store_true\")\n parser.add_argument(\"--pklpath\", default=\"result.pkl\", help=\"path to saved reconstruction\")\n parser.add_argument(\"--show_plot\", action=\"store_true\", help=\"plot the trajectory during running\")\n \n args = parser.parse_args()\n args.skip_edge = eval(args.skip_edge)\n\n args.stereo = False\n dbaf = None\n torch.multiprocessing.set_start_method('spawn')\n\n \"\"\" Load reference trajectory (for visualization) \"\"\"\n all_gt ={}\n try:\n fp = open(args.gtpath,'rt')\n while True:\n line = fp.readline().strip()\n if line == '':break\n if line[0] == '#' : continue\n line = re.sub('\\s\\s+',' ',line)\n elem = line.split(',')\n sod = float(elem[0])/1e9\n if sod not in all_gt.keys():\n all_gt[sod] ={}\n R = quaternion.as_rotation_matrix(quaternion.from_float_array([float(elem[4]),\\\n float(elem[5]),\\\n float(elem[6]),\\\n float(elem[7])]))\n TTT = np.eye(4,4)\n TTT[0:3,0:3] = R\n TTT[0:3,3] = np.array([ float(elem[1]), float(elem[2]), float(elem[3])])\n all_gt[sod]['T'] = TTT\n all_gt_keys =sorted(all_gt.keys())\n fp.close()\n except:\n pass\n\n \"\"\" Load IMU data \"\"\"\n all_imu = np.loadtxt(args.imupath,delimiter=',',comments='#',skiprows=1)\n all_imu[:,0] /= 1e9\n # all_imu[:,1:4] *= 180/math.pi\n all_imu_new = np.zeros_like(all_imu)\n all_imu_new[:,0] = all_imu[:,0]\n all_imu_new[:,1:4] = all_imu[:,5:8] * 180/math.pi\n all_imu_new[:,4:7] = all_imu[:,8:11]\n all_imu_new = all_imu_new[:,:7]\n \n tstamps = []\n\n \"\"\" Load images \"\"\"\n for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp, args.enable_h5,\\\n args.h5path, args.calib, args.stride)):\n if args.show_plot:\n show_image(image[0])\n if dbaf is None:\n args.image_size = [image.shape[2], image.shape[3]]\n dbaf = DBAFusion(args)\n dbaf.frontend.all_imu = all_imu_new\n dbaf.frontend.all_gnss = []\n dbaf.frontend.all_odo = []\n dbaf.frontend.all_stamp = np.loadtxt(args.imagestamp,str,delimiter=',')\n dbaf.frontend.all_stamp = dbaf.frontend.all_stamp.astype(np.float64)[2000::args.stride][None].transpose(1,0)/1e9\n if len(all_gt) > 0:\n dbaf.frontend.all_gt = all_gt\n dbaf.frontend.all_gt_keys = all_gt_keys\n \n # IMU-Camera Extrinsics\n dbaf.video.Ti1c = np.array(\n [-0.04279531, -0.00237969, 0.99908103, 0.19499356,\n -0.99880330, -0.02359508, -0.04283961, 0.04340662,\n 0.02367534, -0.99971877, -0.00136708, -0.01782382,\n 0.00000000, 0.00000000, 0.00000000, 1.00000000]).reshape([4,4])\n dbaf.video.Tbc = gtsam.Pose3(dbaf.video.Ti1c)\n \n # IMU parameters\n dbaf.video.state.set_imu_params((np.array([ 0.0003924 * 25,0.000205689024915 * 25, 0.004905 * 10, 0.000001454441043 * 500])*1.0).tolist())\n dbaf.video.init_pose_sigma = np.array([1, 1, 0.0001, 0.0001,0.0001,0.0001])\n dbaf.video.init_bias_sigma = np.array([1.0,1.0,1.0,1.0,1.0,1.0])\n dbaf.frontend.translation_threshold = args.translation_threshold\n dbaf.frontend.graph.mask_threshold = args.mask_threshold\n\n dbaf.track(t, image, intrinsics=intrinsics)\n\n if args.save_pkl:\n dbaf.save_vis_easy()\n\n dbaf.terminate()\n"
},
{
"path": "demo_vio_tumvi.py",
"content": "import sys\nsys.path.append('dbaf')\n\nfrom tqdm import tqdm\nimport numpy as np\nimport torch\nimport cv2\nimport os\nimport argparse\nfrom dbaf import DBAFusion\n\nimport h5py\nimport pickle\nimport re\nimport math\nimport quaternion\nimport gtsam\n\ndef show_image(image):\n image = image.permute(1, 2, 0).cpu().numpy()\n cv2.imshow('image', image / 255.0)\n cv2.waitKey(1)\n\ndef image_stream(imagedir, imagestamp, enable_h5, h5path, calib, stride):\n \"\"\" image generator \"\"\"\n\n calib = np.loadtxt(calib, delimiter=\" \")\n fx, fy, cx, cy = calib[:4]\n\n K = np.eye(3)\n K[0,0] = fx\n K[0,2] = cx\n K[1,1] = fy\n K[1,2] = cy\n\n Kn = np.eye(3)\n Kn[0,0] = fx \n Kn[0,2] = cx \n Kn[1,1] = fy \n Kn[1,2] = cy\n\n if not enable_h5:\n image_list = sorted(os.listdir(imagedir))[::stride]\n image_stamps = np.loadtxt(imagestamp,str,delimiter=',')\n image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))\n for t, imfile in enumerate(image_list):\n image = cv2.imread(os.path.join(imagedir, imfile))\n\n if len(calib) > 4:\n m1, m2 = cv2.fisheye.initUndistortRectifyMap(K,calib[4:],np.eye(3),Kn,(512,512),cv2.CV_32FC1)\n image = cv2.remap(image, m1, m2, interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_CONSTANT)\n\n tt = float(image_dict[imfile]) /1e9\n\n h0, w0, _ = image.shape\n h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))\n w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))\n\n image = cv2.resize(image, (w1, h1))\n image = image[:h1-h1%8, :w1-w1%8]\n image = torch.as_tensor(image).permute(2, 0, 1)\n\n intrinsics = torch.as_tensor([fx, fy, cx, cy ])\n intrinsics[0::2] *= (w1 / w0)\n intrinsics[1::2] *= (h1 / h0)\n\n yield tt, image[None], intrinsics\n else:\n ccount = 0\n h5_f = h5py.File(h5path,'r')\n all_keys = sorted(list(h5_f.keys()))\n for key in all_keys:\n ccount += 1\n yield pickle.loads(np.array(h5_f[key]))\n\nif __name__ == '__main__':\n\n print(torch.cuda.device_count())\n print(torch.cuda.is_available())\n print(torch.cuda.current_device())\n\n parser = argparse.ArgumentParser()\n parser.add_argument(\"--imagedir\", type=str, help=\"path to image directory\")\n parser.add_argument(\"--imagestamp\", type=str, help=\"\")\n parser.add_argument(\"--imupath\", type=str, help=\"\")\n parser.add_argument(\"--gtpath\", type=str, help=\"\")\n parser.add_argument(\"--enable_h5\", action=\"store_true\", help=\"\")\n parser.add_argument(\"--h5path\", type=str, help=\"\")\n parser.add_argument(\"--resultpath\", type=str, default=\"result.txt\", help=\"\")\n\n parser.add_argument(\"--calib\", type=str, help=\"path to calibration file\")\n parser.add_argument(\"--t0\", default=0, type=int, help=\"starting frame\")\n parser.add_argument(\"--stride\", default=3, type=int, help=\"frame stride\")\n\n parser.add_argument(\"--weights\", default=\"droid.pth\")\n parser.add_argument(\"--buffer\", type=int, default=80)\n parser.add_argument(\"--image_size\", default=[240, 320])\n\n parser.add_argument(\"--max_factors\", type=int, default=48, help=\"maximum active edges (which determines the GPU memory usage)\")\n parser.add_argument(\"--beta\", type=float, default=0.3, help=\"weight for translation / rotation components of flow\")\n parser.add_argument(\"--filter_thresh\", type=float, default=2.4, help=\"how much motion before considering new keyframe\")\n parser.add_argument(\"--warmup\", type=int, default=8, help=\"number of warmup frames\")\n parser.add_argument(\"--keyframe_thresh\", type=float, default=3.0, help=\"threshold to create a new keyframe\")\n parser.add_argument(\"--frontend_thresh\", type=float, default=16.0, help=\"add edges between frames whithin this distance\")\n parser.add_argument(\"--frontend_window\", type=int, default=25, help=\"frontend optimization window\")\n parser.add_argument(\"--active_window\", type=int, default=8, help=\"maximum frames involved in DBA\")\n parser.add_argument(\"--inac_range\", type=int, default=3, help=\"maximum inactive frames (whose flow wouldn't be updated) involved in DBA\")\n parser.add_argument(\"--frontend_radius\", type=int, default=2, help=\"force edges between frames within radius\")\n parser.add_argument(\"--frontend_nms\", type=int, default=1, help=\"non-maximal supression of edges\")\n parser.add_argument(\"--backend_thresh\", type=float, default=22.0)\n parser.add_argument(\"--backend_radius\", type=int, default=2)\n parser.add_argument(\"--backend_nms\", type=int, default=3)\n parser.add_argument(\"--upsample\", action=\"store_true\")\n parser.add_argument(\"--visual_only\", type=int,default=0, help=\"wheter to disbale the IMU\")\n parser.add_argument(\"--far_threshold\", type=float, default=0.02, help=\"far pixels would be downweighted (unit: m^-1)\")\n parser.add_argument(\"--translation_threshold\", type=float, default=0.2, help=\"avoid the insertion of too close keyframes (unit: m)\")\n parser.add_argument(\"--mask_threshold\", type=float, default=-1, help=\"downweight too close edges (unit: m)\")\n parser.add_argument(\"--skip_edge\", type = str, default =\"[]\", help=\"whether to add 'skip' edges in the graph (for example, [-4,-5,-6] relative to the oldest active frame)\")\n parser.add_argument(\"--save_pkl\", action=\"store_true\")\n parser.add_argument(\"--pklpath\", default=\"result.pkl\", help=\"path to saved reconstruction\")\n parser.add_argument(\"--show_plot\", action=\"store_true\", help=\"plot the trajectory during running\")\n \n args = parser.parse_args()\n args.skip_edge = eval(args.skip_edge)\n\n args.stereo = False\n dbaf = None\n torch.multiprocessing.set_start_method('spawn')\n\n \"\"\" Load reference trajectory (for visualization) \"\"\"\n all_gt ={}\n try:\n fp = open(args.gtpath,'rt')\n while True:\n line = fp.readline().strip()\n if line == '':break\n if line[0] == '#' : continue\n line = re.sub('\\s\\s+',' ',line)\n elem = line.split(',')\n sod = float(elem[0])/1e9\n if sod not in all_gt.keys():\n all_gt[sod] ={}\n R = quaternion.as_rotation_matrix(quaternion.from_float_array([float(elem[4]),\\\n float(elem[5]),\\\n float(elem[6]),\\\n float(elem[7])]))\n TTT = np.eye(4,4)\n TTT[0:3,0:3] = R\n TTT[0:3,3] = np.array([ float(elem[1]), float(elem[2]), float(elem[3])])\n all_gt[sod]['T'] = TTT\n all_gt_keys =sorted(all_gt.keys())\n fp.close()\n except:\n pass\n\n \"\"\" Load IMU data \"\"\"\n all_imu = np.loadtxt(args.imupath,delimiter=',')\n all_imu[:,0] /= 1e9\n all_imu[:,1:4] *= 180/math.pi\n \n tstamps = []\n\n \"\"\" Load images \"\"\"\n clahe = cv2.createCLAHE(2.0,tileGridSize=(8, 8))\n for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp, args.enable_h5,\\\n args.h5path, args.calib, args.stride)):\n mm = clahe.apply(image[0][0].numpy())\n image[0] = torch.tensor(mm[None].repeat(3,0))\n if args.show_plot:\n show_image(image[0])\n if dbaf is None:\n args.image_size = [image.shape[2], image.shape[3]]\n dbaf = DBAFusion(args)\n dbaf.frontend.all_imu = all_imu\n dbaf.frontend.all_gnss = []\n dbaf.frontend.all_odo = []\n dbaf.frontend.all_stamp = np.loadtxt(args.imagestamp,str,delimiter=',')\n dbaf.frontend.all_stamp = dbaf.frontend.all_stamp[:,0].astype(np.float64)[None].transpose(1,0)/1e9\n if len(all_gt) > 0:\n dbaf.frontend.all_gt = all_gt\n dbaf.frontend.all_gt_keys = all_gt_keys\n \n # IMU-Camera Extrinsics\n dbaf.video.Ti1c = np.array(\n [-0.9995250378696743, 0.029615343885863205, -0.008522328211654736, 0.04727988224914392,\n 0.0075019185074052044, -0.03439736061393144, -0.9993800792498829, -0.047443232143367084,\n -0.02989013031643309, -0.998969345370175, 0.03415885127385616, -0.0681999605066297,\n 0.0, 0.0, 0.0, 1.0]).reshape([4,4])\n dbaf.video.Ti1c = np.linalg.inv(dbaf.video.Ti1c)\n dbaf.video.Tbc = gtsam.Pose3(dbaf.video.Ti1c)\n \n # IMU parameters\n dbaf.video.state.set_imu_params((np.array([ 0.0003924 * 25,0.000205689024915 * 25, 0.004905 * 10, 0.000001454441043 * 5000])*1.0).tolist())\n dbaf.video.init_pose_sigma = np.array([0.1, 0.1, 0.0001, 0.0001,0.0001,0.0001])\n dbaf.video.init_bias_sigma = np.array([1.0,1.0,1.0, 1.0,1.0,1.0])\n dbaf.frontend.translation_threshold = args.translation_threshold\n dbaf.frontend.graph.mask_threshold = args.mask_threshold\n\n dbaf.track(t, image, intrinsics=intrinsics)\n\n if args.save_pkl:\n dbaf.save_vis_easy()\n\n dbaf.terminate()\n"
