Repository: GREAT-WHU/DBA-Fusion
Branch: main
Commit: 2c658a3be747
Files: 71
Total size: 377.7 KB
Directory structure:
gitextract_ny3kcvri/
├── .gitignore
├── .gitmodules
├── LICENSE
├── README.md
├── batch_kitti360.py
├── batch_subt.py
├── batch_tumvi.py
├── batch_whu.py
├── calib/
│ ├── 0412.txt
│ ├── 0412new.txt
│ ├── 1012.txt
│ ├── barn.txt
│ ├── carla.txt
│ ├── eth.txt
│ ├── euroc.txt
│ ├── handheld.txt
│ ├── kitti_360.txt
│ ├── subt.txt
│ ├── tartan.txt
│ ├── tum3.txt
│ └── tumvi.txt
├── dataset/
│ ├── euroc_to_hdf5.py
│ ├── kitti360_to_hdf5.py
│ └── tumvi_to_hdf5.py
├── dbaf/
│ ├── covisible_graph.py
│ ├── data_readers/
│ │ ├── __init__.py
│ │ ├── augmentation.py
│ │ ├── base.py
│ │ ├── factory.py
│ │ ├── rgbd_utils.py
│ │ ├── stream.py
│ │ ├── tartan.py
│ │ └── tartan_test.txt
│ ├── dbaf.py
│ ├── dbaf_frontend.py
│ ├── depth_video.py
│ ├── droid_net.py
│ ├── geoFunc/
│ │ ├── __init__.py
│ │ ├── const_value.py
│ │ └── trans.py
│ ├── geom/
│ │ ├── __init__.py
│ │ ├── ba.py
│ │ ├── chol.py
│ │ ├── graph_utils.py
│ │ ├── losses.py
│ │ └── projective_ops.py
│ ├── modules/
│ │ ├── __init__.py
│ │ ├── clipping.py
│ │ ├── corr.py
│ │ ├── extractor.py
│ │ └── gru.py
│ ├── motion_filter.py
│ └── multi_sensor.py
├── demo_vio_kitti360.py
├── demo_vio_subt.py
├── demo_vio_tumvi.py
├── demo_vio_whu.py
├── evaluation_scripts/
│ ├── batch_tumvi.py
│ ├── evaluate_kitti.py
│ └── evaluate_tumvi.py
├── results/
│ └── PLACEHOLDER
├── setup.py
├── src/
│ ├── altcorr_kernel.cu
│ ├── bacore.h
│ ├── correlation_kernels.cu
│ ├── droid.cpp
│ └── droid_kernels.cu
└── visualization/
├── check_reconstruction_kitti.py
├── check_reconstruction_kitti_animation.py
├── check_reconstruction_tumvi.py
└── check_reconstruction_tumvi_animation.py
================================================
FILE CONTENTS
================================================
================================================
FILE: .gitignore
================================================
a# Byte-compiled / optimized / DLL files
__pycache__/
*.py[cod]
*$py.class
# C extensions
*.so
# Distribution / packaging
.Python
build/
develop-eggs/
dist/
downloads/
eggs/
.eggs/
lib/
lib64/
parts/
sdist/
var/
wheels/
share/python-wheels/
*.egg-info/
.installed.cfg
*.egg
MANIFEST
# PyInstaller
# Usually these files are written by a python script from a template
# before PyInstaller builds the exe, so as to inject date/other infos into it.
*.manifest
*.spec
# Installer logs
pip-log.txt
pip-delete-this-directory.txt
# Unit test / coverage reports
htmlcov/
.tox/
.nox/
.coverage
.coverage.*
.cache
nosetests.xml
coverage.xml
*.cover
*.py,cover
.hypothesis/
.pytest_cache/
cover/
# Translations
*.mo
*.pot
# Django stuff:
*.log
local_settings.py
db.sqlite3
db.sqlite3-journal
# Flask stuff:
instance/
.webassets-cache
# Scrapy stuff:
.scrapy
# Sphinx documentation
docs/_build/
# PyBuilder
.pybuilder/
target/
# Jupyter Notebook
.ipynb_checkpoints
# IPython
profile_default/
ipython_config.py
# pyenv
# For a library or package, you might want to ignore these files since the code is
# intended to run in multiple environments; otherwise, check them in:
# .python-version
# pipenv
# According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
# However, in case of collaboration, if having platform-specific dependencies or dependencies
# having no cross-platform support, pipenv may install dependencies that don't work, or not
# install all needed dependencies.
#Pipfile.lock
# PEP 582; used by e.g. github.com/David-OConnor/pyflow
__pypackages__/
# Celery stuff
celerybeat-schedule
celerybeat.pid
# SageMath parsed files
*.sage.py
# Environments
.env
.venv
env/
venv/
ENV/
env.bak/
venv.bak/
# Spyder project settings
.spyderproject
.spyproject
# Rope project settings
.ropeproject
# mkdocs documentation
/site
# mypy
.mypy_cache/
.dmypy.json
dmypy.json
# Pyre type checker
.pyre/
# pytype static type analyzer
.pytype/
# Cython debug symbols
cython_debug/
__pycache__
build
dist
*.egg-info
*.vscode/
*.pth
tests
checkpoints
datasets
runs
cache
*.out
*.o
data
figures/*.pdf
================================================
FILE: .gitmodules
================================================
[submodule "thirdparty/lietorch"]
path = thirdparty/lietorch
url = https://github.com/princeton-vl/lietorch
[submodule "thirdparty/eigen"]
path = thirdparty/eigen
url = https://gitlab.com/libeigen/eigen.git
================================================
FILE: LICENSE
================================================
GNU GENERAL PUBLIC LICENSE
Version 3, 29 June 2007
Copyright (C) 2007 Free Software Foundation, Inc.
Everyone is permitted to copy and distribute verbatim copies
of this license document, but changing it is not allowed.
Preamble
The GNU General Public License is a free, copyleft license for
software and other kinds of works.
The licenses for most software and other practical works are designed
to take away your freedom to share and change the works. By contrast,
the GNU General Public License is intended to guarantee your freedom to
share and change all versions of a program--to make sure it remains free
software for all its users. We, the Free Software Foundation, use the
GNU General Public License for most of our software; it applies also to
any other work released this way by its authors. You can apply it to
your programs, too.
When we speak of free software, we are referring to freedom, not
price. Our General Public Licenses are designed to make sure that you
have the freedom to distribute copies of free software (and charge for
them if you wish), that you receive source code or can get it if you
want it, that you can change the software or use pieces of it in new
free programs, and that you know you can do these things.
To protect your rights, we need to prevent others from denying you
these rights or asking you to surrender the rights. Therefore, you have
certain responsibilities if you distribute copies of the software, or if
you modify it: responsibilities to respect the freedom of others.
For example, if you distribute copies of such a program, whether
gratis or for a fee, you must pass on to the recipients the same
freedoms that you received. You must make sure that they, too, receive
or can get the source code. And you must show them these terms so they
know their rights.
Developers that use the GNU GPL protect your rights with two steps:
(1) assert copyright on the software, and (2) offer you this License
giving you legal permission to copy, distribute and/or modify it.
For the developers' and authors' protection, the GPL clearly explains
that there is no warranty for this free software. For both users' and
authors' sake, the GPL requires that modified versions be marked as
changed, so that their problems will not be attributed erroneously to
authors of previous versions.
Some devices are designed to deny users access to install or run
modified versions of the software inside them, although the manufacturer
can do so. This is fundamentally incompatible with the aim of
protecting users' freedom to change the software. The systematic
pattern of such abuse occurs in the area of products for individuals to
use, which is precisely where it is most unacceptable. Therefore, we
have designed this version of the GPL to prohibit the practice for those
products. If such problems arise substantially in other domains, we
stand ready to extend this provision to those domains in future versions
of the GPL, as needed to protect the freedom of users.
Finally, every program is threatened constantly by software patents.
States should not allow patents to restrict development and use of
software on general-purpose computers, but in those that do, we wish to
avoid the special danger that patents applied to a free program could
make it effectively proprietary. To prevent this, the GPL assures that
patents cannot be used to render the program non-free.
The precise terms and conditions for copying, distribution and
modification follow.
TERMS AND CONDITIONS
0. Definitions.
"This License" refers to version 3 of the GNU General Public License.
"Copyright" also means copyright-like laws that apply to other kinds of
works, such as semiconductor masks.
"The Program" refers to any copyrightable work licensed under this
License. Each licensee is addressed as "you". "Licensees" and
"recipients" may be individuals or organizations.
To "modify" a work means to copy from or adapt all or part of the work
in a fashion requiring copyright permission, other than the making of an
exact copy. The resulting work is called a "modified version" of the
earlier work or a work "based on" the earlier work.
A "covered work" means either the unmodified Program or a work based
on the Program.
To "propagate" a work means to do anything with it that, without
permission, would make you directly or secondarily liable for
infringement under applicable copyright law, except executing it on a
computer or modifying a private copy. Propagation includes copying,
distribution (with or without modification), making available to the
public, and in some countries other activities as well.
To "convey" a work means any kind of propagation that enables other
parties to make or receive copies. Mere interaction with a user through
a computer network, with no transfer of a copy, is not conveying.
An interactive user interface displays "Appropriate Legal Notices"
to the extent that it includes a convenient and prominently visible
feature that (1) displays an appropriate copyright notice, and (2)
tells the user that there is no warranty for the work (except to the
extent that warranties are provided), that licensees may convey the
work under this License, and how to view a copy of this License. If
the interface presents a list of user commands or options, such as a
menu, a prominent item in the list meets this criterion.
1. Source Code.
The "source code" for a work means the preferred form of the work
for making modifications to it. "Object code" means any non-source
form of a work.
A "Standard Interface" means an interface that either is an official
standard defined by a recognized standards body, or, in the case of
interfaces specified for a particular programming language, one that
is widely used among developers working in that language.
The "System Libraries" of an executable work include anything, other
than the work as a whole, that (a) is included in the normal form of
packaging a Major Component, but which is not part of that Major
Component, and (b) serves only to enable use of the work with that
Major Component, or to implement a Standard Interface for which an
implementation is available to the public in source code form. A
"Major Component", in this context, means a major essential component
(kernel, window system, and so on) of the specific operating system
(if any) on which the executable work runs, or a compiler used to
produce the work, or an object code interpreter used to run it.
The "Corresponding Source" for a work in object code form means all
the source code needed to generate, install, and (for an executable
work) run the object code and to modify the work, including scripts to
control those activities. However, it does not include the work's
System Libraries, or general-purpose tools or generally available free
programs which are used unmodified in performing those activities but
which are not part of the work. For example, Corresponding Source
includes interface definition files associated with source files for
the work, and the source code for shared libraries and dynamically
linked subprograms that the work is specifically designed to require,
such as by intimate data communication or control flow between those
subprograms and other parts of the work.
The Corresponding Source need not include anything that users
can regenerate automatically from other parts of the Corresponding
Source.
The Corresponding Source for a work in source code form is that
same work.
2. Basic Permissions.
All rights granted under this License are granted for the term of
copyright on the Program, and are irrevocable provided the stated
conditions are met. This License explicitly affirms your unlimited
permission to run the unmodified Program. The output from running a
covered work is covered by this License only if the output, given its
content, constitutes a covered work. This License acknowledges your
rights of fair use or other equivalent, as provided by copyright law.