},
{
"path": "demo_vio_whu.py",
"content": "import sys\nsys.path.append('dbaf')\nsys.path.append('dbaf/geoFunc')\nfrom tqdm import tqdm\nimport numpy as np\nimport torch\nimport cv2\nimport os\nimport argparse\nfrom dbaf import DBAFusion\n\nimport h5py\nimport pickle\nimport re\nimport math\nimport gtsam\nimport geoFunc.trans as trans\n\ndef show_image(image):\n image = image.permute(1, 2, 0).cpu().numpy()\n cv2.imshow('image', image / 255.0)\n cv2.waitKey(1)\n\ndef image_stream(imagedir, imagestamp, enable_h5, h5path, calib, stride):\n \"\"\" image generator \"\"\"\n\n calib = np.loadtxt(calib, delimiter=\" \")\n fx, fy, cx, cy = calib[:4]\n\n K = np.eye(3)\n K[0,0] = fx\n K[0,2] = cx\n K[1,1] = fy\n K[1,2] = cy\n \n\n if not enable_h5:\n image_stamps = np.loadtxt(imagestamp,str,delimiter=',')\n image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))\n image_list = list(image_dict)\n ccount = 0\n for t, imfile in enumerate(image_list):\n tt = float(image_dict[imfile])\n if int(tt*10)%2 == 1: continue\n\n ccount += 1\n\n image = cv2.imread(os.path.join(imagedir, imfile))\n\n if len(calib) > 4:\n image = cv2.undistort(image, K, calib[4:])\n\n h0, w0, _ = image.shape\n h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))\n w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))\n\n image = cv2.resize(image, (w1, h1))\n image = image[:h1-h1%8, :w1-w1%8]\n image = torch.as_tensor(image).permute(2, 0, 1)\n\n intrinsics = torch.as_tensor([fx, fy, cx, cy])\n intrinsics[0::2] *= (w1 / w0)\n intrinsics[1::2] *= (h1 / h0)\n\n yield tt, image[None], intrinsics\n else:\n raise Exception()\n\n\nif __name__ == '__main__':\n print(torch.cuda.device_count())\n print(torch.cuda.is_available())\n print(torch.cuda.current_device())\n\n parser = argparse.ArgumentParser()\n parser.add_argument(\"--imagedir\", type=str, help=\"path to image directory\")\n parser.add_argument(\"--imagestamp\", type=str, help=\"\")\n parser.add_argument(\"--imupath\", type=str, help=\"\")\n parser.add_argument(\"--gtpath\", type=str, help=\"\")\n parser.add_argument(\"--enable_h5\", action=\"store_true\", help=\"\")\n parser.add_argument(\"--h5path\", type=str, help=\"\")\n parser.add_argument(\"--resultpath\", type=str, help=\"\")\n\n parser.add_argument(\"--calib\", type=str, help=\"path to calibration file\")\n parser.add_argument(\"--t0\", default=0, type=int, help=\"starting frame\")\n parser.add_argument(\"--stride\", default=3, type=int, help=\"frame stride\")\n\n parser.add_argument(\"--weights\", default=\"droid.pth\")\n parser.add_argument(\"--buffer\", type=int, default=80)\n parser.add_argument(\"--image_size\", default=[240, 320])\n parser.add_argument(\"--max_factors\", type=int, default=48, help=\"maximum active edges (which determines the GPU memory usage)\")\n parser.add_argument(\"--beta\", type=float, default=0.3, help=\"weight for translation / rotation components of flow\")\n parser.add_argument(\"--filter_thresh\", type=float, default=0.00, help=\"how much motion before considering new keyframe\")\n parser.add_argument(\"--warmup\", type=int, default=8, help=\"number of warmup frames\")\n parser.add_argument(\"--vi_warmup\", type=int, default=15, help=\"\")\n parser.add_argument(\"--keyframe_thresh\", type=float, default=3.0, help=\"threshold to create a new keyframe\")\n parser.add_argument(\"--frontend_thresh\", type=float, default=16.0, help=\"add edges between frames whithin this distance\")\n parser.add_argument(\"--frontend_window\", type=int, default=25, help=\"frontend optimization window\")\n parser.add_argument(\"--active_window\", type=int, default=8)\n parser.add_argument(\"--inac_range\", type=int, default=3)\n parser.add_argument(\"--frontend_radius\", type=int, default=2, help=\"force edges between frames within radius\")\n parser.add_argument(\"--frontend_nms\", type=int, default=1, help=\"non-maximal supression of edges\")\n parser.add_argument(\"--backend_thresh\", type=float, default=22.0)\n parser.add_argument(\"--backend_radius\", type=int, default=2)\n parser.add_argument(\"--backend_nms\", type=int, default=3)\n parser.add_argument(\"--upsample\", action=\"store_true\")\n parser.add_argument(\"--visual_only\", type=int,default=0, help=\"wheter to disbale the IMU\")\n parser.add_argument(\"--far_threshold\", type=float, default=0.02, help=\"far pixels would be downweighted (unit: m^-1)\")\n parser.add_argument(\"--translation_threshold\", type=float, default=0.2, help=\"avoid the insertion of too close keyframes (unit: m)\")\n parser.add_argument(\"--mask_threshold\", type=float, default=-1, help=\"downweight too close edges (unit: m)\")\n parser.add_argument(\"--skip_edge\", type = str, default =\"[]\", help=\"whether to add 'skip' edges in the graph (for example, [-4,-5,-6] relative to the oldest active frame)\")\n parser.add_argument(\"--save_pkl\", action=\"store_true\")\n parser.add_argument(\"--pklpath\", default=\"result.pkl\", help=\"path to saved reconstruction\")\n parser.add_argument(\"--show_plot\", action=\"store_true\", help=\"plot the image/trajectory during running\")\n parser.add_argument(\"--use_gnss\", action=\"store_true\")\n parser.add_argument(\"--gnsspath\", type=str, help=\"\")\n parser.add_argument(\"--use_odo\", action=\"store_true\")\n parser.add_argument(\"--odopath\", type=str, help=\"\")\n parser.add_argument(\"--use_zupt\", action=\"store_true\")\n args = parser.parse_args()\n args.skip_edge = eval(args.skip_edge)\n\n args.stereo = False\n dbaf = None\n\n\n all_gt ={}\n Ti0i1 =np.array([[ 9.99902524e-01, 1.39619889e-02, -7.31054713e-05, 1.00000000e-02],\n [-1.39621803e-02, 9.99888818e-01, -5.23545345e-03, -2.05000000e-01],\n [ 0.00000000e+00, 5.23596383e-03, 9.99986292e-01, -5.00000000e-02],\n [ 0.00000000e+00, 0.00000000e+00, 0.00000000e+00, 1.00000000e+00]])\n Ten0 = None\n is_ref_set = False\n fp = open(args.gtpath,'rt')\n while True:\n line = fp.readline().strip()\n if line == '':break\n if line[0] == '#' :continue\n line = re.sub('\\s\\s+',' ',line)\n elem = line.split(' ')\n sod = float(elem[1])\n if sod not in all_gt.keys():\n all_gt[sod] ={}\n all_gt[sod]['X0'] = float(elem[2])\n all_gt[sod]['Y0'] = float(elem[3])\n all_gt[sod]['Z0'] = float(elem[4])\n all_gt[sod]['VX0'] = float(elem[15])\n all_gt[sod]['VY0'] = float(elem[16])\n all_gt[sod]['VZ0'] = float(elem[17])\n all_gt[sod]['ATTX0']= float(elem[25])\n all_gt[sod]['ATTY0']= float(elem[26])\n all_gt[sod]['ATTZ0']= -float(elem[24])\n Ren = trans.Cen([all_gt[sod]['X0'],all_gt[sod]['Y0'],all_gt[sod]['Z0']])\n ani0 = [all_gt[sod]['ATTX0']/180*math.pi,\\\n all_gt[sod]['ATTY0']/180*math.pi,\\\n all_gt[sod]['ATTZ0']/180*math.pi]\n Rni0 = trans.att2m(ani0)\n Rei0 = np.matmul(Ren,Rni0)\n tei0 = np.array([all_gt[sod]['X0'],all_gt[sod]['Y0'],all_gt[sod]['Z0']])\n Tei0 = np.eye(4,4)\n Tei0[0:3,0:3] = Rei0\n Tei0[0:3,3] = tei0\n if not is_ref_set:\n is_ref_set = True\n Ten0 = np.eye(4,4)\n Ten0[0:3,0:3] = trans.Cen(tei0)\n Ten0[0:3,3] = tei0\n Tn0i0 = np.matmul(np.linalg.inv(Ten0),Tei0)\n Tn0i1 = np.matmul(Tn0i0,Ti0i1)\n all_gt[sod]['T'] = Tn0i1\n all_gt_keys =sorted(all_gt.keys())\n fp.close()\n\n all_imu = np.loadtxt(args.imupath,delimiter=' ')\n\n if args.use_gnss and os.path.isfile(args.gnsspath):\n fix_map = {b'Fixed':1.0,b'Float':0.0}\n all_gnss = np.genfromtxt(args.gnsspath,converters={16: lambda x: fix_map[x]})\n else:\n all_gnss = []\n if args.use_odo and os.path.isfile(args.odopath):\n all_odo = np.genfromtxt(args.odopath)\n all_odo = all_odo[np.fabs(all_odo[:,0] - np.round(all_odo[:,0]))<0.001]\n np.random.seed(12345)\n all_odo[:,1:] += np.random.randn(all_odo.shape[0],3)*0.05\n else:\n all_odo = []\n\n tstamps = []\n # try:\n for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp, args.enable_h5,\\\n args.h5path, args.calib, args.stride)):\n \n if args.show_plot:\n show_image(image[0])\n if dbaf is None:\n args.image_size = [image.shape[2], image.shape[3]]\n dbaf = DBAFusion(args)\n dbaf.frontend.all_imu = all_imu\n dbaf.frontend.all_stamp = np.loadtxt(args.imagestamp,str,delimiter=',')\n dbaf.frontend.all_gnss = all_gnss\n dbaf.frontend.all_odo = all_odo\n if len(all_gt) > 0:\n dbaf.frontend.all_gt = all_gt\n dbaf.frontend.all_gt_keys = all_gt_keys\n dbaf.video.Ti1c = np.array(\n [0.99988370,-0.00563944,-0.01418468,-0.15590000,\n 0.01424932,0.01159187,0.99983149,0.63466000,\n -0.00547407,-0.99991712,0.01167088,0.04605000,\n 0.00000000,0.00000000,0.00000000,1.00000000]).reshape([4,4])\n dbaf.video.tbg = np.array([-0.0125, -0.26, 0.2091])\n\n dbaf.video.Tbc = gtsam.Pose3(dbaf.video.Ti1c)\n dbaf.video.state.set_imu_params([ 0.0003924 * 25,0.000205689024915 * 25, 0.004905 * 10, 0.000001454441043 * 25])\n if args.use_gnss:\n dbaf.video.init_pose_sigma = np.array([1.0, 1.0, 10.0,10.0,10.0,10.0])\n else:\n dbaf.video.init_pose_sigma = np.array([[0.1, 0.1, 0.0001, 0.0001,0.0001,0.0001],\n [1.0, 1.0, 0.0001, 10.0, 10.0, 10.0]])\n dbaf.video.init_bias_sigma = np.array([1.0,1.0,1.0, 0.1, 0.1, 0.1])\n dbaf.frontend.translation_threshold = args.translation_threshold\n dbaf.frontend.graph.mask_threshold = args.mask_threshold\n dbaf.track(t, image, intrinsics=intrinsics)\n dbaf.save_vis_easy()\n dbaf.terminate()\n"