You may make, run and propagate covered works that you do not
convey, without conditions so long as your license otherwise remains
in force. You may convey covered works to others for the sole purpose
of having them make modifications exclusively for you, or provide you
with facilities for running those works, provided that you comply with
the terms of this License in conveying all material for which you do
not control copyright. Those thus making or running the covered works
for you must do so exclusively on your behalf, under your direction
and control, on terms that prohibit them from making any copies of
your copyrighted material outside their relationship with you.
Conveying under any other circumstances is permitted solely under
the conditions stated below. Sublicensing is not allowed; section 10
makes it unnecessary.
3. Protecting Users' Legal Rights From Anti-Circumvention Law.
No covered work shall be deemed part of an effective technological
measure under any applicable law fulfilling obligations under article
11 of the WIPO copyright treaty adopted on 20 December 1996, or
similar laws prohibiting or restricting circumvention of such
measures.
When you convey a covered work, you waive any legal power to forbid
circumvention of technological measures to the extent such circumvention
is effected by exercising rights under this License with respect to
the covered work, and you disclaim any intention to limit operation or
modification of the work as a means of enforcing, against the work's
users, your or third parties' legal rights to forbid circumvention of
technological measures.
4. Conveying Verbatim Copies.
You may convey verbatim copies of the Program's source code as you
receive it, in any medium, provided that you conspicuously and
appropriately publish on each copy an appropriate copyright notice;
keep intact all notices stating that this License and any
non-permissive terms added in accord with section 7 apply to the code;
keep intact all notices of the absence of any warranty; and give all
recipients a copy of this License along with the Program.
You may charge any price or no price for each copy that you convey,
and you may offer support or warranty protection for a fee.
5. Conveying Modified Source Versions.
You may convey a work based on the Program, or the modifications to
produce it from the Program, in the form of source code under the
terms of section 4, provided that you also meet all of these conditions:
a) The work must carry prominent notices stating that you modified
it, and giving a relevant date.
b) The work must carry prominent notices stating that it is
released under this License and any conditions added under section
7. This requirement modifies the requirement in section 4 to
"keep intact all notices".
c) You must license the entire work, as a whole, under this
License to anyone who comes into possession of a copy. This
License will therefore apply, along with any applicable section 7
additional terms, to the whole of the work, and all its parts,
regardless of how they are packaged. This License gives no
permission to license the work in any other way, but it does not
invalidate such permission if you have separately received it.
d) If the work has interactive user interfaces, each must display
Appropriate Legal Notices; however, if the Program has interactive
interfaces that do not display Appropriate Legal Notices, your
work need not make them do so.
A compilation of a covered work with other separate and independent
works, which are not by their nature extensions of the covered work,
and which are not combined with it such as to form a larger program,
in or on a volume of a storage or distribution medium, is called an
"aggregate" if the compilation and its resulting copyright are not
used to limit the access or legal rights of the compilation's users
beyond what the individual works permit. Inclusion of a covered work
in an aggregate does not cause this License to apply to the other
parts of the aggregate.
6. Conveying Non-Source Forms.
You may convey a covered work in object code form under the terms
of sections 4 and 5, provided that you also convey the
machine-readable Corresponding Source under the terms of this License,
in one of these ways:
a) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by the
Corresponding Source fixed on a durable physical medium
customarily used for software interchange.
b) Convey the object code in, or embodied in, a physical product
(including a physical distribution medium), accompanied by a
written offer, valid for at least three years and valid for as
long as you offer spare parts or customer support for that product
model, to give anyone who possesses the object code either (1) a
copy of the Corresponding Source for all the software in the
product that is covered by this License, on a durable physical
medium customarily used for software interchange, for a price no
more than your reasonable cost of physically performing this
conveying of source, or (2) access to copy the
Corresponding Source from a network server at no charge.
c) Convey individual copies of the object code with a copy of the
written offer to provide the Corresponding Source. This
alternative is allowed only occasionally and noncommercially, and
only if you received the object code with such an offer, in accord
with subsection 6b.
d) Convey the object code by offering access from a designated
place (gratis or for a charge), and offer equivalent access to the
Corresponding Source in the same way through the same place at no
further charge. You need not require recipients to copy the
Corresponding Source along with the object code. If the place to
copy the object code is a network server, the Corresponding Source
may be on a different server (operated by you or a third party)
that supports equivalent copying facilities, provided you maintain
clear directions next to the object code saying where to find the
Corresponding Source. Regardless of what server hosts the
Corresponding Source, you remain obligated to ensure that it is
available for as long as needed to satisfy these requirements.
e) Convey the object code using peer-to-peer transmission, provided
you inform other peers where the object code and Corresponding
Source of the work are being offered to the general public at no
charge under subsection 6d.
A separable portion of the object code, whose source code is excluded
from the Corresponding Source as a System Library, need not be
included in conveying the object code work.
A "User Product" is either (1) a "consumer product", which means any
tangible personal property which is normally used for personal, family,
or household purposes, or (2) anything designed or sold for incorporation
into a dwelling. In determining whether a product is a consumer product,
doubtful cases shall be resolved in favor of coverage. For a particular
product received by a particular user, "normally used" refers to a
typical or common use of that class of product, regardless of the status
of the particular user or of the way in which the particular user
actually uses, or expects or is expected to use, the product. A product
is a consumer product regardless of whether the product has substantial
commercial, industrial or non-consumer uses, unless such uses represent
the only significant mode of use of the product.
"Installation Information" for a User Product means any methods,
procedures, authorization keys, or other information required to install
and execute modified versions of a covered work in that User Product from
a modified version of its Corresponding Source. The information must
suffice to ensure that the continued functioning of the modified object
code is in no case prevented or interfered with solely because
modification has been made.
If you convey an object code work under this section in, or with, or
specifically for use in, a User Product, and the conveying occurs as
part of a transaction in which the right of possession and use of the
User Product is transferred to the recipient in perpetuity or for a
fixed term (regardless of how the transaction is characterized), the
Corresponding Source conveyed under this section must be accompanied
by the Installation Information. But this requirement does not apply
if neither you nor any third party retains the ability to install
modified object code on the User Product (for example, the work has
been installed in ROM).
The requirement to provide Installation Information does not include a
requirement to continue to provide support service, warranty, or updates
for a work that has been modified or installed by the recipient, or for
the User Product in which it has been modified or installed. Access to a
network may be denied when the modification itself materially and
adversely affects the operation of the network or violates the rules and
protocols for communication across the network.
Corresponding Source conveyed, and Installation Information provided,
in accord with this section must be in a format that is publicly
documented (and with an implementation available to the public in
source code form), and must require no special password or key for
unpacking, reading or copying.
7. Additional Terms.
"Additional permissions" are terms that supplement the terms of this
License by making exceptions from one or more of its conditions.
Additional permissions that are applicable to the entire Program shall
be treated as though they were included in this License, to the extent
that they are valid under applicable law. If additional permissions
apply only to part of the Program, that part may be used separately
under those permissions, but the entire Program remains governed by
this License without regard to the additional permissions.
When you convey a copy of a covered work, you may at your option
remove any additional permissions from that copy, or from any part of
it. (Additional permissions may be written to require their own
removal in certain cases when you modify the work.) You may place
additional permissions on material, added by you to a covered work,
for which you have or can give appropriate copyright permission.
Notwithstanding any other provision of this License, for material you
add to a covered work, you may (if authorized by the copyright holders of
that material) supplement the terms of this License with terms:
a) Disclaiming warranty or limiting liability differently from the
terms of sections 15 and 16 of this License; or
b) Requiring preservation of specified reasonable legal notices or
author attributions in that material or in the Appropriate Legal
Notices displayed by works containing it; or
c) Prohibiting misrepresentation of the origin of that material, or
requiring that modified versions of such material be marked in
reasonable ways as different from the original version; or
d) Limiting the use for publicity purposes of names of licensors or
authors of the material; or
e) Declining to grant rights under trademark law for use of some
trade names, trademarks, or service marks; or
f) Requiring indemnification of licensors and authors of that
material by anyone who conveys the material (or modified versions of
it) with contractual assumptions of liability to the recipient, for
any liability that these contractual assumptions directly impose on
those licensors and authors.
All other non-permissive additional terms are considered "further
restrictions" within the meaning of section 10. If the Program as you
received it, or any part of it, contains a notice stating that it is
governed by this License along with a term that is a further
restriction, you may remove that term. If a license document contains
a further restriction but permits relicensing or conveying under this
License, you may add to a covered work material governed by the terms
of that license document, provided that the further restriction does
not survive such relicensing or conveying.
If you add terms to a covered work in accord with this section, you
must place, in the relevant source files, a statement of the
additional terms that apply to those files, or a notice indicating
where to find the applicable terms.
Additional terms, permissive or non-permissive, may be stated in the
form of a separately written license, or stated as exceptions;
the above requirements apply either way.
8. Termination.
You may not propagate or modify a covered work except as expressly
provided under this License. Any attempt otherwise to propagate or
modify it is void, and will automatically terminate your rights under
this License (including any patent licenses granted under the third
paragraph of section 11).
However, if you cease all violation of this License, then your
license from a particular copyright holder is reinstated (a)
provisionally, unless and until the copyright holder explicitly and
finally terminates your license, and (b) permanently, if the copyright
holder fails to notify you of the violation by some reasonable means
prior to 60 days after the cessation.
Moreover, your license from a particular copyright holder is
reinstated permanently if the copyright holder notifies you of the
violation by some reasonable means, this is the first time you have
received notice of violation of this License (for any work) from that
copyright holder, and you cure the violation prior to 30 days after
your receipt of the notice.
Termination of your rights under this section does not terminate the
licenses of parties who have received copies or rights from you under
this License. If your rights have been terminated and not permanently
reinstated, you do not qualify to receive new licenses for the same
material under section 10.
9. Acceptance Not Required for Having Copies.
You are not required to accept this License in order to receive or
run a copy of the Program. Ancillary propagation of a covered work
occurring solely as a consequence of using peer-to-peer transmission
to receive a copy likewise does not require acceptance. However,
nothing other than this License grants you permission to propagate or
modify any covered work. These actions infringe copyright if you do
not accept this License. Therefore, by modifying or propagating a
covered work, you indicate your acceptance of this License to do so.
10. Automatic Licensing of Downstream Recipients.
Each time you convey a covered work, the recipient automatically
receives a license from the original licensors, to run, modify and
propagate that work, subject to this License. You are not responsible
for enforcing compliance by third parties with this License.
An "entity transaction" is a transaction transferring control of an
organization, or substantially all assets of one, or subdividing an
organization, or merging organizations. If propagation of a covered
work results from an entity transaction, each party to that
transaction who receives a copy of the work also receives whatever
licenses to the work the party's predecessor in interest had or could
give under the previous paragraph, plus a right to possession of the
Corresponding Source of the work from the predecessor in interest, if
the predecessor has it or can get it with reasonable efforts.
You may not impose any further restrictions on the exercise of the
rights granted or affirmed under this License. For example, you may
not impose a license fee, royalty, or other charge for exercise of
rights granted under this License, and you may not initiate litigation
(including a cross-claim or counterclaim in a lawsuit) alleging that
any patent claim is infringed by making, using, selling, offering for
sale, or importing the Program or any portion of it.
11. Patents.
A "contributor" is a copyright holder who authorizes use under this
License of the Program or a work on which the Program is based. The
work thus licensed is called the contributor's "contributor version".
A contributor's "essential patent claims" are all patent claims
owned or controlled by the contributor, whether already acquired or
hereafter acquired, that would be infringed by some manner, permitted
by this License, of making, using, or selling its contributor version,
but do not include claims that would be infringed only as a
consequence of further modification of the contributor version. For
purposes of this definition, "control" includes the right to grant
patent sublicenses in a manner consistent with the requirements of
this License.
Each contributor grants you a non-exclusive, worldwide, royalty-free
patent license under the contributor's essential patent claims, to
make, use, sell, offer for sale, import and otherwise run, modify and
propagate the contents of its contributor version.
In the following three paragraphs, a "patent license" is any express
agreement or commitment, however denominated, not to enforce a patent
(such as an express permission to practice a patent or covenant not to
sue for patent infringement). To "grant" such a patent license to a
party means to make such an agreement or commitment not to enforce a
patent against the party.
If you convey a covered work, knowingly relying on a patent license,
and the Corresponding Source of the work is not available for anyone
to copy, free of charge and under the terms of this License, through a
publicly available network server or other readily accessible means,
then you must either (1) cause the Corresponding Source to be so
available, or (2) arrange to deprive yourself of the benefit of the
patent license for this particular work, or (3) arrange, in a manner
consistent with the requirements of this License, to extend the patent
license to downstream recipients. "Knowingly relying" means you have
actual knowledge that, but for the patent license, your conveying the
covered work in a country, or your recipient's use of the covered work
in a country, would infringe one or more identifiable patents in that
country that you have reason to believe are valid.
If, pursuant to or in connection with a single transaction or
arrangement, you convey, or propagate by procuring conveyance of, a
covered work, and grant a patent license to some of the parties
receiving the covered work authorizing them to use, propagate, modify
or convey a specific copy of the covered work, then the patent license
you grant is automatically extended to all recipients of the covered
work and works based on it.
A patent license is "discriminatory" if it does not include within
the scope of its coverage, prohibits the exercise of, or is
conditioned on the non-exercise of one or more of the rights that are
specifically granted under this License. You may not convey a covered
work if you are a party to an arrangement with a third party that is
in the business of distributing software, under which you make payment
to the third party based on the extent of your activity of conveying
the work, and under which the third party grants, to any of the
parties who would receive the covered work from you, a discriminatory
patent license (a) in connection with copies of the covered work
conveyed by you (or copies made from those copies), or (b) primarily
for and in connection with specific products or compilations that
contain the covered work, unless you entered into that arrangement,
or that patent license was granted, prior to 28 March 2007.
Nothing in this License shall be construed as excluding or limiting
any implied license or other defenses to infringement that may
otherwise be available to you under applicable patent law.
12. No Surrender of Others' Freedom.
If conditions are imposed on you (whether by court order, agreement or
otherwise) that contradict the conditions of this License, they do not
excuse you from the conditions of this License. If you cannot convey a
covered work so as to satisfy simultaneously your obligations under this
License and any other pertinent obligations, then as a consequence you may
not convey it at all. For example, if you agree to terms that obligate you
to collect a royalty for further conveying from those to whom you convey
the Program, the only way you could satisfy both those terms and this
License would be to refrain entirely from conveying the Program.
13. Use with the GNU Affero General Public License.
Notwithstanding any other provision of this License, you have
permission to link or combine any covered work with a work licensed
under version 3 of the GNU Affero General Public License into a single
combined work, and to convey the resulting work. The terms of this
License will continue to apply to the part which is the covered work,
but the special requirements of the GNU Affero General Public License,
section 13, concerning interaction through a network will apply to the
combination as such.
14. Revised Versions of this License.
The Free Software Foundation may publish revised and/or new versions of
the GNU General Public License from time to time. Such new versions will
be similar in spirit to the present version, but may differ in detail to
address new problems or concerns.
Each version is given a distinguishing version number. If the
Program specifies that a certain numbered version of the GNU General
Public License "or any later version" applies to it, you have the
option of following the terms and conditions either of that numbered
version or of any later version published by the Free Software
Foundation. If the Program does not specify a version number of the
GNU General Public License, you may choose any version ever published
by the Free Software Foundation.
If the Program specifies that a proxy can decide which future
versions of the GNU General Public License can be used, that proxy's
public statement of acceptance of a version permanently authorizes you
to choose that version for the Program.
Later license versions may give you additional or different
permissions. However, no additional obligations are imposed on any
author or copyright holder as a result of your choosing to follow a
later version.
15. Disclaimer of Warranty.
THERE IS NO WARRANTY FOR THE PROGRAM, TO THE EXTENT PERMITTED BY
APPLICABLE LAW. EXCEPT WHEN OTHERWISE STATED IN WRITING THE COPYRIGHT
HOLDERS AND/OR OTHER PARTIES PROVIDE THE PROGRAM "AS IS" WITHOUT WARRANTY
OF ANY KIND, EITHER EXPRESSED OR IMPLIED, INCLUDING, BUT NOT LIMITED TO,
THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
PURPOSE. THE ENTIRE RISK AS TO THE QUALITY AND PERFORMANCE OF THE PROGRAM
IS WITH YOU. SHOULD THE PROGRAM PROVE DEFECTIVE, YOU ASSUME THE COST OF
ALL NECESSARY SERVICING, REPAIR OR CORRECTION.
16. Limitation of Liability.
IN NO EVENT UNLESS REQUIRED BY APPLICABLE LAW OR AGREED TO IN WRITING
WILL ANY COPYRIGHT HOLDER, OR ANY OTHER PARTY WHO MODIFIES AND/OR CONVEYS
THE PROGRAM AS PERMITTED ABOVE, BE LIABLE TO YOU FOR DAMAGES, INCLUDING ANY
GENERAL, SPECIAL, INCIDENTAL OR CONSEQUENTIAL DAMAGES ARISING OUT OF THE
USE OR INABILITY TO USE THE PROGRAM (INCLUDING BUT NOT LIMITED TO LOSS OF
DATA OR DATA BEING RENDERED INACCURATE OR LOSSES SUSTAINED BY YOU OR THIRD
PARTIES OR A FAILURE OF THE PROGRAM TO OPERATE WITH ANY OTHER PROGRAMS),
EVEN IF SUCH HOLDER OR OTHER PARTY HAS BEEN ADVISED OF THE POSSIBILITY OF
SUCH DAMAGES.
17. Interpretation of Sections 15 and 16.
If the disclaimer of warranty and limitation of liability provided
above cannot be given local legal effect according to their terms,
reviewing courts shall apply local law that most closely approximates
an absolute waiver of all civil liability in connection with the
Program, unless a warranty or assumption of liability accompanies a
copy of the Program in return for a fee.
END OF TERMS AND CONDITIONS
How to Apply These Terms to Your New Programs
If you develop a new program, and you want it to be of the greatest
possible use to the public, the best way to achieve this is to make it
free software which everyone can redistribute and change under these terms.
To do so, attach the following notices to the program. It is safest
to attach them to the start of each source file to most effectively
state the exclusion of warranty; and each file should have at least
the "copyright" line and a pointer to where the full notice is found.
Copyright (C)
This program is free software: you can redistribute it and/or modify
it under the terms of the GNU General Public License as published by
the Free Software Foundation, either version 3 of the License, or
(at your option) any later version.
This program is distributed in the hope that it will be useful,
but WITHOUT ANY WARRANTY; without even the implied warranty of
MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
GNU General Public License for more details.
You should have received a copy of the GNU General Public License
along with this program. If not, see .
Also add information on how to contact you by electronic and paper mail.
If the program does terminal interaction, make it output a short
notice like this when it starts in an interactive mode:
Copyright (C)
This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
This is free software, and you are welcome to redistribute it
under certain conditions; type `show c' for details.
The hypothetical commands `show w' and `show c' should show the appropriate
parts of the General Public License. Of course, your program's commands
might be different; for a GUI interface, you would use an "about box".
You should also get your employer (if you work as a programmer) or school,
if any, to sign a "copyright disclaimer" for the program, if necessary.
For more information on this, and how to apply and follow the GNU GPL, see
.
The GNU General Public License does not permit incorporating your program
into proprietary programs. If your program is a subroutine library, you
may consider it more useful to permit linking proprietary applications with
the library. If this is what you want to do, use the GNU Lesser General
Public License instead of this License. But first, please read
.
================================================
FILE: README.md
================================================
# DBA-Fusion
>Tightly Integrating Deep Dense Visual Bundle Adjustment with Multiple Sensors for Large-Scale Localization and Mapping
[[Paper](https://arxiv.org/abs/2403.13714)] [[Video](https://www.bilibili.com/video/BV1yeNEecEwR/?share_source=copy_web&vd_source=a659a573a520a1151e294d0c8b9c842a)]
## What is this?
**DBA-Fusion** is basically a VIO system which fuses DROID-SLAM-like dense bundle adjustment (DBA) with classic factor graph optimization. This work enables **realtime metric-scale localization and dense mapping** with excellent accuracy and robustness. Besides, this framework supports the **flexible fusion of multiple sensors** like GNSS or wheel speed sensors, extending its applicability to large-scale scenarios.
## Update log
- [x] Code Upload (2024/02/28)
- [x] Monocular VIO Examples (2024/02/28)
- [x] Multi-sensor data sequence (WUH1012) used in the manuscript is available [here](https://drive.google.com/file/d/1w7UsAwreou_9YRYHz13QIGu6jOJGpdg5/view?usp=sharing).
- [x] Multi-Sensor Fusion Examples
- [ ] Stereo/RGB-D VIO Support
## Installation
The pipeline of the work is based on python, and the computation part is mainly based on Pytorch (with CUDA) and GTSAM.
Use the following commands to set up the python environment.
```Bash
conda create -n dbaf python=3.10.11
conda activate dbaf
# Other CUDA versions should also be fine.
pip install torch==1.11.0+cu113 torchvision==0.12.0+cu113 torchaudio==0.11.0 --extra-index-url https://download.pytorch.org/whl/cu113
pip install torch-scatter==2.0.9 -f https://data.pyg.org/whl/torch-1.11.0+cu113.html
pip install gdown tqdm numpy==1.25.0 numpy-quaternion==2022.4.3 opencv-python==4.7.0.72 scipy pyparsing matplotlib h5py
pip install evo --upgrade --no-binary evo
pip install open3d # optional for visualization
```
As for GTSAM, we make some modifications to it to extend the python wrapper APIs, clone it from the following repository and install it under your python environment.
```Bash
git clone https://github.com/yuxuanzhou97/gtsam
cd gtsam
mkdir build
cd build
cmake .. -DGTSAM_BUILD_PYTHON=1 -DGTSAM_PYTHON_VERSION=3.10.11
make python-install
```
Finally, run the following command to build DBA-Fusion.
```Bash
git clone --recurse-submodules https://github.com/GREAT-WHU/DBA-Fusion.git
cd DBA-Fusion
python setup.py install
```
## Run DBA-Fusion
We don't modify the model of DROID-SLAM so you can directly employ the weight trained for DROID-SLAM. Here we use the [model](https://drive.google.com/file/d/1PpqVt1H4maBa_GbPJp4NwxRsd9jk-elh/view?usp=sharing) pre-trained on TartanAir (provided by [DROID-SLAM](https://github.com/princeton-vl/DROID-SLAM?tab=readme-ov-file)), which shows great zero-shot performance on real-world datasets.
**(Attention!!!)**
For the default configurations, around ~10GB GPU memory is needed. Lower the "max_factors" argument to 36 or lower could help reduce the memory usage to ~6GB.