},
{
"path": "evaluation_scripts/batch_tumvi.py",
"content": "import subprocess\n\nfor seq in ['magistrale1','magistrale2','magistrale3','magistrale4','magistrale5','magistrale6',\\\n 'outdoors1','outdoors2','outdoors3','outdoors4','outdoors5','outdoors6','outdoors7','outdoors8']:\n p = subprocess.Popen('python ./evaluation_scripts/evaluate_tumvi.py --batch --seq=%s | grep rmse' % seq, shell = True)\n p.wait()"
},
{
"path": "evaluation_scripts/evaluate_kitti.py",
"content": "import argparse\nimport logging\nimport typing\n\nimport numpy as np\n\nimport evo.common_ape_rpe as common\nfrom evo.core import lie_algebra, sync, metrics\nfrom evo.core.result import Result\nfrom evo.core.trajectory import PosePath3D, PoseTrajectory3D\nfrom evo.tools import file_interface, log\nfrom evo.tools.settings import SETTINGS\n\nimport matplotlib.pyplot as plt\nimport copy\nfrom scipy.spatial.transform import Rotation\nimport bisect\nimport math\nimport time\n\nlogger = logging.getLogger(__name__)\n\nSEP = \"-\" * 80 # separator line\n\ndef ape(traj_ref: PosePath3D, traj_est: PosePath3D,\n pose_relation: metrics.PoseRelation, align: bool = False,\n correct_scale: bool = False, n_to_align: int = -1,\n align_origin: bool = False, ref_name: str = \"reference\",\n est_name: str = \"estimate\",\n change_unit: typing.Optional[metrics.Unit] = None) -> Result:\n if n_to_align >0 : \n print('>>>>> only use the starting segment')\n n_to_align = np.where((np.array(traj_ref.timestamps)[1:]-np.array(traj_ref.timestamps)[0:-1])>100)[0][0]-1\n\n # Align the trajectories.\n only_scale = correct_scale and not align\n alignment_transformation = None\n if align or correct_scale:\n logger.debug(SEP)\n alignment_transformation = lie_algebra.sim3(\n *traj_est.align(traj_ref, correct_scale, only_scale, n=n_to_align))\n elif align_origin:\n logger.debug(SEP)\n alignment_transformation = traj_est.align_origin(traj_ref)\n\n # Calculate APE.\n logger.debug(SEP)\n data = (traj_ref, traj_est)\n ape_metric = metrics.APE(pose_relation)\n ape_metric.process_data(data)\n\n if change_unit:\n ape_metric.change_unit(change_unit)\n\n title = str(ape_metric)\n if align and not correct_scale:\n title += \"\\n(with SE(3) Umeyama alignment)\"\n elif align and correct_scale:\n title += \"\\n(with Sim(3) Umeyama alignment)\"\n elif only_scale:\n title += \"\\n(scale corrected)\"\n elif align_origin:\n title += \"\\n(with origin alignment)\"\n else:\n title += \"\\n(not aligned)\"\n if (align or correct_scale) and n_to_align != -1:\n title += \" (aligned poses: {})\".format(n_to_align)\n\n ape_result = ape_metric.get_result(ref_name, est_name)\n ape_result.info[\"title\"] = title\n\n logger.debug(SEP)\n logger.info(ape_result.pretty_str())\n\n ape_result.add_trajectory(ref_name, traj_ref)\n ape_result.add_trajectory(est_name, traj_est)\n if isinstance(traj_est, PoseTrajectory3D):\n seconds_from_start = np.array(\n [t - traj_est.timestamps[0] for t in traj_est.timestamps])\n ape_result.add_np_array(\"seconds_from_start\", seconds_from_start)\n ape_result.add_np_array(\"timestamps\", traj_est.timestamps)\n ape_result.add_np_array(\"distances_from_start\", traj_ref.distances)\n ape_result.add_np_array(\"distances\", traj_est.distances)\n\n if alignment_transformation is not None:\n ape_result.add_np_array(\"alignment_transformation_sim3\",\n alignment_transformation)\n\n return ape_result\n\n\nif __name__ == '__main__':\n color_list = [[0,0,1],[1,0.6,1],[1,0,0]]\n plt.figure('1',figsize=[6,6])\n parser = argparse.ArgumentParser()\n parser.add_argument('--seq', type=str, help='seq',default='0010')\n args = parser.parse_args()\n args.subcommand = 'tum'\n seq = args.seq\n args.ref_file = '/home/zhouyuxuan/data/2013_05_28_drive_%s_sync/gt_local.txt' % seq\n args.pose_relation = 'trans_part'\n args.align = True\n args.correct_scale = False\n args.n_to_align = 1\n args.align_origin = False\n args.plot_mode = 'xyz'\n args.plot_x_dimension = 'seconds'\n args.plot_colormap_min = None\n args.plot_colormap_max = None\n args.plot_colormap_max_percentile = None\n args.ros_map_yaml = None\n args.plot = True\n args.est_files = ['results/result_%s.txt' %seq]\n label_list = ['DBA-Fusion (M)']\n args.save_plot = False\n args.serialize_plot = False\n for iii in range(len(args.est_files)):\n args.est_file = args.est_files[iii]\n\n if args.est_file.find('visual') != -1:\n args.correct_scale = True\n else:\n args.correct_scale = False\n traj_ref, traj_est, ref_name, est_name = common.load_trajectories(args)\n traj_ref_sel, traj_est_sel = sync.associate_trajectories(\n traj_ref, traj_est, 0.01,0.0,\n first_name=ref_name, snd_name=est_name)\n args.n_to_align = -1\n pose_relation = common.get_pose_relation(args)\n result = ape(traj_ref=traj_ref_sel, traj_est=traj_est_sel,\n pose_relation=pose_relation, align=args.align,\n correct_scale=args.correct_scale, n_to_align=args.n_to_align,\n align_origin=args.align_origin, ref_name=ref_name,\n est_name=est_name)\n traj_est_sel = copy.deepcopy(result.trajectories[est_name])\n T01 = result.np_arrays['alignment_transformation_sim3']\n print(T01)\n result = ape(traj_ref=traj_ref_sel, traj_est=traj_est_sel,\n pose_relation=pose_relation, align=args.align,\n correct_scale=False, n_to_align=-1,\n align_origin=args.align_origin, ref_name=ref_name,\n est_name=est_name)\n print(result)\n traj_est.transform(T01)\n \n if iii == 0:\n x_series=[]\n y_series=[]\n z_series=[]\n for i in range(len(traj_ref.poses_se3)):\n TTT = traj_ref.poses_se3[i]\n x_series.append(TTT[0,3])\n y_series.append(TTT[1,3])\n z_series.append(TTT[2,3])\n plt.plot(x_series,y_series,c=[0,0,0],linestyle = '--')\n\n x_series=[]\n y_series=[]\n z_series=[]\n for i in range(len(traj_est.poses_se3)):\n TTT = traj_est.poses_se3[i]\n x_series.append(TTT[0,3])\n y_series.append(TTT[1,3])\n z_series.append(TTT[2,3])\n ppp = TTT[0:3,3]\n qqq = Rotation.from_matrix(TTT[:3, :3]/np.power(np.linalg.det(TTT[:3, :3]),1.0/3)).as_quat()\n plt.plot(x_series,y_series,c=color_list[iii],label = label_list[iii])\n \n # t_series=[]\n # x_series=[]\n # y_series=[]\n # z_series=[]\n # for i in range(len(traj_ref_sel.timestamps)):\n # T0 = traj_ref_sel.poses_se3[i]\n # T1 = traj_est_sel.poses_se3[i]\n # T01 = np.matmul(np.linalg.inv(T0),T1)\n # att = Rotation.from_matrix(T01[0:3,0:3]).as_rotvec()\n # t_series.append(traj_ref_sel.timestamps[i])\n # x_series.append(att[0])\n # y_series.append(att[1])\n # z_series.append(att[2])\n # plt.figure()\n # plt.plot(t_series,x_series)\n # plt.plot(t_series,y_series)\n # plt.plot(t_series,z_series)\n # plt.show()\n\n print('Evaluating relative pose error ...')\n subtraj_length = [100,200,300,400,500,600,700,800]\n max_dist_difH=1\n rel_trans_error_dist = []\n rel_att_error_dist = []\n for i in range(8):\n subsection_index=[]\n max_dist_diff=0.2*subtraj_length[i]\n traj_len = len(traj_ref_sel.timestamps)\n for j in range(traj_len-2):\n k = bisect.bisect(traj_ref_sel.distances,traj_ref_sel.distances[j]+subtraj_length[i]-max_dist_difH)\n if k > 0 and k < traj_len and math.fabs(traj_ref_sel.distances[k] - (traj_ref_sel.distances[j]+subtraj_length[i]))< max_dist_difH:\n subsection_index.append([j,k])\n print(\"The trajectory at %dm have %d matching points... \" %(subtraj_length[i],len(subsection_index)))\n rel_tran_errors = []\n rel_att_errors = []\n for ii in subsection_index:\n T_gt_1 =traj_ref_sel.poses_se3[ii[0]]\n T_gt_2 =traj_ref_sel.poses_se3[ii[1]]\n T_est_1 =traj_est_sel.poses_se3[ii[0]]\n T_est_2 =traj_est_sel.poses_se3[ii[1]]\n T_gt_12=np.matmul(np.linalg.inv(T_gt_1),T_gt_2)\n T_est_12=np.matmul(np.linalg.inv(T_est_1),T_est_2)\n T_error=np.matmul(np.linalg.inv(T_gt_12),T_est_12)\n rel_tran_error = np.linalg.norm(T_error[0:3,3])\n rel_att_error = np.linalg.norm(Rotation.from_matrix(T_error[0:3,0:3]).as_rotvec())\n rel_tran_errors.append(rel_tran_error/subtraj_length[i]*100)\n rel_att_errors.append(rel_att_error/subtraj_length[i]*100/math.pi*180)\n rel_trans_error_dist.append(np.mean(np.array(rel_tran_errors)))\n rel_att_error_dist.append(np.mean(np.array(rel_att_errors)))\n print('Relative Translation Error: %f%%' % np.mean(np.array(rel_trans_error_dist)))\n print('Relative Rotation Error: %f deg / 100 m' % np.mean(np.array(rel_att_error_dist)))\n plt.show()"
},
{
"path": "evaluation_scripts/evaluate_tumvi.py",