### 1. TUM-VI
1.1 Download the [TUM-VI](https://cvg.cit.tum.de/data/datasets/visual-inertial-dataset) datasets (512*512).
**(Optional)**
For better speed performance, it is recommended to convert the PNG images to a single HDF5 file through
```Bash
python dataset/tumvi_to_hdf5.py --imagedir=${DATASET_DIR}/dataset-${SEQ}_512_16/mav0/cam0/data --imagestamp=${DATASET_DIR}/dataset-${SEQ}_512_16/mav0/cam0/data.csv --h5path=${SEQ}.h5 --calib=calib/tumvi.txt --stride 4
```
1.2 Specify the data path in [batch_tumvi.py](../batch_tumvi.py) (if you use HDF5 file, activate the "--enable_h5" and "--h5_path" arguments), run the following command
```Bash
python batch_tumvi.py # This would trigger demo_vio_tumvi.py automatically.
```
Look into [demo_vio_tumvi.py](../demo_vio_tumvi.py) to learn about the arguments. Data loading and almost all the parameters are specified in this **one** file.
1.3 The outputs of the program includes **a text file** which contains real-time navigation results and **a .pkl file** which contains all keyframe poses and point clouds.
To evaluate the realtime pose estimation performance, run the following command (notice to change the file paths in *evaluate_kitti.py*)
```Bash
python evaluation_scripts/evaluate_tumvi.py --seq ${SEQ}
```
or
```Bash
python evaluation_scripts/batch_evaluate_tumvi.py
```
For 3D visualization, currently we haven't handled the realtime visualization functionality. Run the scripts in the **"visualization"** folder for post-time visualization.
```Bash
python visualization/check_reconstruction_tumvi.py
```
### 2. KITTI-360
2.1 Download the [KITTI-360](https://www.cvlibs.net/datasets/kitti-360/index.php) datasets. Notice that we use the **unrectified perspective images** for the evaluation (named like "2013_05_28_drive_XXXX_sync/image_00/data_rgb").
For **IMU** data and IMU-centered **ground-truth poses**, we transform the axises to **Right-Forward-Up (RFU)** and check the synchronization. Besides, we use [OpenVINS](https://github.com/rpng/open_vins/) (in stereo VIO mode) to online refine the Camera-IMU extrinsics and time offsets (whose pre-calibrated values seem not very accurate) on the sequences. The refined parameters are used for for all the tests.
**To reproduce the results**, just download the our prepared IMU and ground-truth data from [here](https://drive.google.com/file/d/1BO8zGvoey7IdwbWXmAdlhGPr6hiCFJ6Y/view?usp=drive_link), then uncompress it to the data path.
**(Optional)**
Similar to the TUM-VI part, you can use the following script to generate a HDF5 file for best speed performance.
```Bash
python dataset/kitti360_to_hdf5.py --imagedir=${DATASET_DIR}/2013_05_28_drive_%s_sync/image_00/data_rgb --imagestamp=${DATASET_DIR}/2013_05_28_drive_%s_sync/camstamp.txt --h5path=${SEQ}.h5 --calib=calib/kitti360.txt --stride 2
```
2.2 Run the following command
```Bash
python batch_kitti360.py
```
Dataloading and parameters are specified in [demo_vio_kitti360.py](../demo_vio_kitti360.py).
2.3 For evaluation and visualization, run
```Bash
python evaluation_scripts/evaluate_kitti360.py --seq ${SEQ}
python visualization/check_reconstruction_kitti360.py
```
### 3. WUH1012
Download our self-collected data sequence from [here](https://drive.google.com/file/d/1w7UsAwreou_9YRYHz13QIGu6jOJGpdg5/view?usp=sharing).
See [batch_whu.py](../batch_whu.py) for multi-sensor fusion in different modes (VIO + wheel speed/GNSS), as described in the manuscript.
### 4. Run on Your Own Dataset
To run monocular VIO on your own dataset,
* Duplicate a script from [demo_vio_kitti360.py](../demo_vio_kitti360.py) or [demo_vio_tumvi.py](../demo_vio_tumvi.py).
* In the script, specify the data loading procedure of IMU data and images.
* Specify the camera intrinsics and camera-IMU extrinsics in the script.
* Try it!
## Some Results
- Visual point cloud map compared to accumulated LiDAR point clouds.
- Further processing on the visual point clouds. (P.S. For 3-D GS, the point positions and number are fixed. The training time is around 3 minutes on RTX4080 laptop. )
## Acknowledgement
DBA-Fusion is developed by [GREAT](http://igmas.users.sgg.whu.edu.cn/group) (GNSS+ REsearch, Application and Teaching) Group, School of Geodesy and Geomatics, Wuhan University.
This 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).
================================================
FILE: batch_kitti360.py
================================================
import os
import subprocess
for i in ['0000','0002','0003','0004','0005','0006','0009','0010']:
p = subprocess.Popen("python demo_vio_kitti360.py" +\
" --imagedir=/home/zhouyuxuan/data/2013_05_28_drive_%s_sync/image_00/data_rgb"%i +\
" --imagestamp=/home/zhouyuxuan/data/2013_05_28_drive_%s_sync/camstamp.txt"%i +\
" --imupath=/home/zhouyuxuan/data/2013_05_28_drive_%s_sync/imu.txt"%i +\
" --gtpath=/home/zhouyuxuan/data/2013_05_28_drive_%s_sync/gt_local.txt"%i +\
# " --enable_h5" +\
# " --h5path=/home/zhouyuxuan/DROID-SLAM/%s.h5"%i +\
" --resultpath=results/result_%s.txt"%i +\
" --calib=calib/kitti_360.txt" +\
" --stride=2" +\
" --max_factors=48" +\
" --active_window=12" +\
" --frontend_window=5" +\
" --frontend_radius=2" +\
" --frontend_nms=1" +\
" --inac_range=3" +\
" --visual_only=0" +\
" --far_threshold=-1" +\
" --translation_threshold=0.5" +\
" --mask_threshold=1.0" +\
" --skip_edge=[-4,-5,-6]" +\
" --save_pkl" +\
" --pklpath=results/%s.pkl"%i +\
" --show_plot",shell=True)
p.wait()
================================================
FILE: batch_subt.py
================================================
import os
import subprocess
for i in [\
# 'Handheld1_Folder',\
'Handheld2_Folder',\
]:
p = subprocess.Popen("python demo_vio_subt.py" +\
" --imagedir=/mnt/e/subt/%s/cam_0"%i +\
" --imagestamp=/mnt/e/subt/%s/cam_0/timestamps.txt"%i +\
" --imupath=/mnt/e/subt/%s/imu/imu_data.csv"%i +\
" --resultpath=results/result_%s.txt"%i +\
" --calib=calib/subt.txt" +\
" --stride=8" +\
" --max_factors=48" +\
" --active_window=12" +\
" --frontend_window=5" +\
" --frontend_radius=2" +\
" --frontend_nms=1" +\
" --far_threshold=0.02" +\
" --inac_range=3" +\
" --visual_only=0" +\
" --translation_threshold=0.2" +\
" --mask_threshold=-1.0" +\
" --skip_edge=[-4,-5,-6]" +\
" --save_pkl" +\
" --pklpath=results/%s.pkl"%i +\
" --show_plot",shell=True)
p.wait()
================================================
FILE: batch_tumvi.py
================================================
import os
import subprocess
for i in [\
'outdoors1',\
'outdoors2',\
'outdoors3',\
'outdoors4',\
'outdoors5',\
'outdoors6',\
'outdoors7',\
'outdoors8',\
'magistrale1',\
'magistrale2',\
'magistrale3',\
'magistrale4',\
'magistrale5',\
'magistrale6'
]:
p = subprocess.Popen("python demo_vio_tumvi.py" +\
" --imagedir=/mnt/z/tum-vi/dataset-%s_512_16/mav0/cam0/data"%i +\
" --imagestamp=/mnt/z/tum-vi/dataset-%s_512_16/mav0/cam0/data.csv"%i +\
" --imupath=/mnt/z/tum-vi/dataset-%s_512_16/mav0/imu0/data.csv"%i +\
" --gtpath=/mnt/z/tum-vi/dataset-%s_512_16/dso/gt_imu.csv"%i +\
# " --enable_h5" +\
# " --h5path=/home/zhouyuxuan/DROID-SLAM/%s.h5"%i +\
" --resultpath=results/result_%s.txt"%i +\
" --calib=calib/tumvi.txt" +\
" --stride=4" +\
" --max_factors=48" +\
" --active_window=12" +\
" --frontend_window=5" +\
" --frontend_radius=2" +\
" --frontend_nms=1" +\
" --far_threshold=0.02" +\
" --inac_range=3" +\
" --visual_only=0" +\
" --translation_threshold=0.2" +\
" --mask_threshold=-1.0" +\
" --skip_edge=[-4,-5,-6]" +\
" --save_pkl" +\
" --pklpath=results/%s.pkl"%i +\
" --show_plot",shell=True)
p.wait()
================================================
FILE: batch_whu.py
================================================
import os
import subprocess
# VIO
p = subprocess.Popen("python demo_vio_whu.py" +\
" --imagedir=/home/zhouyuxuan/data/WUH1012/cam0" +\
" --imagestamp=/home/zhouyuxuan/data/WUH1012/camstamp.txt" +\
" --imupath=/home/zhouyuxuan/data/WUH1012/imu.txt" +\
" --gtpath=/home/zhouyuxuan/data/WUH1012/IE.txt" +\
" --resultpath=results/result_whu_vio.txt" +\
" --calib=calib/1012.txt" +\
" --stride=2" +\
" --max_factors=48" +\
" --active_window=12" +\
" --frontend_window=5" +\
" --frontend_radius=2" +\
" --frontend_nms=1" +\
" --inac_range=3" +\
" --visual_only=0" +\
" --far_threshold=-1" +\
" --translation_threshold=0.25" +\
" --mask_threshold=0.0" +\
" --skip_edge=[]" +\
" --save_pkl" +\
" --use_zupt" +\
" --pklpath=results/whu.pkl" +\
" --show_plot",
shell=True)
p.wait()
# VIO + wheel
p = subprocess.Popen("python demo_vio_whu.py" +\
" --imagedir=/home/zhouyuxuan/data/WUH1012/cam0" +\
" --imagestamp=/home/zhouyuxuan/data/WUH1012/camstamp.txt" +\
" --imupath=/home/zhouyuxuan/data/WUH1012/imu.txt" +\
" --gtpath=/home/zhouyuxuan/data/WUH1012/IE.txt" +\
" --resultpath=results/result_whu_viow.txt" +\
" --calib=calib/1012.txt" +\
" --stride=2" +\
" --max_factors=48" +\
" --active_window=12" +\
" --frontend_window=5" +\
" --frontend_radius=2" +\
" --frontend_nms=1" +\
" --inac_range=3" +\
" --visual_only=0" +\
" --far_threshold=-1" +\
" --translation_threshold=0.25" +\
" --mask_threshold=0.0" +\
" --skip_edge=[]" +\
" --save_pkl" +\
" --use_odo" +\
" --odopath=/home/zhouyuxuan/data/WUH1012/odo_synthesis.txt" +\
" --pklpath=results/whu.pkl" +\
" --show_plot",
shell=True)
p.wait()
# VIO + GNSS
p = subprocess.Popen("python demo_vio_whu.py" +\
" --imagedir=/home/zhouyuxuan/data/WUH1012/cam0" +\
" --imagestamp=/home/zhouyuxuan/data/WUH1012/camstamp.txt" +\
" --imupath=/home/zhouyuxuan/data/WUH1012/imu.txt" +\
" --gtpath=/home/zhouyuxuan/data/WUH1012/IE.txt" +\
" --resultpath=results/result_whu_viog.txt" +\
" --calib=calib/1012.txt" +\
" --stride=2" +\
" --max_factors=48" +\
" --active_window=12" +\
" --frontend_window=5" +\
" --frontend_radius=2" +\
" --frontend_nms=1" +\
" --inac_range=3" +\
" --visual_only=0" +\
" --far_threshold=-1" +\
" --translation_threshold=0.25" +\
" --mask_threshold=0.0" +\
" --skip_edge=[]" +\
" --save_pkl" +\
" --use_gnss" +\