"content": "import argparse\nimport logging\nimport typing\n\nimport numpy as np\n\nimport evo.common_ape_rpe as common\nfrom evo.core import lie_algebra, sync, metrics\nfrom evo.core.result import Result\nfrom evo.core.trajectory import PosePath3D, PoseTrajectory3D\nfrom evo.tools import file_interface, log\nfrom evo.tools.settings import SETTINGS\n\nimport matplotlib.pyplot as plt\nimport matplotlib\nimport copy\nimport os\n\nmatplotlib.rcParams['mathtext.fontset'] = 'custom'\nmatplotlib.rcParams['mathtext.rm'] = 'Times New Roman'\nmatplotlib.rcParams['mathtext.it'] = 'Times New Roman:italic'\nmatplotlib.rcParams['mathtext.bf'] = 'Times New Roman:bold'\nmatplotlib.rcParams['font.family'] = 'Arial'\n\nfont0={'family':'Arial',\n 'style':'normal',\n 'weight':'bold',\n 'color':'black',\n 'size':6\n}\nfont1={'family':'Arial',\n 'style':'normal',\n 'weight':'bold',\n 'color':'black',\n 'size':8\n}\n\n\nlogger = logging.getLogger(__name__)\n\nSEP = \"-\" * 80 # separator line\n\ndef ape(traj_ref: PosePath3D, traj_est: PosePath3D,\n pose_relation: metrics.PoseRelation, align: bool = False,\n correct_scale: bool = False, n_to_align: int = -1,\n align_origin: bool = False, ref_name: str = \"reference\",\n est_name: str = \"estimate\",\n change_unit: typing.Optional[metrics.Unit] = None) -> Result:\n if n_to_align >0 : \n print('>>>>> only use the starting segment')\n n_to_align = np.where((np.array(traj_ref.timestamps)[1:]-np.array(traj_ref.timestamps)[0:-1])>100)[0][0]-1\n\n # Align the trajectories.\n only_scale = correct_scale and not align\n alignment_transformation = None\n if align or correct_scale:\n logger.debug(SEP)\n alignment_transformation = lie_algebra.sim3(\n *traj_est.align(traj_ref, correct_scale, only_scale, n=n_to_align))\n elif align_origin:\n logger.debug(SEP)\n alignment_transformation = traj_est.align_origin(traj_ref)\n\n # Calculate APE.\n logger.debug(SEP)\n data = (traj_ref, traj_est)\n ape_metric = metrics.APE(pose_relation)\n ape_metric.process_data(data)\n\n if change_unit:\n ape_metric.change_unit(change_unit)\n\n title = str(ape_metric)\n if align and not correct_scale:\n title += \"\\n(with SE(3) Umeyama alignment)\"\n elif align and correct_scale:\n title += \"\\n(with Sim(3) Umeyama alignment)\"\n elif only_scale:\n title += \"\\n(scale corrected)\"\n elif align_origin:\n title += \"\\n(with origin alignment)\"\n else:\n title += \"\\n(not aligned)\"\n if (align or correct_scale) and n_to_align != -1:\n title += \" (aligned poses: {})\".format(n_to_align)\n\n ape_result = ape_metric.get_result(ref_name, est_name)\n ape_result.info[\"title\"] = title\n\n logger.debug(SEP)\n logger.info(ape_result.pretty_str())\n\n ape_result.add_trajectory(ref_name, traj_ref)\n ape_result.add_trajectory(est_name, traj_est)\n if isinstance(traj_est, PoseTrajectory3D):\n seconds_from_start = np.array(\n [t - traj_est.timestamps[0] for t in traj_est.timestamps])\n ape_result.add_np_array(\"seconds_from_start\", seconds_from_start)\n ape_result.add_np_array(\"timestamps\", traj_est.timestamps)\n ape_result.add_np_array(\"distances_from_start\", traj_ref.distances)\n ape_result.add_np_array(\"distances\", traj_est.distances)\n\n if alignment_transformation is not None:\n ape_result.add_np_array(\"alignment_transformation_sim3\",\n alignment_transformation)\n\n return ape_result\n\n\nif __name__ == '__main__':\n color_list = [np.array([133,164,195])/255.0,[1,0.6,1],[1,0,0]]\n plt.figure('1',figsize=[6,6])\n parser = argparse.ArgumentParser()\n parser.add_argument('--seq', type=str, help='seq', default='outdoors1')\n parser.add_argument('--batch', action=\"store_true\")\n args = parser.parse_args()\n args.subcommand = 'tum'\n seq = args.seq\n\n args.ref_file = '/mnt/z/tum-vi/dataset-%s_512_16/dso/gt_imu.csv' % seq\n # Convert GT format\n dd = np.loadtxt(args.ref_file,delimiter=',',comments='#')\n dd_new = np.copy(dd)\n dd_new[:,0] /= 1e9\n dd_new[:,4] = dd[:,7] \n dd_new[:,5] = dd[:,4] \n dd_new[:,6] = dd[:,5] \n dd_new[:,7] = dd[:,6] \n args.ref_file = os.path.join(os.path.dirname(args.ref_file),'gt_imu_temp.txt')\n np.savetxt(args.ref_file, dd_new,delimiter=' ')\n\n args.pose_relation = 'trans_part'\n args.align = True\n args.correct_scale = False\n args.n_to_align = 1\n args.align_origin = False\n args.plot_mode = 'xyz'\n args.plot_x_dimension = 'seconds'\n args.plot_colormap_min = None\n args.plot_colormap_max = None\n args.plot_colormap_max_percentile = None\n args.ros_map_yaml = None\n args.plot = True\n args.est_files = ['results/result_%s.txt' % seq]\n label_list = ['DBA-Fusion (M)']\n args.save_plot = False\n args.serialize_plot = False\n for iii in range(len(args.est_files)):\n args.est_file = args.est_files[iii]\n\n if args.est_file.find('visual') != -1:\n args.correct_scale = True\n else:\n args.correct_scale = False\n\n traj_ref, traj_est, ref_name, est_name = common.load_trajectories(args)\n traj_ref_sel, traj_est_sel = sync.associate_trajectories(\n traj_ref, traj_est, 0.01,0.0,\n first_name=ref_name, snd_name=est_name)\n\n # use the starting part for scale estimation\n args.n_to_align = np.where((np.array(traj_ref_sel.timestamps)[1:]-np.array(traj_ref_sel.timestamps)[0:-1])>100)[0][0]-1\n pose_relation = common.get_pose_relation(args)\n result = ape(traj_ref=traj_ref_sel, traj_est=traj_est_sel,\n pose_relation=pose_relation, align=args.align,\n correct_scale=args.correct_scale, n_to_align=args.n_to_align,\n align_origin=args.align_origin, ref_name=ref_name,\n est_name=est_name)\n traj_est_sel = copy.deepcopy(result.trajectories[est_name])\n T01 = result.np_arrays['alignment_transformation_sim3']\n\n # metric-scale APE calculation\n result = ape(traj_ref=traj_ref_sel, traj_est=traj_est_sel,\n pose_relation=pose_relation, align=args.align,\n correct_scale=False, n_to_align=-1,\n align_origin=args.align_origin, ref_name=ref_name,\n est_name=est_name)\n print(result)\n\n if args.batch:\n quit()\n\n # visualization\n traj_est.transform(T01)\n \n if iii == 0:\n x_series=[]\n y_series=[]\n z_series=[]\n for i in range(len(traj_ref.poses_se3)):\n TTT = traj_ref.poses_se3[i]\n x_series.append(TTT[0,3])\n y_series.append(TTT[1,3])\n z_series.append(TTT[2,3])\n plt.plot(x_series,y_series,c=[0,0,0],linestyle = '--',linewidth=0.5)\n\n x_series=[]\n y_series=[]\n z_series=[]\n for i in range(len(traj_est.poses_se3)):\n TTT = traj_est.poses_se3[i]\n x_series.append(TTT[0,3])\n y_series.append(TTT[1,3])\n z_series.append(TTT[2,3])\n plt.plot(x_series,y_series,c=color_list[iii],label = label_list[iii],linewidth=0.5)\n\n ll = max(max(x_series)-min(x_series),max(y_series)-min(y_series))\n plt.xlim([(max(x_series)+min(x_series))/2 - 0.65*ll,(max(x_series)+min(x_series))/2+0.65*ll])\n plt.ylim([(max(y_series)+min(y_series))/2 - 0.65*ll,(max(y_series)+min(y_series))/2+0.65*ll])\n # plt.xlabel('X [m]')\n # plt.ylabel('Y [m]')\n plt.tick_params(labelsize=6,direction='in')\n lg = plt.legend(loc='upper right',markerscale=3,fontsize=5,framealpha=1,ncol=1,columnspacing=0.3,handletextpad=0.3,edgecolor='black',fancybox=False)\n lg.set_zorder(200)\n lg.get_frame().set_linewidth(0.8)\n plt.gca().yaxis.set_label_coords(-.1, .5)\n plt.show()\n "
},
{
"path": "results/PLACEHOLDER",
"content": ""
},
{
"path": "setup.py",
"content": "from setuptools import setup\nfrom torch.utils.cpp_extension import BuildExtension, CUDAExtension\n\nimport os.path as osp\nROOT = osp.dirname(osp.abspath(__file__))\n\nsetup(\n name='droid_backends',\n ext_modules=[\n CUDAExtension('droid_backends',\n include_dirs=[osp.join(ROOT, 'thirdparty/eigen')],\n sources=[\n 'src/droid.cpp', \n 'src/droid_kernels.cu',\n 'src/correlation_kernels.cu',\n 'src/altcorr_kernel.cu',\n ],\n extra_compile_args={\n 'cxx': ['-O3'],\n 'nvcc': ['-O3',\n '-gencode=arch=compute_60,code=sm_60',\n '-gencode=arch=compute_61,code=sm_61',\n '-gencode=arch=compute_70,code=sm_70',\n '-gencode=arch=compute_75,code=sm_75',\n '-gencode=arch=compute_80,code=sm_80',\n '-gencode=arch=compute_86,code=sm_86',\n ]\n }),\n ],\n cmdclass={ 'build_ext' : BuildExtension }\n)\n\nsetup(\n name='lietorch',\n version='0.2',\n description='Lie Groups for PyTorch',\n packages=['lietorch'],\n package_dir={'': 'thirdparty/lietorch'},\n ext_modules=[\n CUDAExtension('lietorch_backends', \n include_dirs=[\n osp.join(ROOT, 'thirdparty/lietorch/lietorch/include'), \n osp.join(ROOT, 'thirdparty/eigen')],\n sources=[\n 'thirdparty/lietorch/lietorch/src/lietorch.cpp', \n 'thirdparty/lietorch/lietorch/src/lietorch_gpu.cu',\n 'thirdparty/lietorch/lietorch/src/lietorch_cpu.cpp'],\n extra_compile_args={\n 'cxx': ['-O2'], \n 'nvcc': ['-O2',\n '-gencode=arch=compute_60,code=sm_60', \n '-gencode=arch=compute_61,code=sm_61', \n '-gencode=arch=compute_70,code=sm_70', \n '-gencode=arch=compute_75,code=sm_75',\n '-gencode=arch=compute_80,code=sm_80',\n '-gencode=arch=compute_86,code=sm_86', \n ]\n }),\n ],\n cmdclass={ 'build_ext' : BuildExtension }\n)\n"
},
{
"path": "src/altcorr_kernel.cu",
"content": "#include \n#include \n#include \n#include \n#include \n#include \n\n\n#include \n#include \n#include \n#include \n\n\n\n#define BLOCK_H 4\n#define BLOCK_W 8\n#define BLOCK_HW BLOCK_H * BLOCK_W\n#define CHANNEL_STRIDE 32\n\n\n__forceinline__ __device__\nbool within_bounds(int h, int w, int H, int W) {\n return h >= 0 && h < H && w >= 0 && w < W;\n}\n\ntemplate \n__global__ void altcorr_forward_kernel(\n const torch::PackedTensorAccessor32 fmap1,\n const torch::PackedTensorAccessor32 fmap2,\n const torch::PackedTensorAccessor32 coords,\n torch::PackedTensorAccessor32 corr,\n int r)\n{\n const int b = blockIdx.x;\n const int h0 = blockIdx.y * blockDim.x;\n const int w0 = blockIdx.z * blockDim.y;\n const int tid = threadIdx.x * blockDim.y + threadIdx.y;\n\n const int H1 = fmap1.size(1);\n const int W1 = fmap1.size(2);\n const int H2 = fmap2.size(1);\n const int W2 = fmap2.size(2);\n const int N = coords.size(1);\n const int C = fmap1.size(3);\n\n __shared__ scalar_t f1[CHANNEL_STRIDE][BLOCK_HW];\n __shared__ scalar_t f2[CHANNEL_STRIDE][BLOCK_HW];\n \n __shared__ float x2s[BLOCK_HW];\n __shared__ float y2s[BLOCK_HW];\n\n for (int c=0; c(floor(y2s[k1])) - r + iy;\n int w2 = static_cast(floor(x2s[k1])) - r + ix;\n int c2 = tid % CHANNEL_STRIDE;\n\n if (within_bounds(h2, w2, H2, W2))\n f2[c2][k1] = fmap2[b][h2][w2][c+c2];\n \n else\n f2[c2][k1] = static_cast(0.0);\n }\n\n __syncthreads();\n \n scalar_t s = 0.0;\n for (int k=0; k((dy) * (dx));\n scalar_t ne = s * static_cast((dy) * (1-dx));\n scalar_t sw = s * static_cast((1-dy) * (dx));\n scalar_t se = s * static_cast((1-dy) * (1-dx));\n\n // if (iy > 0 && ix > 0 && within_bounds(h1, w1, H1, W1))\n // corr[b][n][ix_nw][h1][w1] += nw;\n\n // if (iy > 0 && ix < rd && within_bounds(h1, w1, H1, W1))\n // corr[b][n][ix_ne][h1][w1] += ne;\n\n // if (iy < rd && ix > 0 && within_bounds(h1, w1, H1, W1))\n // corr[b][n][ix_sw][h1][w1] += sw;\n\n // if (iy < rd && ix < rd && within_bounds(h1, w1, H1, W1))\n // corr[b][n][ix_se][h1][w1] += se;\n\n\n scalar_t* corr_ptr = &corr[b][n][0][h1][w1];\n\n if (iy > 0 && ix > 0 && within_bounds(h1, w1, H1, W1))\n *(corr_ptr + ix_nw) += nw;\n\n if (iy > 0 && ix < rd && within_bounds(h1, w1, H1, W1))\n *(corr_ptr + ix_ne) += ne;\n\n if (iy < rd && ix > 0 && within_bounds(h1, w1, H1, W1))\n *(corr_ptr + ix_sw) += sw;\n\n if (iy < rd && ix < rd && within_bounds(h1, w1, H1, W1))\n *(corr_ptr + ix_se) += se;\n\n\n }\n } \n }\n }\n}\n\n\ntemplate \n__global__ void altcorr_backward_kernel(\n const torch::PackedTensorAccessor32 fmap1,\n const torch::PackedTensorAccessor32 fmap2,\n const torch::PackedTensorAccessor32 coords,\n const torch::PackedTensorAccessor32 corr_grad,\n torch::PackedTensorAccessor32 fmap1_grad,\n torch::PackedTensorAccessor32 fmap2_grad,\n torch::PackedTensorAccessor32 coords_grad,\n int r)\n{\n\n const int b = blockIdx.x;\n const int h0 = blockIdx.y * blockDim.x;\n const int w0 = blockIdx.z * blockDim.y;\n const int tid = threadIdx.x * blockDim.y + threadIdx.y;\n\n const int H1 = fmap1.size(1);\n const int W1 = fmap1.size(2);\n const int H2 = fmap2.size(1);\n const int W2 = fmap2.size(2);\n const int N = coords.size(1);\n const int C = fmap1.size(3);\n\n __shared__ scalar_t f1[CHANNEL_STRIDE][BLOCK_HW+1];\n __shared__ scalar_t f2[CHANNEL_STRIDE][BLOCK_HW+1];\n\n __shared__ scalar_t f1_grad[CHANNEL_STRIDE][BLOCK_HW+1];\n __shared__ scalar_t f2_grad[CHANNEL_STRIDE][BLOCK_HW+1];\n\n __shared__ scalar_t x2s[BLOCK_HW];\n __shared__ scalar_t y2s[BLOCK_HW];\n\n for (int c=0; c(floor(y2s[k1]))-r+iy;\n int w2 = static_cast(floor(x2s[k1]))-r+ix;\n int c2 = tid % CHANNEL_STRIDE;\n\n auto fptr = fmap2[b][h2][w2];\n if (within_bounds(h2, w2, H2, W2))\n f2[c2][k1] = fptr[c+c2];\n else\n f2[c2][k1] = 0.0;\n\n f2_grad[c2][k1] = 0.0;\n }\n\n __syncthreads();\n \n const scalar_t* grad_ptr = &corr_grad[b][n][0][h1][w1];\n scalar_t g = 0.0;\n\n int ix_nw = H1*W1*((iy-1) + rd*(ix-1));\n int ix_ne = H1*W1*((iy-1) + rd*ix);\n int ix_sw = H1*W1*(iy + rd*(ix-1));\n int ix_se = H1*W1*(iy + rd*ix);\n\n if (iy > 0 && ix > 0 && within_bounds(h1, w1, H1, W1))\n g += *(grad_ptr + ix_nw) * dy * dx;\n\n if (iy > 0 && ix < rd && within_bounds(h1, w1, H1, W1))\n g += *(grad_ptr + ix_ne) * dy * (1-dx);\n\n if (iy < rd && ix > 0 && within_bounds(h1, w1, H1, W1))\n g += *(grad_ptr + ix_sw) * (1-dy) * dx;\n\n if (iy < rd && ix < rd && within_bounds(h1, w1, H1, W1))\n g += *(grad_ptr + ix_se) * (1-dy) * (1-dx);\n \n for (int k=0; k(floor(y2s[k1]))-r+iy;\n int w2 = static_cast(floor(x2s[k1]))-r+ix;\n int c2 = tid % CHANNEL_STRIDE;\n\n scalar_t* fptr = &fmap2_grad[b][h2][w2][0];\n if (within_bounds(h2, w2, H2, W2))\n atomicAdd(fptr+c+c2, f2_grad[c2][k1]);\n }\n }\n } \n }\n __syncthreads();\n\n\n for (int k=0; k altcorr_cuda_forward(\n torch::Tensor fmap1,\n torch::Tensor fmap2,\n torch::Tensor coords,\n int radius)\n{\n const auto B = coords.size(0);\n const auto N = coords.size(1);\n const auto H = coords.size(2);\n const auto W = coords.size(3);\n\n const auto rd = 2 * radius + 1;\n auto opts = fmap1.options();\n auto corr = torch::zeros({B, N, rd*rd, H, W}, opts);\n \n const dim3 blocks(B, (H+BLOCK_H-1)/BLOCK_H, (W+BLOCK_W-1)/BLOCK_W);\n const dim3 threads(BLOCK_H, BLOCK_W);\n\n\n AT_DISPATCH_FLOATING_TYPES_AND_HALF(fmap1.type(), \"altcorr_forward_kernel\", ([&] {\n altcorr_forward_kernel<<>>(\n fmap1.packed_accessor32(),\n fmap2.packed_accessor32(),\n coords.packed_accessor32(),\n corr.packed_accessor32(),\n radius);\n }));\n\n return {corr};\n}\n\nstd::vector altcorr_cuda_backward(\n torch::Tensor fmap1,\n torch::Tensor fmap2,\n torch::Tensor coords,\n torch::Tensor corr_grad,\n int radius)\n{\n const auto B = coords.size(0);\n const auto N = coords.size(1);\n\n const auto H1 = fmap1.size(1);\n const auto W1 = fmap1.size(2);\n const auto H2 = fmap2.size(1);\n const auto W2 = fmap2.size(2);\n const auto C = fmap1.size(3);\n\n auto opts = fmap1.options();\n auto fmap1_grad = torch::zeros({B, H1, W1, C}, opts);\n auto fmap2_grad = torch::zeros({B, H2, W2, C}, opts);\n auto coords_grad = torch::zeros({B, N, H1, W1, 2}, opts);\n \n const dim3 blocks(B, (H1+BLOCK_H-1)/BLOCK_H, (W1+BLOCK_W-1)/BLOCK_W);\n const dim3 threads(BLOCK_H, BLOCK_W);\n\n altcorr_backward_kernel<<>>(\n fmap1.packed_accessor32(),\n fmap2.packed_accessor32(),\n coords.packed_accessor32(),\n corr_grad.packed_accessor32(),\n fmap1_grad.packed_accessor32(),\n fmap2_grad.packed_accessor32(),\n coords_grad.packed_accessor32(),\n radius);\n\n return {fmap1_grad, fmap2_grad, coords_grad};\n}"
},
{
"path": "src/bacore.h",
"content": "#include \n#include \n\nclass BACore\n{\npublic:\n BACore(){}\n ~BACore(){}\npublic:\n void init(torch::Tensor _poses,\n torch::Tensor _disps,\n torch::Tensor _intrinsics,\n torch::Tensor _disps_sens,\n torch::Tensor _targets,\n torch::Tensor _weights,\n torch::Tensor _eta,\n torch::Tensor _ii,\n torch::Tensor _jj,\n const int t0,\n const int t1,\n const int iterations,\n const float lm,\n const float ep,\n const bool motion_only);\n void hessian(torch::Tensor H, torch::Tensor v);\n void optimize(torch::Tensor H, torch::Tensor v);\n std::vector retract(torch::Tensor _dx);\n \npublic:\n torch::Tensor poses;\n torch::Tensor disps;\n torch::Tensor intrinsics;\n torch::Tensor disps_sens;\n torch::Tensor targets;\n torch::Tensor weights;\n torch::Tensor eta;\n torch::Tensor ii;\n torch::Tensor jj;\n int t0,t1;\n float lm, ep;\n\n torch::Tensor ts;\n torch::Tensor ii_exp;\n torch::Tensor jj_exp;\n\n std::tuple kuniq;\n\n torch::Tensor kx;\n torch::Tensor kk_exp; // 不重复元素的索引\n \n torch::Tensor dx;\n torch::Tensor dz;\n\n // initialize buffers\n torch::Tensor Hs; \n torch::Tensor vs; \n torch::Tensor Eii; \n torch::Tensor Eij; \n torch::Tensor Cii; \n torch::Tensor wi; \n\n torch::Tensor m ;\n torch::Tensor C ;\n torch::Tensor w ;\n torch::Tensor Q ;\n torch::Tensor Ei;\n torch::Tensor E ;\n\n\n};\n"
},
{
"path": "src/correlation_kernels.cu",
"content": "#include \n#include \n#include \n#include \n#include \n#include \n\n\n#include \n#include \n#include \n\n#define BLOCK 16\n\n__forceinline__ __device__ bool within_bounds(int h, int w, int H, int W) {\n return h >= 0 && h < H && w >= 0 && w < W;\n}\n\ntemplate \n__global__ void corr_index_forward_kernel(\n const torch::PackedTensorAccessor32 volume,\n const torch::PackedTensorAccessor32 coords,\n torch::PackedTensorAccessor32 corr,\n int r)\n{\n // batch index\n const int x = blockIdx.x * blockDim.x + threadIdx.x;\n const int y = blockIdx.y * blockDim.y + threadIdx.y;\n const int n = blockIdx.z;\n\n const int h1 = volume.size(1);\n const int w1 = volume.size(2);\n const int h2 = volume.size(3);\n const int w2 = volume.size(4);\n\n if (!within_bounds(y, x, h1, w1)) {\n return;\n }\n\n float x0 = coords[n][0][y][x];\n float y0 = coords[n][1][y][x];\n\n float dx = x0 - floor(x0);\n float dy = y0 - floor(y0);\n\n int rd = 2*r + 1;\n for (int i=0; i(floor(x0)) - r + i;\n int y1 = static_cast(floor(y0)) - r + j;\n\n if (within_bounds(y1, x1, h2, w2)) {\n scalar_t s = volume[n][y][x][y1][x1];\n\n if (i > 0 && j > 0)\n corr[n][i-1][j-1][y][x] += s * scalar_t(dx * dy);\n\n if (i > 0 && j < rd)\n corr[n][i-1][j][y][x] += s * scalar_t(dx * (1.0f-dy));\n\n if (i < rd && j > 0)\n corr[n][i][j-1][y][x] += s * scalar_t((1.0f-dx) * dy);\n\n if (i < rd && j < rd)\n corr[n][i][j][y][x] += s * scalar_t((1.0f-dx) * (1.0f-dy));\n\n }\n }\n }\n}\n\n\ntemplate \n__global__ void corr_index_backward_kernel(\n const torch::PackedTensorAccessor32 coords,\n const torch::PackedTensorAccessor32 corr_grad,\n torch::PackedTensorAccessor32 volume_grad,\n int r)\n{\n // batch index\n const int x = blockIdx.x * blockDim.x + threadIdx.x;\n const int y = blockIdx.y * blockDim.y + threadIdx.y;\n const int n = blockIdx.z;\n\n const int h1 = volume_grad.size(1);\n const int w1 = volume_grad.size(2);\n const int h2 = volume_grad.size(3);\n const int w2 = volume_grad.size(4);\n\n if (!within_bounds(y, x, h1, w1)) {\n return;\n }\n\n float x0 = coords[n][0][y][x];\n float y0 = coords[n][1][y][x];\n\n float dx = x0 - floor(x0);\n float dy = y0 - floor(y0);\n\n int rd = 2*r + 1;\n for (int i=0; i(floor(x0)) - r + i;\n int y1 = static_cast(floor(y0)) - r + j;\n\n if (within_bounds(y1, x1, h2, w2)) {\n scalar_t g = 0.0;\n if (i > 0 && j > 0)\n g += corr_grad[n][i-1][j-1][y][x] * scalar_t(dx * dy);\n\n if (i > 0 && j < rd)\n g += corr_grad[n][i-1][j][y][x] * scalar_t(dx * (1.0f-dy));\n\n if (i < rd && j > 0)\n g += corr_grad[n][i][j-1][y][x] * scalar_t((1.0f-dx) * dy);\n\n if (i < rd && j < rd)\n g += corr_grad[n][i][j][y][x] * scalar_t((1.0f-dx) * (1.0f-dy));\n\n volume_grad[n][y][x][y1][x1] += g;\n }\n }\n }\n}\n\nstd::vector corr_index_cuda_forward(\n torch::Tensor volume,\n torch::Tensor coords,\n int radius)\n{\n const auto batch_size = volume.size(0);\n const auto ht = volume.size(1);\n const auto wd = volume.size(2);\n\n const dim3 blocks((wd + BLOCK - 1) / BLOCK, \n (ht + BLOCK - 1) / BLOCK, \n batch_size);\n \n const dim3 threads(BLOCK, BLOCK);\n\n auto opts = volume.options();\n torch::Tensor corr = torch::zeros(\n {batch_size, 2*radius+1, 2*radius+1, ht, wd}, opts);\n\n AT_DISPATCH_FLOATING_TYPES_AND_HALF(volume.type(), \"sampler_forward_kernel\", ([&] {\n corr_index_forward_kernel<<>>(\n volume.packed_accessor32(),\n coords.packed_accessor32(),\n corr.packed_accessor32(),\n radius);\n }));\n\n return {corr};\n\n}\n\nstd::vector corr_index_cuda_backward(\n torch::Tensor volume,\n torch::Tensor coords,\n torch::Tensor corr_grad,\n int radius)\n{\n const auto batch_size = volume.size(0);\n const auto ht = volume.size(1);\n const auto wd = volume.size(2);\n\n auto volume_grad = torch::zeros_like(volume);\n\n const dim3 blocks((wd + BLOCK - 1) / BLOCK, \n (ht + BLOCK - 1) / BLOCK, \n batch_size);\n\n const dim3 threads(BLOCK, BLOCK);\n\n\n AT_DISPATCH_FLOATING_TYPES_AND_HALF(volume.type(), \"sampler_backward_kernel\", ([&] {\n corr_index_backward_kernel<<>>(\n coords.packed_accessor32(),\n corr_grad.packed_accessor32(),\n volume_grad.packed_accessor32(),\n radius);\n }));\n\n return {volume_grad};\n}"