" --gnsspath=/home/zhouyuxuan/data/data_20221012103154/SEPT-PVT.flt" +\
" --pklpath=results/whu.pkl" +\
" --show_plot",
shell=True)
p.wait()
================================================
FILE: calib/0412.txt
================================================
889.32868436 889.32868436 515.73648834 202.43873596
================================================
FILE: calib/0412new.txt
================================================
885.839465 882.512623 505.509972 389.860117 -0.125551 0.065179 -0.000074 -0.000698
================================================
FILE: calib/1012.txt
================================================
890.21388839 889.56330572 512.88196119 381.38486858 -0.13095809 0.06640391 -0.00094794 0.0003389
================================================
FILE: calib/barn.txt
================================================
1161.545689 1161.545689 960.000000 540.000000 -0.025158 0.0 0.0 0.0
================================================
FILE: calib/carla.txt
================================================
886.8100 886.8100 512 256
================================================
FILE: calib/eth.txt
================================================
726.21081542969 726.21081542969 359.2048034668 202.47247314453
================================================
FILE: calib/euroc.txt
================================================
458.654 457.296 367.215 248.375 -0.28340811 0.07395907 0.00019359 1.76187114e-05
================================================
FILE: calib/handheld.txt
================================================
531.0895358407821 530.9183032386885 511.3708876141611 399.1276554093305 -0.3367182787437319 0.10679061024072911 0.0003055063102509499 0.0009756613499403765
================================================
FILE: calib/kitti_360.txt
================================================
788.629315 786.382230 687.158398 317.752196 -0.344441 0.141678 0.000414 -0.000222 -0.029608
================================================
FILE: calib/subt.txt
================================================
758.3153257832925 676.6492212772476 318.27111164892506 239.79816832491474 1.583106303248484 -0.059098218173967695 0.1793477408661115 0.0016819528105368057 -0.0005887999624264534
================================================
FILE: calib/tartan.txt
================================================
320.0 320.0 320.0 240.0
================================================
FILE: calib/tum3.txt
================================================
535.4 539.2 320.1 247.6
================================================
FILE: calib/tumvi.txt
================================================
190.97847715128717 190.9733070521226 254.93170605935475 256.8974428996504 0.0034823894022493434 0.0007150348452162257 -0.0020532361418706202 0.00020293673591811182
================================================
FILE: dataset/euroc_to_hdf5.py
================================================
from tqdm import tqdm
import numpy as np
import torch
import cv2
import os
import argparse
import h5py
import pickle
def show_image(image):
image = image.permute(1, 2, 0).cpu().numpy()
cv2.imshow('image', image / 255.0)
cv2.waitKey(1)
def image_stream(imagedir, imagestamp, h5path, calib, stride):
""" image generator """
calib = np.loadtxt(calib, delimiter=" ")
fx, fy, cx, cy = calib[:4]
K = np.eye(3)
K[0,0] = fx
K[0,2] = cx
K[1,1] = fy
K[1,2] = cy
Kn = np.eye(3)
Kn[0,0] = fx
Kn[0,2] = cx
Kn[1,1] = fy
Kn[1,2] = cy
image_list = sorted(os.listdir(imagedir))[::stride]
image_stamps = np.loadtxt(imagestamp,str,delimiter=',')
image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))
h5_f = h5py.File(h5path,'w')
for t, imfile in enumerate(image_list):
image = cv2.imread(os.path.join(imagedir, imfile))
if len(calib) > 4:
m1, m2 = cv2.initUndistortRectifyMap(K,calib[4:],np.eye(3),Kn,(image.shape[1],image.shape[0]),cv2.CV_32FC1)
image = cv2.remap(image, m1, m2, interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_CONSTANT)
tt = float(image_dict[imfile]) /1e9
h0, w0, _ = image.shape
h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))
w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))
image = cv2.resize(image, (w1, h1))
image = image[:h1-h1%8, :w1-w1%8]
image = torch.as_tensor(image).permute(2, 0, 1)
intrinsics = torch.as_tensor([fx, fy, cx, cy ])
intrinsics[0::2] *= (w1 / w0)
intrinsics[1::2] *= (h1 / h0)
h5_f.create_dataset('%.10f'%tt,data = np.fromstring(pickle.dumps((tt, image[None], intrinsics)),dtype='uint8'))
yield tt, image[None], intrinsics
h5_f.close()
if __name__ == '__main__':
print(torch.cuda.device_count())
print(torch.cuda.is_available())
print(torch.cuda.current_device())
parser = argparse.ArgumentParser()
parser.add_argument("--imagedir", type=str, help="path to image directory")
parser.add_argument("--imagestamp", type=str, help="")
parser.add_argument("--h5path", type=str, help="")
parser.add_argument("--calib", type=str, help="path to calibration file")
parser.add_argument("--stride", default=4, type=int, help="frame stride")
parser.add_argument("--show_plot", action="store_true", help="")
args = parser.parse_args()
for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp,\
args.h5path, args.calib, args.stride)):
if args.show_plot:
show_image(image[0])
================================================
FILE: dataset/kitti360_to_hdf5.py
================================================
from tqdm import tqdm
import numpy as np
import torch
import cv2
import os
import argparse
import h5py
import pickle
def show_image(image):
image = image.permute(1, 2, 0).cpu().numpy()
cv2.imshow('image', image / 255.0)
cv2.waitKey(1)
def image_stream(imagedir, imagestamp, h5path, calib, stride):
""" image generator """
calib = np.loadtxt(calib, delimiter=" ")
fx, fy, cx, cy = calib[:4]
K = np.eye(3)
K[0,0] = fx
K[0,2] = cx
K[1,1] = fy
K[1,2] = cy
Kn = np.eye(3)
Kn[0,0] = fx
Kn[0,2] = cx
Kn[1,1] = fy
Kn[1,2] = cy
image_list = sorted(os.listdir(imagedir))[::stride]
image_stamps = np.loadtxt(imagestamp,str)
image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))
h5_f = h5py.File(h5path,'w')
for t, imfile in enumerate(image_list):
image = cv2.imread(os.path.join(imagedir, imfile))
if len(calib) > 4:
image = cv2.undistort(image, K, calib[4:])
tt = float(image_dict[imfile])
h0, w0, _ = image.shape
h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))
w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))
image = cv2.resize(image, (w1, h1))
image = image[:h1-h1%8, :w1-w1%8]
image = torch.as_tensor(image).permute(2, 0, 1)
intrinsics = torch.as_tensor([fx, fy, cx, cy])
intrinsics[0::2] *= (w1 / w0)
intrinsics[1::2] *= (h1 / h0)
h5_f.create_dataset('%.10f'%tt,data = np.fromstring(pickle.dumps((tt, image[None], intrinsics)),dtype='uint8'))
yield tt, image[None], intrinsics
h5_f.close()
if __name__ == '__main__':
print(torch.cuda.device_count())
print(torch.cuda.is_available())
print(torch.cuda.current_device())
parser = argparse.ArgumentParser()
parser.add_argument("--imagedir", type=str, help="path to image directory")
parser.add_argument("--imagestamp", type=str, help="")
parser.add_argument("--h5path", type=str, help="")
parser.add_argument("--calib", type=str, help="path to calibration file")
parser.add_argument("--stride", default=2, type=int, help="frame stride")
parser.add_argument("--show_plot", action="store_true", help="")
args = parser.parse_args()
for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp,\
args.h5path, args.calib, args.stride)):
if args.show_plot:
show_image(image[0])
================================================
FILE: dataset/tumvi_to_hdf5.py
================================================
from tqdm import tqdm
import numpy as np
import torch
import cv2
import os
import argparse
import h5py
import pickle
def show_image(image):
image = image.permute(1, 2, 0).cpu().numpy()
cv2.imshow('image', image / 255.0)
cv2.waitKey(1)
def image_stream(imagedir, imagestamp, h5path, calib, stride):
""" image generator """
calib = np.loadtxt(calib, delimiter=" ")
fx, fy, cx, cy = calib[:4]
K = np.eye(3)
K[0,0] = fx
K[0,2] = cx
K[1,1] = fy
K[1,2] = cy
Kn = np.eye(3)
Kn[0,0] = fx
Kn[0,2] = cx
Kn[1,1] = fy
Kn[1,2] = cy
image_list = sorted(os.listdir(imagedir))[::stride]
image_stamps = np.loadtxt(imagestamp,str,delimiter=',')
image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))
h5_f = h5py.File(h5path,'w')
for t, imfile in enumerate(image_list):
image = cv2.imread(os.path.join(imagedir, imfile))
if len(calib) > 4:
m1, m2 = cv2.fisheye.initUndistortRectifyMap(K,calib[4:],np.eye(3),Kn,(512,512),cv2.CV_32FC1)
image = cv2.remap(image, m1, m2, interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_CONSTANT)
tt = float(image_dict[imfile]) /1e9
h0, w0, _ = image.shape
h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))
w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))
image = cv2.resize(image, (w1, h1))
image = image[:h1-h1%8, :w1-w1%8]
image = torch.as_tensor(image).permute(2, 0, 1)
intrinsics = torch.as_tensor([fx, fy, cx, cy ])
intrinsics[0::2] *= (w1 / w0)
intrinsics[1::2] *= (h1 / h0)
h5_f.create_dataset('%.10f'%tt,data = np.fromstring(pickle.dumps((tt, image[None], intrinsics)),dtype='uint8'))
yield tt, image[None], intrinsics
h5_f.close()
if __name__ == '__main__':
print(torch.cuda.device_count())
print(torch.cuda.is_available())
print(torch.cuda.current_device())
parser = argparse.ArgumentParser()
parser.add_argument("--imagedir", type=str, help="path to image directory")
parser.add_argument("--imagestamp", type=str, help="")
parser.add_argument("--h5path", type=str, help="")
parser.add_argument("--calib", type=str, help="path to calibration file")
parser.add_argument("--stride", default=4, type=int, help="frame stride")
parser.add_argument("--show_plot", action="store_true", help="")
args = parser.parse_args()
for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp,\
args.h5path, args.calib, args.stride)):