},
{
"path": "src/droid.cpp",
"content": "#include \n#include \n\n#include \"bacore.h\"\n\n// CUDA forward declarations\nstd::vector projective_transform_cuda(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor ii,\n torch::Tensor jj);\n\ntorch::Tensor depth_filter_cuda(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor ix,\n torch::Tensor thresh);\n\n\ntorch::Tensor frame_distance_cuda(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor ii,\n torch::Tensor jj,\n const float beta);\n\nstd::vector projmap_cuda(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor ii,\n torch::Tensor jj);\n\ntorch::Tensor iproj_cuda(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics);\n\nstd::vector ba_cuda(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor disps_sens,\n torch::Tensor targets,\n torch::Tensor weights,\n torch::Tensor eta,\n torch::Tensor ii,\n torch::Tensor jj,\n const int t0,\n const int t1,\n const int iterations,\n const float lm,\n const float ep,\n const bool motion_only);\n\nstd::vector ba_cuda_extend(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor disps_sens,\n torch::Tensor targets,\n torch::Tensor weights,\n torch::Tensor eta,\n torch::Tensor ii,\n torch::Tensor jj,\n torch::Tensor H,\n torch::Tensor v,\n torch::Tensor A_prior,\n const int t0,\n const int t1,\n const int iterations,\n const float lm,\n const float ep,\n const bool motion_only,\n const bool skip_solve);\n\nstd::vector corr_index_cuda_forward(\n torch::Tensor volume,\n torch::Tensor coords,\n int radius);\n\nstd::vector corr_index_cuda_backward(\n torch::Tensor volume,\n torch::Tensor coords,\n torch::Tensor corr_grad,\n int radius);\n\nstd::vector altcorr_cuda_forward(\n torch::Tensor fmap1,\n torch::Tensor fmap2,\n torch::Tensor coords,\n int radius);\n\nstd::vector altcorr_cuda_backward(\n torch::Tensor fmap1,\n torch::Tensor fmap2,\n torch::Tensor coords,\n torch::Tensor corr_grad,\n int radius);\n\n\n#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x \" must be contiguous\")\n#define CHECK_INPUT(x) CHECK_CONTIGUOUS(x)\n\n\nstd::vector ba(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor disps_sens,\n torch::Tensor targets,\n torch::Tensor weights,\n torch::Tensor eta,\n torch::Tensor ii,\n torch::Tensor jj,\n const int t0,\n const int t1,\n const int iterations,\n const float lm,\n const float ep,\n const bool motion_only) {\n\n CHECK_INPUT(targets);\n CHECK_INPUT(weights);\n CHECK_INPUT(poses);\n CHECK_INPUT(disps);\n CHECK_INPUT(intrinsics);\n CHECK_INPUT(disps_sens);\n CHECK_INPUT(ii);\n CHECK_INPUT(jj);\n\n return ba_cuda(poses, disps, intrinsics, disps_sens, targets, weights,\n eta, ii, jj, t0, t1, iterations, lm, ep, motion_only);\n\n}\n\nstd::vector ba_extend(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor disps_sens,\n torch::Tensor targets,\n torch::Tensor weights,\n torch::Tensor eta,\n torch::Tensor ii,\n torch::Tensor jj,\n torch::Tensor H,\n torch::Tensor v,\n torch::Tensor A_prior,\n const int t0,\n const int t1,\n const int iterations,\n const float lm,\n const float ep,\n const bool motion_only,\n const bool skip_solve) {\n\n CHECK_INPUT(targets);\n CHECK_INPUT(weights);\n CHECK_INPUT(poses);\n CHECK_INPUT(disps);\n CHECK_INPUT(intrinsics);\n CHECK_INPUT(disps_sens);\n CHECK_INPUT(ii);\n CHECK_INPUT(jj);\n\n CHECK_INPUT(H);\n CHECK_INPUT(v);\n CHECK_INPUT(A_prior);\n\n std::vector dx_dz = ba_cuda_extend(poses, disps, intrinsics, disps_sens, targets, weights,\n eta, ii, jj, H, v, A_prior, t0, t1, iterations, lm, ep, motion_only, skip_solve);\n return dx_dz;\n\n}\n\n\ntorch::Tensor frame_distance(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor ii,\n torch::Tensor jj,\n const float beta) {\n\n CHECK_INPUT(poses);\n CHECK_INPUT(disps);\n CHECK_INPUT(intrinsics);\n CHECK_INPUT(ii);\n CHECK_INPUT(jj);\n\n return frame_distance_cuda(poses, disps, intrinsics, ii, jj, beta);\n\n}\n\n\nstd::vector projmap(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor ii,\n torch::Tensor jj) {\n\n CHECK_INPUT(poses);\n CHECK_INPUT(disps);\n CHECK_INPUT(intrinsics);\n CHECK_INPUT(ii);\n CHECK_INPUT(jj);\n\n return projmap_cuda(poses, disps, intrinsics, ii, jj);\n\n}\n\n\ntorch::Tensor iproj(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics) {\n CHECK_INPUT(poses);\n CHECK_INPUT(disps);\n CHECK_INPUT(intrinsics);\n\n return iproj_cuda(poses, disps, intrinsics);\n}\n\n\n// c++ python binding\nstd::vector corr_index_forward(\n torch::Tensor volume,\n torch::Tensor coords,\n int radius) {\n CHECK_INPUT(volume);\n CHECK_INPUT(coords);\n\n return corr_index_cuda_forward(volume, coords, radius);\n}\n\nstd::vector corr_index_backward(\n torch::Tensor volume,\n torch::Tensor coords,\n torch::Tensor corr_grad,\n int radius) {\n CHECK_INPUT(volume);\n CHECK_INPUT(coords);\n CHECK_INPUT(corr_grad);\n\n auto volume_grad = corr_index_cuda_backward(volume, coords, corr_grad, radius);\n return {volume_grad};\n}\n\nstd::vector altcorr_forward(\n torch::Tensor fmap1,\n torch::Tensor fmap2,\n torch::Tensor coords,\n int radius) {\n CHECK_INPUT(fmap1);\n CHECK_INPUT(fmap2);\n CHECK_INPUT(coords);\n\n return altcorr_cuda_forward(fmap1, fmap2, coords, radius);\n}\n\nstd::vector altcorr_backward(\n torch::Tensor fmap1,\n torch::Tensor fmap2,\n torch::Tensor coords,\n torch::Tensor corr_grad,\n int radius) {\n CHECK_INPUT(fmap1);\n CHECK_INPUT(fmap2);\n CHECK_INPUT(coords);\n CHECK_INPUT(corr_grad);\n\n return altcorr_cuda_backward(fmap1, fmap2, coords, corr_grad, radius);\n}\n\n\ntorch::Tensor depth_filter(\n torch::Tensor poses,\n torch::Tensor disps,\n torch::Tensor intrinsics,\n torch::Tensor ix,\n torch::Tensor thresh) {\n\n CHECK_INPUT(poses);\n CHECK_INPUT(disps);\n CHECK_INPUT(intrinsics);\n CHECK_INPUT(ix);\n CHECK_INPUT(thresh);\n\n return depth_filter_cuda(poses, disps, intrinsics, ix, thresh);\n}\n\nPYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {\n // bundle adjustment kernels\n m.def(\"ba\", &ba, \"bundle adjustment\");\n m.def(\"ba_extend\", &ba_extend, \"bundle adjustment (extended)\");\n m.def(\"frame_distance\", &frame_distance, \"frame_distance\");\n m.def(\"projmap\", &projmap, \"projmap\");\n m.def(\"depth_filter\", &depth_filter, \"depth_filter\");\n m.def(\"iproj\", &iproj, \"back projection\");\n\n // correlation volume kernels\n m.def(\"altcorr_forward\", &altcorr_forward, \"ALTCORR forward\");\n m.def(\"altcorr_backward\", &altcorr_backward, \"ALTCORR backward\");\n m.def(\"corr_index_forward\", &corr_index_forward, \"INDEX forward\");\n m.def(\"corr_index_backward\", &corr_index_backward, \"INDEX backward\");\n py::class_(m, \"BACore\").def(py::init<>())\n .def(\"init\", &BACore::init)\n .def(\"hessian\", &BACore::hessian)\n .def(\"optimize\", &BACore::optimize)\n .def(\"retract\", &BACore::retract);\n}"
},
{
"path": "src/droid_kernels.cu",
"content": "#include \n#include \n#include \n#include \n\n#include \n#include \n\n#include \n#include \n#include \n\n// #include \"utils.cuh\"\n\n#include \n#include \n#include \n#include \n\n#include \"bacore.h\"\n\ntypedef Eigen::SparseMatrix SpMat;\ntypedef Eigen::Triplet T;\ntypedef std::vector> graph_t;\ntypedef std::vector tensor_list_t;\n\n\n\n#define MIN_DEPTH 0.25\n\n#define THREADS 256\n#define NUM_BLOCKS(batch_size) ((batch_size + THREADS - 1) / THREADS)\n\n\n#define GPU_1D_KERNEL_LOOP(k, n) \\\n for (size_t k = threadIdx.x; k 1e-4) {\n float a = (1 - cosf(theta)) / theta_sq;\n crossInplace(phi, tau);\n t[0] += a * tau[0];\n t[1] += a * tau[1];\n t[2] += a * tau[2];\n\n float b = (theta - sinf(theta)) / (theta * theta_sq);\n crossInplace(phi, tau);\n t[0] += b * tau[0];\n t[1] += b * tau[1];\n t[2] += b * tau[2];\n }\n}\n\ntorch::Tensor solveDense(const Eigen::MatrixXd& A, const Eigen::MatrixXd& Aprior, const Eigen::VectorXd& b, \nconst int N, const int M,\nconst float lm=0.0001,const float ep = 0.1){\n torch::Tensor dx;\n Eigen::MatrixXd L = A + Aprior;\n L.diagonal().array() += ep + lm * L.diagonal().array();\n Eigen::LLT solver;\n solver.compute(L);\n if (solver.info() == Eigen::Success) {\n Eigen::VectorXd x = solver.solve(b);\n dx = torch::from_blob(x.data(), {N, M}, torch::TensorOptions()\n .dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);\n }\n else {\n dx = torch::zeros({N, M}, torch::TensorOptions()\n .device(torch::kCUDA).dtype(torch::kFloat32));\n }\n return dx;\n}\n\ntorch::Tensor solveDenseD(const Eigen::MatrixXd& A, const Eigen::VectorXd& b, \nconst int N, const int M,\nconst float lm=0.0001,const float ep = 0.1){\n torch::Tensor dx;\n Eigen::MatrixXd L(A);\n L.diagonal().array() += ep + lm * L.diagonal().array();\n Eigen::LLT solver;\n solver.compute(L);\n if (solver.info() == Eigen::Success) {\n Eigen::VectorXd x = solver.solve(b);\n dx = torch::from_blob(x.data(), {N, M}, torch::TensorOptions()\n .dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);\n }\n else {\n dx = torch::zeros({N, M}, torch::TensorOptions()\n .device(torch::kCUDA).dtype(torch::kFloat32));\n }\n return dx;\n}\n\n__global__ void projective_transform_kernel(\n const torch::PackedTensorAccessor32 target,\n const torch::PackedTensorAccessor32 weight,\n const torch::PackedTensorAccessor32 poses,\n const torch::PackedTensorAccessor32 disps,\n const torch::PackedTensorAccessor32 intrinsics,\n const torch::PackedTensorAccessor32 ii,\n const torch::PackedTensorAccessor32 jj,\n torch::PackedTensorAccessor32 