if args.show_plot:
show_image(image[0])
================================================
FILE: dbaf/covisible_graph.py
================================================
import torch
import lietorch
import numpy as np
import matplotlib.pyplot as plt
from lietorch import SE3
from modules.corr import CorrBlock, AltCorrBlock
import geom.projective_ops as pops
import matplotlib.pyplot as plt
import cv2
from depth_video import DepthVideo
import matplotlib.cm as cm
import matplotlib
class CovisibleGraph:
def __init__(self, video: DepthVideo, update_op, device="cuda:0", corr_impl="volume", args = None):
self.video = video
self.update_op = update_op
self.device = device
self.max_factors = args.max_factors
self.corr_impl = corr_impl
self.upsample = args.upsample
# operator at 1/8 resolution
self.ht = ht = video.ht // 8
self.wd = wd = video.wd // 8
self.coords0 = pops.coords_grid(ht, wd, device=device)
self.ii = torch.as_tensor([], dtype=torch.long, device=device)
self.jj = torch.as_tensor([], dtype=torch.long, device=device)
self.age = torch.as_tensor([], dtype=torch.long, device=device)
self.corr, self.net, self.inp = None, None, None
self.damping = 1e-6 * torch.ones_like(self.video.disps)
self.target = torch.zeros([1, 0, ht, wd, 2], device=device, dtype=torch.float)
self.weight = torch.zeros([1, 0, ht, wd, 2], device=device, dtype=torch.float)
# inactive factors
self.ii_inac = torch.as_tensor([], dtype=torch.long, device=device)
self.jj_inac = torch.as_tensor([], dtype=torch.long, device=device)
self.ii_bad = torch.as_tensor([], dtype=torch.long, device=device)
self.jj_bad = torch.as_tensor([], dtype=torch.long, device=device)
self.target_inac = torch.zeros([1, 0, ht, wd, 2], device=device, dtype=torch.float)
self.weight_inac = torch.zeros([1, 0, ht, wd, 2], device=device, dtype=torch.float)
self.far_threshold = args.far_threshold
self.inac_range = args.inac_range
self.mask_threshold = args.mask_threshold
self.img_count = 0
self.skip_edge = args.skip_edge
self.frontend_window = args.frontend_window
# simple online visualization
self.show_covisible_graph = False
self.show_oldest_disparity = False
self.show_flow_and_weight = False
def __filter_repeated_edges(self, ii, jj):
""" remove duplicate edges """
keep = torch.zeros(ii.shape[0], dtype=torch.bool, device=ii.device)
eset = set(
[(i.item(), j.item()) for i, j in zip(self.ii, self.jj)] +
[(i.item(), j.item()) for i, j in zip(self.ii_inac, self.jj_inac)])
for k, (i, j) in enumerate(zip(ii, jj)):
keep[k] = (i.item(), j.item()) not in eset
return ii[keep], jj[keep]
def print_edges(self):
ii = self.ii.cpu().numpy()
jj = self.jj.cpu().numpy()
ix = np.argsort(ii)
ii = ii[ix]
jj = jj[ix]
w = torch.mean(self.weight, dim=[0,2,3,4]).cpu().numpy()
w = w[ix]
for e in zip(ii, jj, w):
print(e)
print()
def filter_edges(self):
""" remove bad edges """
conf = torch.mean(self.weight, dim=[0,2,3,4])
mask = (torch.abs(self.ii-self.jj) > 2) & (conf < 0.001)
self.ii_bad = torch.cat([self.ii_bad, self.ii[mask]])
self.jj_bad = torch.cat([self.jj_bad, self.jj[mask]])
self.rm_factors(mask, store=False)
def clear_edges(self):
self.rm_factors(self.ii >= 0)
self.net = None
self.inp = None
@torch.cuda.amp.autocast(enabled=True)
def add_factors(self, ii, jj, remove=False):
""" add edges to factor graph """
if not isinstance(ii, torch.Tensor):
ii = torch.as_tensor(ii, dtype=torch.long, device=self.device)
if not isinstance(jj, torch.Tensor):
jj = torch.as_tensor(jj, dtype=torch.long, device=self.device)
# remove duplicate edges
ii, jj = self.__filter_repeated_edges(ii, jj)
if ii.shape[0] == 0:
return
# place limit on number of factors
if self.max_factors > 0 and self.ii.shape[0] + ii.shape[0] > self.max_factors \
and self.corr is not None and remove:
ix = torch.arange(len(self.age))[torch.argsort(self.age).cpu()]
self.rm_factors(ix >= self.max_factors - ii.shape[0], store=True)
net = self.video.nets[ii].to(self.device).unsqueeze(0)
# correlation volume for new edges
if self.corr_impl == "volume":
c = (ii == jj).long()
fmap1 = self.video.fmaps[ii,0].to(self.device).unsqueeze(0)
fmap2 = self.video.fmaps[jj,c].to(self.device).unsqueeze(0)
corr = CorrBlock(fmap1, fmap2)
self.corr = corr if self.corr is None else self.corr.cat(corr)
inp = self.video.inps[ii].to(self.device).unsqueeze(0)
self.inp = inp if self.inp is None else torch.cat([self.inp, inp], 1)
with torch.cuda.amp.autocast(enabled=False):
target, _ = self.video.reproject(ii, jj)
weight = torch.zeros_like(target)
self.ii = torch.cat([self.ii, ii], 0)
self.jj = torch.cat([self.jj, jj], 0)
self.age = torch.cat([self.age, torch.zeros_like(ii)], 0)
# reprojection factors
self.net = net if self.net is None else torch.cat([self.net, net], 1)
self.target = torch.cat([self.target, target], 1)
self.weight = torch.cat([self.weight, weight], 1)
@torch.cuda.amp.autocast(enabled=True)
def rm_factors(self, mask, store=False):
""" drop edges from factor graph """
# store estimated factors
if store:
self.ii_inac = torch.cat([self.ii_inac, self.ii[mask]], 0)
self.jj_inac = torch.cat([self.jj_inac, self.jj[mask]], 0)
self.target_inac = torch.cat([self.target_inac, self.target[:,mask]], 1)
self.weight_inac = torch.cat([self.weight_inac, self.weight[:,mask]], 1)
self.ii = self.ii[~mask]
self.jj = self.jj[~mask]
self.age = self.age[~mask]
if self.corr_impl == "volume":
self.corr = self.corr[~mask]
if self.net is not None:
self.net = self.net[:,~mask]
if self.inp is not None:
self.inp = self.inp[:,~mask]
self.target = self.target[:,~mask]
self.weight = self.weight[:,~mask]
@torch.cuda.amp.autocast(enabled=True)
def rm_keyframe(self, ix):
""" drop edges from factor graph """
with self.video.get_lock():
self.video.images[ix] = self.video.images[ix+1]
self.video.poses[ix] = self.video.poses[ix+1]
self.video.disps[ix] = self.video.disps[ix+1]
self.video.disps_sens[ix] = self.video.disps_sens[ix+1]
self.video.intrinsics[ix] = self.video.intrinsics[ix+1]
self.video.nets[ix] = self.video.nets[ix+1]
self.video.inps[ix] = self.video.inps[ix+1]
self.video.fmaps[ix] = self.video.fmaps[ix+1]
self.video.tstamp[ix] = self.video.tstamp[ix+1] # BUG fix
m = (self.ii_inac == ix) | (self.jj_inac == ix)
self.ii_inac[self.ii_inac >= ix] -= 1
self.jj_inac[self.jj_inac >= ix] -= 1
if torch.any(m):
self.ii_inac = self.ii_inac[~m]
self.jj_inac = self.jj_inac[~m]
self.target_inac = self.target_inac[:,~m]
self.weight_inac = self.weight_inac[:,~m]
m = (self.ii == ix) | (self.jj == ix)
self.ii[self.ii >= ix] -= 1
self.jj[self.jj >= ix] -= 1
self.rm_factors(m, store=False)
@torch.cuda.amp.autocast(enabled=True)
def update(self, t0=None, t1=None, itrs=2, use_inactive=False, EP=1e-7, motion_only=False, marg = False):
""" run update operator on factor graph """
self.video.logger.info('update')
with torch.cuda.amp.autocast(enabled=False):
coords1, mask = self.video.reproject(self.ii, self.jj)
motn = torch.cat([coords1 - self.coords0, self.target - coords1], dim=-1)
motn = motn.permute(0,1,4,2,3).clamp(-64.0, 64.0) # 1,2,4,48,96
corr = self.corr(coords1)
self.net, delta, weight, damping, upmask = \
self.update_op(self.net, self.inp, corr, motn, self.ii, self.jj, self.upsample)
if t0 is None:
t0 = max(1, self.ii.min().item()+1)
self.video.logger.info('ba')
with torch.cuda.amp.autocast(enabled=False):
self.target = coords1 + delta.to(dtype=torch.float)
self.weight = weight.to(dtype=torch.float)
ht, wd = self.coords0.shape[0:2]
if self.upsample:
self.damping[torch.unique(self.ii)] = damping
if use_inactive:
m = (self.ii_inac >= t0 - self.inac_range) & (self.jj_inac >= t0 - self.inac_range)
ii = torch.cat([self.ii_inac[m], self.ii], 0)
jj = torch.cat([self.jj_inac[m], self.jj], 0)
target = torch.cat([self.target_inac[:,m], self.target], 1)
weight = torch.cat([self.weight_inac[:,m], self.weight], 1)
else:
ii, jj, target, weight = self.ii, self.jj, self.target, self.weight
# Some real-time visualization for debugging
# 1) Disparity
if self.show_oldest_disparity:
disp_show_front = self.video.disps[self.ii[0]].cpu().numpy()
disp_show_front = cv2.resize(disp_show_front,[disp_show_front.shape[1]*8,disp_show_front.shape[0]*8],interpolation = cv2.INTER_NEAREST)
disp_show_front= disp_show_front.astype(np.float32)
normalizer = matplotlib.colors.Normalize(vmin=-0.2, vmax=1.0)
mapper = cm.ScalarMappable(norm=normalizer,cmap='magma')
colormapped_im = (mapper.to_rgba(disp_show_front)[:, :, :3] * 255).astype(np.uint8)
colormapped_im = cv2.cvtColor(colormapped_im,cv2.COLOR_RGB2BGR)
cv2.imshow('colormapped_im',colormapped_im)
cv2.waitKey(1)
# 2) Optical flow and weight
if self.show_flow_and_weight:
rgb = self.video.images[torch.max(self.ii)].cpu().numpy().transpose(1,2,0)
new_flow_id = torch.where(torch.logical_and(self.ii==torch.max(self.ii),self.jj==torch.max(self.ii)-5))[0][0].item()
weight_cpu = weight[0,new_flow_id].cpu().numpy().astype(np.float32)
weight_cpu = np.linalg.norm(weight_cpu,axis=2)
normalizer = matplotlib.colors.Normalize(vmin=-0.0, vmax=1.5)
mapper = cm.ScalarMappable(norm=normalizer,cmap='jet')
colormapped_im = (mapper.to_rgba(weight_cpu)[:, :, :3] * 255).astype(np.uint8)
colormapped_im = cv2.cvtColor(colormapped_im,cv2.COLOR_RGB2BGR)
colormapped_im = cv2.resize(colormapped_im,[rgb.shape[1],rgb.shape[0]])
colormapped_im = cv2.addWeighted(rgb,0.5,colormapped_im,0.5,0)
absflow = (self.target[0,new_flow_id] - self.coords0).cpu().numpy()
for iii in range(0,absflow.shape[0],4):
for jjj in range(0,absflow.shape[1],4):
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)
cv2.imshow('weight_cpu',colormapped_im)
self.img_count += 1
# 3) Covisible graph
if self.show_covisible_graph:
i0 = min(ii)
i1 = max(ii)
ppp = SE3(self.video.poses[i0:(i1+1)]).inv().matrix()[:,0:3,3].cpu().numpy()
# [:,:3].cpu().numpy()
scale = max(max(ppp[:,0]) - min(ppp[:,0]),max(ppp[:,1]) - min(ppp[:,1]))
ppp[:,0] -= np.mean(ppp[:,0])
ppp[:,1] = -(ppp[:,1]- np.mean(ppp[:,1]))
ppp *= max(round(1/scale * 200 / 50)*50,50)
ppp += 500
mmm = np.zeros([1000,1000],dtype=np.uint8)
for iii in range(0,i1+1-i0):
mmm = cv2.circle(mmm,(int(round(ppp[iii,0])),int(round(ppp[iii,1]))),4,255,0)
for iii in range(self.ii_inac[m].shape[0]):
iiii = self.ii_inac[m][iii]-i0
jjjj = self.jj_inac[m][iii]-i0
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)
for iii in range(self.ii.shape[0]):
iiii = self.ii[iii]-i0
jjjj = self.jj[iii]-i0
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)
cv2.imshow('window',mmm)
## Tricks for better performance
# 1) downweight far points
if self.far_threshold > 0 and self.video.imu_enabled:
disp_mask = (self.video.disps < self.far_threshold)
mask = disp_mask[ii, :, :]
weight[:, mask] /= 1000.0
# 2) downweight far points
if self.mask_threshold > 0 and self.video.imu_enabled:
pose0 = SE3(self.video.poses[ii])
pose1 = SE3(self.video.poses[jj])
pose01 = pose0*pose1.inv()
mask = torch.norm(pose01.translation()[:,:3],dim=1) < self.mask_threshold
weight[:,mask,:,:,:] /= 1000.0
# 3) downweight edges related to the newest frame
downweight_newframe = True
if downweight_newframe:
weight[:,ii==max(ii)] /= 10.0
weight[:,jj==max(jj)] /= 4.0
damping = .2 * self.damping[torch.unique(ii)].contiguous() + EP
target = target.view(-1, ht, wd, 2).permute(0,3,1,2).contiguous()
weight = weight.view(-1, ht, wd, 2).permute(0,3,1,2).contiguous()
# Dense bundle adjustment
self.video.ba(target, weight, damping, ii, jj, t0, t1,
itrs=itrs, lm=1e-4, ep=0.1, motion_only=motion_only)
if self.upsample:
self.video.upsample(torch.unique(self.ii), upmask)
self.age += 1
def add_neighborhood_factors(self, t0, t1, r=3):
""" add edges between neighboring frames within radius r """
ii, jj = torch.meshgrid(torch.arange(t0,t1), torch.arange(t0,t1))
ii = ii.reshape(-1).to(dtype=torch.long, device=self.device)
jj = jj.reshape(-1).to(dtype=torch.long, device=self.device)
c = 1 if self.video.stereo else 0
keep = ((ii - jj).abs() > c) & ((ii - jj).abs() <= r)
self.add_factors(ii[keep], jj[keep])
def add_proximity_factors(self, t0=0, t1=0, rad=2, nms=2, beta=0.25, thresh=16.0, remove=False):
""" add edges to the factor graph based on distance """
t = self.video.counter.value
ix = torch.arange(t0, t)
jx = torch.arange(t1, t)
ii, jj = torch.meshgrid(ix, jx)
ii = ii.reshape(-1)
jj = jj.reshape(-1)
cc = ii.shape[0]
# Opportunistic "skip" edges in the graph
if self.skip_edge:
if torch.max(ii) - torch.min(ii) == self.frontend_window - 1:
jj_add = torch.min(ii) + torch.tensor(self.skip_edge)
jj_add = jj_add[jj_add>0]
ii_add = torch.zeros_like(jj_add) + torch.max(ii)
jj = torch.cat([jj,jj_add])
ii = torch.cat([ii,ii_add])
d = self.video.distance(ii, jj, beta=beta)
d[ii - rad < jj] = np.inf
d[d > 100] = np.inf
ii1 = torch.cat([self.ii, self.ii_bad, self.ii_inac], 0)
jj1 = torch.cat([self.jj, self.jj_bad, self.jj_inac], 0)
for i, j in zip(ii1.cpu().numpy(), jj1.cpu().numpy()):
for di in range(-nms, nms+1):
for dj in range(-nms, nms+1):
if abs(di) + abs(dj) <= max(min(abs(i-j)-2, nms), 0):
i1 = i + di
j1 = j + dj
if (t0 <= i1 < t) and (t1 <= j1 < t):
d[(i1-t0)*(t-t1) + (j1-t1)] = np.inf
es = []
for i in range(t0, t):
if self.video.stereo:
es.append((i, i))
d[(i-t0)*(t-t1) + (i-t1)] = np.inf
for j in range(max(i-rad-1,0), i):
es.append((i,j))
es.append((j,i))
if (i-t0)*(t-t1) + (j-t1) >=0:
d[(i-t0)*(t-t1) + (j-t1)] = np.inf
ix = torch.argsort(d)
for k in ix:
if k >= cc:
continue
if d[k].item() > thresh:
continue
if len(es) > self.max_factors:
break
i = ii[k]
j = jj[k]
# bidirectional
es.append((i, j))
es.append((j, i))
for di in range(-nms, nms+1):
for dj in range(-nms, nms+1):
if abs(di) + abs(dj) <= max(min(abs(i-j)-2, nms), 0):
i1 = i + di
j1 = j + dj
if (t0 <= i1 < t) and (t1 <= j1 < t):
d[(i1-t0)*(t-t1) + (j1-t1)] = np.inf
if ii.shape[0] > cc:
ix = torch.argsort(d[cc:ii.shape[0]])
if d[cc + ix[0]] < thresh and d[cc + ix[0]] > 0:
es.append((ii[cc+ix[0]],jj[cc+ix[0]]))
es.append((jj[cc+ix[0]],ii[cc+ix[0]]))
ii, jj = torch.as_tensor(es, device=self.device).unbind(dim=-1)
self.add_factors(ii, jj, remove)
================================================
FILE: dbaf/data_readers/__init__.py
================================================
================================================
FILE: dbaf/data_readers/augmentation.py
================================================
import torch
import torchvision.transforms as transforms
import numpy as np
import torch.nn.functional as F
class RGBDAugmentor:
""" perform augmentation on RGB-D video """
def __init__(self, crop_size):
self.crop_size = crop_size
self.augcolor = transforms.Compose([
transforms.ToPILImage(),
transforms.ColorJitter(brightness=0.25, contrast=0.25, saturation=0.25, hue=0.4/3.14),
transforms.RandomGrayscale(p=0.1),
transforms.ToTensor()])
self.max_scale = 0.25
def spatial_transform(self, images, depths, poses, intrinsics):
""" cropping and resizing """
ht, wd = images.shape[2:]
max_scale = self.max_scale
min_scale = np.log2(np.maximum(
(self.crop_size[0] + 1) / float(ht),
(self.crop_size[1] + 1) / float(wd)))
scale = 2 ** np.random.uniform(min_scale, max_scale)
intrinsics = scale * intrinsics
depths = depths.unsqueeze(dim=1)
images = F.interpolate(images, scale_factor=scale, mode='bilinear',
align_corners=False, recompute_scale_factor=True)
depths = F.interpolate(depths, scale_factor=scale, recompute_scale_factor=True)
# always perform center crop (TODO: try non-center crops)
y0 = (images.shape[2] - self.crop_size[0]) // 2
x0 = (images.shape[3] - self.crop_size[1]) // 2
intrinsics = intrinsics - torch.tensor([0.0, 0.0, x0, y0])
images = images[:, :, y0:y0+self.crop_size[0], x0:x0+self.crop_size[1]]
depths = depths[:, :, y0:y0+self.crop_size[0], x0:x0+self.crop_size[1]]
depths = depths.squeeze(dim=1)
return images, poses, depths, intrinsics
def color_transform(self, images):
""" color jittering """
num, ch, ht, wd = images.shape
images = images.permute(1, 2, 3, 0).reshape(ch, ht, wd*num)
images = 255 * self.augcolor(images[[2,1,0]] / 255.0)
return images[[2,1,0]].reshape(ch, ht, wd, num).permute(3,0,1,2).contiguous()
def __call__(self, images, poses, depths, intrinsics):
images = self.color_transform(images)
return self.spatial_transform(images, depths, poses, intrinsics)
================================================
FILE: dbaf/data_readers/base.py
================================================
import numpy as np
import torch
import torch.utils.data as data
import torch.nn.functional as F
import csv
import os
import cv2
import math
import random
import json
import pickle
import os.path as osp
from .augmentation import RGBDAugmentor
from .rgbd_utils import *
class RGBDDataset(data.Dataset):
def __init__(self, name, datapath, n_frames=4, crop_size=[384,512], fmin=8.0, fmax=75.0, do_aug=True):
""" Base class for RGBD dataset """
self.aug = None
self.root = datapath
self.name = name
self.n_frames = n_frames
self.fmin = fmin # exclude very easy examples
self.fmax = fmax # exclude very hard examples
if do_aug:
self.aug = RGBDAugmentor(crop_size=crop_size)
# building dataset is expensive, cache so only needs to be performed once
cur_path = osp.dirname(osp.abspath(__file__))
if not os.path.isdir(osp.join(cur_path, 'cache')):
os.mkdir(osp.join(cur_path, 'cache'))
cache_path = osp.join(cur_path, 'cache', '{}.pickle'.format(self.name))
if osp.isfile(cache_path):
scene_info = pickle.load(open(cache_path, 'rb'))[0]
else:
scene_info = self._build_dataset()
with open(cache_path, 'wb') as cachefile:
pickle.dump((scene_info,), cachefile)
self.scene_info = scene_info
self._build_dataset_index()
def _build_dataset_index(self):
self.dataset_index = []
for scene in self.scene_info:
if not self.__class__.is_test_scene(scene):
graph = self.scene_info[scene]['graph']
for i in graph:
if len(graph[i][0]) > self.n_frames:
self.dataset_index.append((scene, i))
else:
print("Reserving {} for validation".format(scene))
@staticmethod
def image_read(image_file):
return cv2.imread(image_file)
@staticmethod
def depth_read(depth_file):
return np.load(depth_file)
def build_frame_graph(self, poses, depths, intrinsics, f=16, max_flow=256):
""" compute optical flow distance between all pairs of frames """
def read_disp(fn):
depth = self.__class__.depth_read(fn)[f//2::f, f//2::f]
depth[depth < 0.01] = np.mean(depth)
return 1.0 / depth
poses = np.array(poses)
intrinsics = np.array(intrinsics) / f
disps = np.stack(list(map(read_disp, depths)), 0)
d = f * compute_distance_matrix_flow(poses, disps, intrinsics)
# uncomment for nice visualization
# import matplotlib.pyplot as plt
# plt.imshow(d)
# plt.show()
graph = {}
for i in range(d.shape[0]):
j, = np.where(d[i] < max_flow)
graph[i] = (j, d[i,j])
return graph
def __getitem__(self, index):
""" return training video """
index = index % len(self.dataset_index)
scene_id, ix = self.dataset_index[index]
frame_graph = self.scene_info[scene_id]['graph']
images_list = self.scene_info[scene_id]['images']
depths_list = self.scene_info[scene_id]['depths']
poses_list = self.scene_info[scene_id]['poses']
intrinsics_list = self.scene_info[scene_id]['intrinsics']
inds = [ ix ]
while len(inds) < self.n_frames:
# get other frames within flow threshold
k = (frame_graph[ix][1] > self.fmin) & (frame_graph[ix][1] < self.fmax)
frames = frame_graph[ix][0][k]
# prefer frames forward in time
if np.count_nonzero(frames[frames > ix]):
ix = np.random.choice(frames[frames > ix])
elif np.count_nonzero(frames):
ix = np.random.choice(frames)
inds += [ ix ]
images, depths, poses, intrinsics = [], [], [], []
for i in inds:
images.append(self.__class__.image_read(images_list[i]))
depths.append(self.__class__.depth_read(depths_list[i]))
poses.append(poses_list[i])
intrinsics.append(intrinsics_list[i])
images = np.stack(images).astype(np.float32)
depths = np.stack(depths).astype(np.float32)
poses = np.stack(poses).astype(np.float32)
intrinsics = np.stack(intrinsics).astype(np.float32)
images = torch.from_numpy(images).float()