Hs, // Hessian of Poses\n torch::PackedTensorAccessor32 vs, // Residuals (Pose)\n torch::PackedTensorAccessor32 Eii,// Hessian, Disps_i - Pose_i\n torch::PackedTensorAccessor32 Eij,// Hessian, Disps_i - Pose_j\n torch::PackedTensorAccessor32 Cii,// Hessian, Disps_i - Disps_i\n torch::PackedTensorAccessor32 bz) // Residuals (Disp)\n{\n const int block_id = blockIdx.x;\n const int thread_id = threadIdx.x;\n\n const int ht = disps.size(1);\n const int wd = disps.size(2);\n\n int ix = static_cast(ii[block_id]);\n int jx = static_cast(jj[block_id]);\n\n __shared__ float fx;\n __shared__ float fy;\n __shared__ float cx;\n __shared__ float cy;\n\n __shared__ float ti[3], tj[3], tij[3];\n __shared__ float qi[4], qj[4], qij[4];\n\n // load intrinsics from global memory\n if (thread_id == 0) {\n fx = intrinsics[0];\n fy = intrinsics[1];\n cx = intrinsics[2];\n cy = intrinsics[3];\n }\n\n __syncthreads();\n\n // stereo frames\n if (ix == jx) {\n if (thread_id == 0) {\n tij[0] = -0.1;\n tij[1] = 0;\n tij[2] = 0;\n qij[0] = 0;\n qij[1] = 0;\n qij[2] = 0;\n qij[3] = 1;\n }\n }\n\n else {\n\n // load poses from global memory\n if (thread_id < 3) {\n ti[thread_id] = poses[ix][thread_id];\n tj[thread_id] = poses[jx][thread_id];\n }\n\n if (thread_id < 4) {\n qi[thread_id] = poses[ix][thread_id+3];\n qj[thread_id] = poses[jx][thread_id+3];\n }\n\n __syncthreads();\n\n if (thread_id == 0) {\n relSE3(ti, qi, tj, qj, tij, qij);\n }\n }\n\n __syncthreads();\n\n //points \n float Xi[4];\n float Xj[4];\n\n // jacobians\n float Jx[12];\n float Jz;\n\n float* Ji = &Jx[0];\n float* Jj = &Jx[6];\n\n // hessians\n float hij[12*(12+1)/2];\n\n float vi[6], vj[6];\n\n int l;\n for (l=0; l<12*(12+1)/2; l++) {\n hij[l] = 0;\n }\n\n for (int n=0; n<6; n++) {\n vi[n] = 0;\n vj[n] = 0;\n }\n\n __syncthreads();\n\n GPU_1D_KERNEL_LOOP(k, ht*wd) {\n\n const int i = k / wd;\n const int j = k % wd;\n\n const float u = static_cast(j);\n const float v = static_cast(i);\n \n // homogenous coordinates\n Xi[0] = (u - cx) / fx;\n Xi[1] = (v - cy) / fy;\n Xi[2] = 1;\n Xi[3] = disps[ix][i][j];\n\n // transform homogenous point\n actSE3(tij, qij, Xi, Xj);\n\n const float x = Xj[0];\n const float y = Xj[1];\n const float h = Xj[3];\n\n const float d = (Xj[2] < MIN_DEPTH) ? 0.0 : 1.0 / Xj[2];\n const float d2 = d * d;\n\n float wu = (Xj[2] < MIN_DEPTH) ? 0.0 : .001 * weight[block_id][0][i][j];\n float wv = (Xj[2] < MIN_DEPTH) ? 0.0 : .001 * weight[block_id][1][i][j];\n const float ru = target[block_id][0][i][j] - (fx * d * x + cx);\n const float rv = target[block_id][1][i][j] - (fy * d * y + cy);\n\n // x - coordinate\n\n Jj[0] = fx * (h*d);\n Jj[1] = fx * 0;\n Jj[2] = fx * (-x*h*d2);\n Jj[3] = fx * (-x*y*d2);\n Jj[4] = fx * (1 + x*x*d2);\n Jj[5] = fx * (-y*d);\n\n Jz = fx * (tij[0] * d - tij[2] * (x * d2));\n Cii[block_id][k] = wu * Jz * Jz;\n bz[block_id][k] = wu * ru * Jz;\n\n if (ix == jx) wu = 0;\n\n adjSE3(tij, qij, Jj, Ji);\n for (int n=0; n<6; n++) Ji[n] *= -1;\n\n l=0;\n for (int n=0; n<12; n++) {\n for (int m=0; m<=n; m++) {\n hij[l] += wu * Jx[n] * Jx[m];\n l++;\n }\n }\n\n for (int n=0; n<6; n++) {\n vi[n] += wu * ru * Ji[n];\n vj[n] += wu * ru * Jj[n];\n\n Eii[block_id][n][k] = wu * Jz * Ji[n];\n Eij[block_id][n][k] = wu * Jz * Jj[n];\n }\n\n\n Jj[0] = fy * 0;\n Jj[1] = fy * (h*d);\n Jj[2] = fy * (-y*h*d2);\n Jj[3] = fy * (-1 - y*y*d2);\n Jj[4] = fy * (x*y*d2);\n Jj[5] = fy * (x*d);\n\n Jz = fy * (tij[1] * d - tij[2] * (y * d2));\n Cii[block_id][k] += wv * Jz * Jz;\n bz[block_id][k] += wv * rv * Jz;\n\n if (ix == jx) wv = 0;\n\n adjSE3(tij, qij, Jj, Ji);\n for (int n=0; n<6; n++) Ji[n] *= -1;\n\n l=0;\n for (int n=0; n<12; n++) {\n for (int m=0; m<=n; m++) {\n hij[l] += wv * Jx[n] * Jx[m];\n l++;\n }\n }\n\n for (int n=0; n<6; n++) {\n vi[n] += wv * rv * Ji[n];\n vj[n] += wv * rv * Jj[n];\n\n Eii[block_id][n][k] += wv * Jz * Ji[n];\n Eij[block_id][n][k] += wv * Jz * Jj[n];\n }\n\n\n }\n\n __syncthreads();\n\n __shared__ float sdata[THREADS];\n for (int n=0; n<6; n++) {\n sdata[threadIdx.x] = vi[n];\n blockReduce(sdata);\n if (threadIdx.x == 0) {\n vs[0][block_id][n] = sdata[0];\n }\n\n __syncthreads();\n\n sdata[threadIdx.x] = vj[n];\n blockReduce(sdata);\n if (threadIdx.x == 0) {\n vs[1][block_id][n] = sdata[0];\n }\n\n }\n\n l=0;\n for (int n=0; n<12; n++) {\n for (int m=0; m<=n; m++) {\n sdata[threadIdx.x] = hij[l];\n blockReduce(sdata);\n\n if (threadIdx.x == 0) {\n if (n<6 && m<6) {\n Hs[0][block_id][n][m] = sdata[0];\n Hs[0][block_id][m][n] = sdata[0];\n }\n else if (n >=6 && m<6) {\n Hs[1][block_id][m][n-6] = sdata[0];\n Hs[2][block_id][n-6][m] = sdata[0];\n }\n else {\n Hs[3][block_id][n-6][m-6] = sdata[0];\n Hs[3][block_id][m-6][n-6] = sdata[0];\n }\n }\n\n l++;\n }\n }\n}\n\n\n__global__ void projmap_kernel(\n const torch::PackedTensorAccessor32 poses,\n const torch::PackedTensorAccessor32 disps,\n const torch::PackedTensorAccessor32 intrinsics,\n const torch::PackedTensorAccessor32 ii,\n const torch::PackedTensorAccessor32 jj,\n torch::PackedTensorAccessor32 coords,\n torch::PackedTensorAccessor32 valid)\n{\n\n const int block_id = blockIdx.x;\n const int thread_id = threadIdx.x;\n\n const int ht = disps.size(1);\n const int wd = disps.size(2);\n\n __shared__ int ix;\n __shared__ int jx;\n\n __shared__ float fx;\n __shared__ float fy;\n __shared__ float cx;\n __shared__ float cy;\n\n __shared__ float ti[3], tj[3], tij[3];\n __shared__ float qi[4], qj[4], qij[4];\n\n // load intrinsics from global memory\n if (thread_id == 0) {\n ix = static_cast(ii[block_id]);\n jx = static_cast(jj[block_id]);\n fx = intrinsics[0];\n fy = intrinsics[1];\n cx = intrinsics[2];\n cy = intrinsics[3];\n }\n\n __syncthreads();\n\n // load poses from global memory\n if (thread_id < 3) {\n ti[thread_id] = poses[ix][thread_id];\n tj[thread_id] = poses[jx][thread_id];\n }\n\n if (thread_id < 4) {\n qi[thread_id] = poses[ix][thread_id+3];\n qj[thread_id] = poses[jx][thread_id+3];\n }\n\n __syncthreads();\n\n if (thread_id == 0) {\n relSE3(ti, qi, tj, qj, tij, qij);\n }\n\n //points \n float Xi[4];\n float Xj[4];\n\n __syncthreads();\n\n GPU_1D_KERNEL_LOOP(k, ht*wd) {\n const int i = k / wd;\n const int j = k % wd;\n\n const float u = static_cast(j);\n const float v = static_cast(i);\n \n // homogenous coordinates\n Xi[0] = (u - cx) / fx;\n Xi[1] = (v - cy) / fy;\n Xi[2] = 1;\n Xi[3] = disps[ix][i][j];\n\n // transform homogenous point\n actSE3(tij, qij, Xi, Xj);\n\n coords[block_id][i][j][0] = u;\n coords[block_id][i][j][1] = v;\n\n if (Xj[2] > 0.01) {\n coords[block_id][i][j][0] = fx * (Xj[0] / Xj[2]) + cx;\n coords[block_id][i][j][1] = fy * (Xj[1] / Xj[2]) + cy;\n }\n\n valid[block_id][i][j][0] = (Xj[2] > MIN_DEPTH) ? 1.0 : 0.0;\n\n }\n}\n\n__global__ void frame_distance_kernel(\n const torch::PackedTensorAccessor32 poses,\n const torch::PackedTensorAccessor32 disps,\n const torch::PackedTensorAccessor32 intrinsics,\n const torch::PackedTensorAccessor32 ii,\n const torch::PackedTensorAccessor32 jj,\n torch::PackedTensorAccessor32 dist,\n const float beta) {\n\n const int block_id = blockIdx.x;\n const int thread_id = threadIdx.x;\n\n const int ht = disps.size(1);\n const int wd = disps.size(2);\n\n __shared__ int ix;\n __shared__ int jx;\n\n __shared__ float fx;\n __shared__ float fy;\n __shared__ float cx;\n __shared__ float cy;\n\n __shared__ float ti[3], tj[3], tij[3];\n __shared__ float qi[4], qj[4], qij[4];\n\n // load intrinsics from global memory\n if (thread_id == 0) {\n ix = static_cast(ii[block_id]);\n jx = static_cast(jj[block_id]);\n fx = intrinsics[0];\n fy = intrinsics[1];\n cx = intrinsics[2];\n cy = intrinsics[3];\n }\n\n __syncthreads();\n\n\n //points \n float Xi[4];\n float Xj[4];\n\n __shared__ float accum[THREADS]; accum[thread_id] = 0;\n __shared__ float valid[THREADS]; valid[thread_id] = 0;\n __shared__ float total[THREADS]; total[thread_id] = 0;\n\n __syncthreads();\n\n for (int n=0; n<1; n++) {\n\n if (thread_id < 3) {\n ti[thread_id] = poses[ix][thread_id];\n tj[thread_id] = poses[jx][thread_id];\n }\n\n if (thread_id < 4) {\n qi[thread_id] = poses[ix][thread_id+3];\n qj[thread_id] = poses[jx][thread_id+3];\n }\n\n __syncthreads();\n\n\n relSE3(ti, qi, tj, qj, tij, qij);\n\n float d, du, dv;\n\n GPU_1D_KERNEL_LOOP(k, ht*wd) {\n const int i = k / wd;\n const int j = k % wd;\n\n const float u = static_cast(j);\n const float v = static_cast(i);\n\n\n // if (disps[ix][i][j] < 0.01) {\n // continue;\n // }\n \n // homogenous coordinates\n Xi[0] = (u - cx) / fx;\n Xi[1] = (v - cy) / fy;\n Xi[2] = 1;\n Xi[3] = disps[ix][i][j];\n\n // transform homogenous point\n actSE3(tij, qij, Xi, Xj);\n\n du = fx * (Xj[0] / Xj[2]) + cx - u;\n dv = fy * (Xj[1] / Xj[2]) + cy - v;\n d = sqrtf(du*du + dv*dv);\n\n total[threadIdx.x] += beta;\n \n if (Xj[2] > MIN_DEPTH) {\n accum[threadIdx.x] += beta * d;\n valid[threadIdx.x] += beta;\n }\n\n Xi[0] = (u - cx) / fx;\n Xi[1] = (v - cy) / fy;\n Xi[2] = 1;\n Xi[3] = disps[ix][i][j];\n\n Xj[0] = Xi[0] + Xi[3] * tij[0];\n Xj[1] = Xi[1] + Xi[3] * tij[1];\n Xj[2] = Xi[2] + Xi[3] * tij[2];\n\n du = fx * (Xj[0] / Xj[2]) + cx - u;\n dv = fy * (Xj[1] / Xj[2]) + cy - v;\n d = sqrtf(du*du + dv*dv);\n\n total[threadIdx.x] += (1 - beta);\n \n if (Xj[2] > MIN_DEPTH) {\n accum[threadIdx.x] += (1 - beta) * d;\n valid[threadIdx.x] += (1 - beta);\n }\n }\n\n if (threadIdx.x == 0) {\n int tmp = ix;\n ix = jx;\n jx = tmp;\n }\n\n __syncthreads();\n\n }\n __syncthreads(); blockReduce(accum);\n __syncthreads(); blockReduce(total);\n __syncthreads(); blockReduce(valid);\n\n __syncthreads();\n\n if (thread_id == 0) {\n // dist[block_id] = (valid[0] / (total[0] + 1e-8) < 0.75) ? 1000.0 : accum[0] / valid[0];\n dist[block_id] = (valid[0] / (total[0] + 1e-8) < 0.75) ? 1000.0 : accum[0] / valid[0];\n }\n}\n\n\n\n__global__ void depth_filter_kernel(\n const torch::PackedTensorAccessor32 poses,\n const torch::PackedTensorAccessor32 disps,\n const torch::PackedTensorAccessor32 intrinsics,\n const torch::PackedTensorAccessor32 inds,\n const torch::PackedTensorAccessor32 thresh,\n torch::PackedTensorAccessor32 counter)\n{\n\n const int block_id = blockIdx.x;\n const int neigh_id = blockIdx.y;\n const int index = blockIdx.z * blockDim.x + threadIdx.x;\n\n // if (threadIdx.x == 0) {\n // printf(\"%d %d %d %d\\n\", blockIdx.x, blockIdx.y, blockDim.x, threadIdx.x);\n // }\n\n const int num = disps.size(0);\n const int ht = disps.size(1);\n const int wd = disps.size(2);\n\n __shared__ int ix;\n __shared__ int jx;\n\n __shared__ float fx;\n __shared__ float fy;\n __shared__ float cx;\n __shared__ float cy;\n\n __shared__ float ti[3], tj[3], tij[3];\n __shared__ float qi[4], qj[4], qij[4];\n\n if (threadIdx.x == 0) {\n ix = static_cast(inds[block_id]);\n jx = (neigh_id < 3) ? ix - neigh_id - 1 : ix + neigh_id;\n fx = intrinsics[0];\n fy = intrinsics[1];\n cx = intrinsics[2];\n cy = intrinsics[3];\n }\n\n __syncthreads();\n\n if (jx < 0 || jx >= num) {\n return;\n }\n\n const float t = thresh[block_id];\n\n // load poses from global memory\n if (threadIdx.x < 3) {\n ti[threadIdx.x] = poses[ix][threadIdx.x];\n tj[threadIdx.x] = poses[jx][threadIdx.x];\n }\n\n if (threadIdx.x < 4) {\n qi[threadIdx.x] = poses[ix][threadIdx.x+3];\n qj[threadIdx.x] = poses[jx][threadIdx.x+3];\n }\n\n __syncthreads();\n\n if (threadIdx.x == 0) {\n relSE3(ti, qi, tj, qj, tij, qij);\n }\n\n //points \n float Xi[4];\n float Xj[4];\n\n __syncthreads();\n\n if (index < ht*wd) {\n const int i = index / wd;\n const int j = index % wd;\n\n const float ui = static_cast(j);\n const float vi = static_cast(i);\n const float di = disps[ix][i][j];\n \n // homogenous coordinates\n Xi[0] = (ui - cx) / fx;\n Xi[1] = (vi - cy) / fy;\n Xi[2] = 1;\n Xi[3] = di;\n\n // transform homogenous point\n actSE3(tij, qij, Xi, Xj);\n\n const float uj = fx * (Xj[0] / Xj[2]) + cx;\n const float vj = fy * (Xj[1] / Xj[2]) + cy;\n const float dj = Xj[3] / Xj[2];\n\n const int u0 = static_cast(floor(uj));\n const int v0 = static_cast(floor(vj));\n\n if (u0 >= 0 && v0 >= 0 && u0 < wd-1 && v0 < ht-1) {\n const float wx = ceil(uj) - uj;\n const float wy = ceil(vj) - vj;\n\n const float d00 = disps[jx][v0+0][u0+0];\n const float d01 = disps[jx][v0+0][u0+1];\n const float d10 = disps[jx][v0+1][u0+0];\n const float d11 = disps[jx][v0+1][u0+1];\n\n const float dj_hat = wy*wx*d00 + wy*(1-wx)*d01 + (1-wy)*wx*d10 + (1-wy)*(1-wx)*d11;\n\n const float err = abs(1.0/dj - 1.0/dj_hat);\n if (abs(1.0/dj - 1.0/d00) < t) atomicAdd(&counter[block_id][i][j], 1.0f);\n else if (abs(1.0/dj - 1.0/d01) < t) atomicAdd(&counter[block_id][i][j], 1.0f);\n else if (abs(1.0/dj - 1.0/d10) < t) atomicAdd(&counter[block_id][i][j], 1.0f);\n else if (abs(1.0/dj - 1.0/d11) < t) atomicAdd(&counter[block_id][i][j], 1.0f);\n }\n }\n}\n\n\n\n__global__ void iproj_kernel(\n const torch::PackedTensorAccessor32 poses,\n const torch::PackedTensorAccessor32 disps,\n const torch::PackedTensorAccessor32 intrinsics,\n torch::PackedTensorAccessor32 points)\n\n{\n\n const int block_id = blockIdx.x;\n const int index = blockIdx.y * blockDim.x + threadIdx.x;\n\n\n const int num = disps.size(0);\n const int ht = disps.size(1);\n const int wd = disps.size(2);\n\n __shared__ float fx;\n __shared__ float fy;\n __shared__ float cx;\n __shared__ float cy;\n\n __shared__ float t[3];\n __shared__ float q[4];\n\n if (threadIdx.x == 0) {\n fx = intrinsics[0];\n fy = intrinsics[1];\n cx = intrinsics[2];\n cy = intrinsics[3];\n }\n\n __syncthreads();\n\n\n // load poses from global memory\n if (threadIdx.x < 3) {\n t[threadIdx.x] = poses[block_id][threadIdx.x];\n }\n\n if (threadIdx.x < 4) {\n q[threadIdx.x] = poses[block_id][threadIdx.x+3];\n }\n\n __syncthreads();\n\n //points \n float Xi[4];\n float Xj[4];\n\n if (index < ht*wd) {\n const int i = index / wd;\n const int j = index % wd;\n\n const float ui = static_cast(j);\n const float vi = static_cast(i);\n const float di = disps[block_id][i][j];\n \n // homogenous coordinates\n Xi[0] = (ui - cx) / fx;\n Xi[1] = (vi - cy) / fy;\n Xi[2] = 1;\n Xi[3] = di;\n\n // transform homogenous point\n actSE3(t, q, Xi, Xj);\n\n points[block_id][i][j][0] = Xj[0] / Xj[3];\n points[block_id][i][j][1] = Xj[1] / Xj[3];\n points[block_id][i][j][2] = Xj[2] / Xj[3];\n\n }\n}\n\n\n\n__global__ void accum_kernel(\n const torch::PackedTensorAccessor32 inps,\n const torch::PackedTensorAccessor32 ptrs,\n const torch::PackedTensorAccessor32 idxs,\n torch::PackedTensorAccessor32 outs)\n{\n \n const int block_id = blockIdx.x;\n const int D = inps.size(2);\n\n const int start = ptrs[block_id];\n const int end = ptrs[block_id+1];\n\n for (int k=threadIdx.x; k poses,\n const torch::PackedTensorAccessor32 dx,\n const int t0, const int t1) \n{\n\n for (int k=t0+threadIdx.x; k disps,\n const torch::PackedTensorAccessor32 dz,\n const torch::PackedTensorAccessor32 inds) \n{\n const int i = inds[blockIdx.x];\n const int ht = disps.size(1);\n const int wd = disps.size(2);\n\n for (int k=threadIdx.x; k();\n long* jx_data = jx_cpu.data_ptr();\n long* kx_data = inds.data_ptr();\n\n int count = jx.size(0);\n std::vector cols;\n\n torch::Tensor ptrs_cpu = torch::zeros({count+1}, \n torch::TensorOptions().dtype(torch::kInt64));\n \n long* ptrs_data = ptrs_cpu.data_ptr();\n ptrs_data[0] = 0;\n\n int i = 0;\n for (int j=0; j();\n\n for (int i=0; i>>(\n data.packed_accessor32(),\n ptrs.packed_accessor32(),\n idxs.packed_accessor32(),\n out.packed_accessor32());\n\n return out;\n}\n\n\n__global__ void EEt6x6_kernel(\n const torch::PackedTensorAccessor32 E,\n const torch::PackedTensorAccessor32 Q,\n const torch::PackedTensorAccessor32 idx,\n torch::PackedTensorAccessor32 S)\n{\n\n // indicices\n const int ix = idx[blockIdx.x][0];\n const int jx = idx[blockIdx.x][1];\n const int kx = idx[blockIdx.x][2];\n\n const int D = E.size(2);\n\n float dS[6][6];\n float ei[6];\n float ej[6];\n\n for (int i=0; i<6; i++) {\n for (int j=0; j<6; j++) {\n dS[i][j] = 0;\n }\n }\n\n for (int k=threadIdx.x; k E,\n const torch::PackedTensorAccessor32 Q,\n const torch::PackedTensorAccessor32 w,\n const torch::PackedTensorAccessor32 idx,\n torch::PackedTensorAccessor32 v)\n{\n const int D = E.size(2);\n const int kx = idx[blockIdx.x][0];\n\n float b[6];\n for (int n=0; n<6; n++) {\n b[n] = 0.0;\n }\n\n for (int k=threadIdx.x; k E,\n const torch::PackedTensorAccessor32 x,\n const torch::PackedTensorAccessor32 idx,\n torch::PackedTensorAccessor32 w)\n{\n\n const int D = E.size(2);\n const int ix = idx[blockIdx.x];\n\n if (idx[blockIdx.x] <= 0 || idx[blockIdx.x] >= x.size(0))\n return;\n\n for (int k=threadIdx.x; k A;\n Eigen::VectorX b;\n\n SparseBlock(int N, int M) : N(N), M(M) {\n A = Eigen::SparseMatrix(N*M, N*M);\n b = Eigen::VectorXd::Zero(N*M);\n }\n\n SparseBlock(Eigen::SparseMatrix const& A, Eigen::VectorX const& b, \n int N, int M) : A(A), b(b), N(N), M(M) {}\n\n void update_lhs(torch::Tensor As, torch::Tensor ii, torch::Tensor jj) {\n\n auto As_cpu = As.to(torch::kCPU).to(torch::kFloat64);\n auto ii_cpu = ii.to(torch::kCPU).to(torch::kInt64);\n auto jj_cpu = jj.to(torch::kCPU).to(torch::kInt64);\n\n auto As_acc = As_cpu.accessor();\n auto ii_acc = ii_cpu.accessor();\n auto jj_acc = jj_cpu.accessor();\n\n std::vector tripletList;\n for (int n=0; n= 0 && j >= 0) {\n for (int k=0; k();\n auto ii_acc = ii_cpu.accessor();\n\n for (int n=0; n= 0) {\n for (int j=0; j get_dense() {\n Eigen::MatrixXd Ad = Eigen::MatrixXd(A);\n\n torch::Tensor H = torch::from_blob(Ad.data(), {N*M, N*M}, torch::TensorOptions()\n .dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);\n\n torch::Tensor v = torch::from_blob(b.data(), {N*M, 1}, torch::TensorOptions()\n .dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);\n\n return std::make_tuple(H, v);\n }\n\n void get_dense_extend(torch::PackedTensorAccessor32 H,\n torch::PackedTensorAccessor32 v)\n {\n // Eigen::MatrixXd Ad = Eigen::MatrixXd(A);\n // std::cerr< L(A);\n L.diagonal().array() += ep + lm * L.diagonal().array();\n\n Eigen::SimplicialLLT> solver;\n solver.compute(L);\n\n if (solver.info() == Eigen::Success) {\n Eigen::VectorXd x = solver.solve(b);\n dx = torch::from_blob(x.data(), {N, M}, torch::TensorOptions()\n .dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);\n }\n else {\n dx = torch::zeros({N, M}, torch::TensorOptions()\n .device(torch::kCUDA).dtype(torch::kFloat32));\n }\n \n return dx;\n }\n\n torch::Tensor solve_dense(const float lm=0.0001,const float ep = 0.1){\n torch::Tensor dx;\n Eigen::MatrixXd L(A);\n L.diagonal().array() += ep + lm * L.diagonal().array();\n Eigen::LLT solver;\n solver.compute(L);\n\n if (solver.info() == Eigen::Success) {\n Eigen::VectorXd x = solver.solve(b);\n dx = torch::from_blob(x.data(), {N, M}, torch::TensorOptions()\n .dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);\n }\n else {\n dx = torch::zeros({N, M}, torch::TensorOptions()\n .device(torch::kCUDA).dtype(torch::kFloat32));\n }\n return dx;\n }\n\n private:\n const int N;\n const int M;\n\n};\n\n\nSparseBlock schur_block(torch::Tensor E,\n torch::Tensor Q,\n torch::Tensor w,\n torch::Tensor ii,\n torch::Tensor jj,\n torch::Tensor kk,\n const int t0,\n const int t1)\n{\n\n torch::Tensor ii_cpu = ii.to(torch::kCPU);\n torch::Tensor jj_cpu = jj.to(torch::kCPU);\n torch::Tensor kk_cpu = kk.to(torch::kCPU);\n\n const int P = t1 - t0;\n const long* ii_data = ii_cpu.data_ptr();\n const long* jj_data = jj_cpu.data_ptr();\n const long* kk_data = kk_cpu.data_ptr();\n\n std::vector> graph(P);\n std::vector> index(P);\n\n for (int n=0; n= t0 && j <= t1) {\n const int t = j - t0;\n graph[t].push_back(k);\n index[t].push_back(n);\n }\n }\n\n std::vector ii_list, jj_list, idx, jdx;\n\n for (int i=0; i>>(\n E.packed_accessor32(),\n Q.packed_accessor32(),\n ix_cuda.packed_accessor32(),\n S.packed_accessor32