images = images.permute(0, 3, 1, 2)
disps = torch.from_numpy(1.0 / depths)
poses = torch.from_numpy(poses)
intrinsics = torch.from_numpy(intrinsics)
if self.aug is not None:
images, poses, disps, intrinsics = \
self.aug(images, poses, disps, intrinsics)
# scale scene
if len(disps[disps>0.01]) > 0:
s = disps[disps>0.01].mean()
disps = disps / s
poses[...,:3] *= s
return images, poses, disps, intrinsics
def __len__(self):
return len(self.dataset_index)
def __imul__(self, x):
self.dataset_index *= x
return self
================================================
FILE: dbaf/data_readers/factory.py
================================================
import pickle
import os
import os.path as osp
# RGBD-Dataset
from .tartan import TartanAir
from .stream import ImageStream
from .stream import StereoStream
from .stream import RGBDStream
# streaming datasets for inference
from .tartan import TartanAirStream
from .tartan import TartanAirTestStream
def dataset_factory(dataset_list, **kwargs):
""" create a combined dataset """
from torch.utils.data import ConcatDataset
dataset_map = { 'tartan': (TartanAir, ) }
db_list = []
for key in dataset_list:
# cache datasets for faster future loading
db = dataset_map[key][0](**kwargs)
print("Dataset {} has {} images".format(key, len(db)))
db_list.append(db)
return ConcatDataset(db_list)
def create_datastream(dataset_path, **kwargs):
""" create data_loader to stream images 1 by 1 """
from torch.utils.data import DataLoader
if osp.isfile(osp.join(dataset_path, 'calibration.txt')):
db = ETH3DStream(dataset_path, **kwargs)
elif osp.isdir(osp.join(dataset_path, 'image_left')):
db = TartanAirStream(dataset_path, **kwargs)
elif osp.isfile(osp.join(dataset_path, 'rgb.txt')):
db = TUMStream(dataset_path, **kwargs)
elif osp.isdir(osp.join(dataset_path, 'mav0')):
db = EurocStream(dataset_path, **kwargs)
elif osp.isfile(osp.join(dataset_path, 'calib.txt')):
db = KITTIStream(dataset_path, **kwargs)
else:
# db = TartanAirStream(dataset_path, **kwargs)
db = TartanAirTestStream(dataset_path, **kwargs)
stream = DataLoader(db, shuffle=False, batch_size=1, num_workers=4)
return stream
def create_imagestream(dataset_path, **kwargs):
""" create data_loader to stream images 1 by 1 """
from torch.utils.data import DataLoader
db = ImageStream(dataset_path, **kwargs)
return DataLoader(db, shuffle=False, batch_size=1, num_workers=4)
def create_stereostream(dataset_path, **kwargs):
""" create data_loader to stream images 1 by 1 """
from torch.utils.data import DataLoader
db = StereoStream(dataset_path, **kwargs)
return DataLoader(db, shuffle=False, batch_size=1, num_workers=4)
def create_rgbdstream(dataset_path, **kwargs):
""" create data_loader to stream images 1 by 1 """
from torch.utils.data import DataLoader
db = RGBDStream(dataset_path, **kwargs)
return DataLoader(db, shuffle=False, batch_size=1, num_workers=4)
================================================
FILE: dbaf/data_readers/rgbd_utils.py
================================================
import numpy as np
import os.path as osp
import torch
from lietorch import SE3
import geom.projective_ops as pops
from scipy.spatial.transform import Rotation
def parse_list(filepath, skiprows=0):
""" read list data """
data = np.loadtxt(filepath, delimiter=' ', dtype=np.unicode_, skiprows=skiprows)
return data
def associate_frames(tstamp_image, tstamp_depth, tstamp_pose, max_dt=1.0):
""" pair images, depths, and poses """
associations = []
for i, t in enumerate(tstamp_image):
if tstamp_pose is None:
j = np.argmin(np.abs(tstamp_depth - t))
if (np.abs(tstamp_depth[j] - t) < max_dt):
associations.append((i, j))
else:
j = np.argmin(np.abs(tstamp_depth - t))
k = np.argmin(np.abs(tstamp_pose - t))
if (np.abs(tstamp_depth[j] - t) < max_dt) and \
(np.abs(tstamp_pose[k] - t) < max_dt):
associations.append((i, j, k))
return associations
def loadtum(datapath, frame_rate=-1):
""" read video data in tum-rgbd format """
if osp.isfile(osp.join(datapath, 'groundtruth.txt')):
pose_list = osp.join(datapath, 'groundtruth.txt')
elif osp.isfile(osp.join(datapath, 'pose.txt')):
pose_list = osp.join(datapath, 'pose.txt')
else:
return None, None, None, None
image_list = osp.join(datapath, 'rgb.txt')
depth_list = osp.join(datapath, 'depth.txt')
calib_path = osp.join(datapath, 'calibration.txt')
intrinsic = None
if osp.isfile(calib_path):
intrinsic = np.loadtxt(calib_path, delimiter=' ')
intrinsic = intrinsic.astype(np.float64)
image_data = parse_list(image_list)
depth_data = parse_list(depth_list)
pose_data = parse_list(pose_list, skiprows=1)
pose_vecs = pose_data[:,1:].astype(np.float64)
tstamp_image = image_data[:,0].astype(np.float64)
tstamp_depth = depth_data[:,0].astype(np.float64)
tstamp_pose = pose_data[:,0].astype(np.float64)
associations = associate_frames(tstamp_image, tstamp_depth, tstamp_pose)
# print(len(tstamp_image))
# print(len(associations))
indicies = range(len(associations))[::5]
# indicies = [ 0 ]
# for i in range(1, len(associations)):
# t0 = tstamp_image[associations[indicies[-1]][0]]
# t1 = tstamp_image[associations[i][0]]
# if t1 - t0 > 1.0 / frame_rate:
# indicies += [ i ]
images, poses, depths, intrinsics, tstamps = [], [], [], [], []
for ix in indicies:
(i, j, k) = associations[ix]
images += [ osp.join(datapath, image_data[i,1]) ]
depths += [ osp.join(datapath, depth_data[j,1]) ]
poses += [ pose_vecs[k] ]
tstamps += [ tstamp_image[i] ]
if intrinsic is not None:
intrinsics += [ intrinsic ]
return images, depths, poses, intrinsics, tstamps
def all_pairs_distance_matrix(poses, beta=2.5):
""" compute distance matrix between all pairs of poses """
poses = np.array(poses, dtype=np.float32)
poses[:,:3] *= beta # scale to balence rot + trans
poses = SE3(torch.from_numpy(poses))
r = (poses[:,None].inv() * poses[None,:]).log()
return r.norm(dim=-1).cpu().numpy()
def pose_matrix_to_quaternion(pose):
""" convert 4x4 pose matrix to (t, q) """
q = Rotation.from_matrix(pose[:3, :3]).as_quat()
return np.concatenate([pose[:3, 3], q], axis=0)
def compute_distance_matrix_flow(poses, disps, intrinsics):
""" compute flow magnitude between all pairs of frames """
if not isinstance(poses, SE3):
poses = torch.from_numpy(poses).float().cuda()[None]
poses = SE3(poses).inv()
disps = torch.from_numpy(disps).float().cuda()[None]
intrinsics = torch.from_numpy(intrinsics).float().cuda()[None]
N = poses.shape[1]
ii, jj = torch.meshgrid(torch.arange(N), torch.arange(N))
ii = ii.reshape(-1).cuda()
jj = jj.reshape(-1).cuda()
MAX_FLOW = 100.0
matrix = np.zeros((N, N), dtype=np.float32)
s = 2048
for i in range(0, ii.shape[0], s):
flow1, val1 = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s])
flow2, val2 = pops.induced_flow(poses, disps, intrinsics, jj[i:i+s], ii[i:i+s])
flow = torch.stack([flow1, flow2], dim=2)
val = torch.stack([val1, val2], dim=2)
mag = flow.norm(dim=-1).clamp(max=MAX_FLOW)
mag = mag.view(mag.shape[1], -1)
val = val.view(val.shape[1], -1)
mag = (mag * val).mean(-1) / val.mean(-1)
mag[val.mean(-1) < 0.7] = np.inf
i1 = ii[i:i+s].cpu().numpy()
j1 = jj[i:i+s].cpu().numpy()
matrix[i1, j1] = mag.cpu().numpy()
return matrix
def compute_distance_matrix_flow2(poses, disps, intrinsics, beta=0.4):
""" compute flow magnitude between all pairs of frames """
# if not isinstance(poses, SE3):
# poses = torch.from_numpy(poses).float().cuda()[None]
# poses = SE3(poses).inv()
# disps = torch.from_numpy(disps).float().cuda()[None]
# intrinsics = torch.from_numpy(intrinsics).float().cuda()[None]
N = poses.shape[1]
ii, jj = torch.meshgrid(torch.arange(N), torch.arange(N))
ii = ii.reshape(-1)
jj = jj.reshape(-1)
MAX_FLOW = 128.0
matrix = np.zeros((N, N), dtype=np.float32)
s = 2048
for i in range(0, ii.shape[0], s):
flow1a, val1a = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s], tonly=True)
flow1b, val1b = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s])
flow2a, val2a = pops.induced_flow(poses, disps, intrinsics, jj[i:i+s], ii[i:i+s], tonly=True)
flow2b, val2b = pops.induced_flow(poses, disps, intrinsics, ii[i:i+s], jj[i:i+s])
flow1 = flow1a + beta * flow1b
val1 = val1a * val2b
flow2 = flow2a + beta * flow2b
val2 = val2a * val2b
flow = torch.stack([flow1, flow2], dim=2)
val = torch.stack([val1, val2], dim=2)
mag = flow.norm(dim=-1).clamp(max=MAX_FLOW)
mag = mag.view(mag.shape[1], -1)
val = val.view(val.shape[1], -1)
mag = (mag * val).mean(-1) / val.mean(-1)
mag[val.mean(-1) < 0.8] = np.inf
i1 = ii[i:i+s].cpu().numpy()
j1 = jj[i:i+s].cpu().numpy()
matrix[i1, j1] = mag.cpu().numpy()
return matrix
================================================
FILE: dbaf/data_readers/stream.py
================================================
import numpy as np
import torch
import torch.utils.data as data
import torch.nn.functional as F
import csv
import os
import cv2
import math
import random
import json
import pickle
import os.path as osp
from .rgbd_utils import *
class RGBDStream(data.Dataset):
def __init__(self, datapath, frame_rate=-1, image_size=[384,512], crop_size=[0,0]):
self.datapath = datapath
self.frame_rate = frame_rate
self.image_size = image_size
self.crop_size = crop_size
self._build_dataset_index()
@staticmethod
def image_read(image_file):
return cv2.imread(image_file)
@staticmethod
def depth_read(depth_file):
return np.load(depth_file)
def __len__(self):
return len(self.images)
def __getitem__(self, index):
""" return training video """
image = self.__class__.image_read(self.images[index])
image = torch.from_numpy(image).float()
image = image.permute(2, 0, 1)
try:
tstamp = self.tstamps[index]
except:
tstamp = index
pose = torch.from_numpy(self.poses[index]).float()
intrinsic = torch.from_numpy(self.intrinsics[index]).float()
# resize image
sx = self.image_size[1] / image.shape[2]
sy = self.image_size[0] / image.shape[1]
image = F.interpolate(image[None], self.image_size, mode='bilinear', align_corners=False)[0]
fx, fy, cx, cy = intrinsic.unbind(dim=0)
fx, cx = sx * fx, sx * cx
fy, cy = sy * fy, sy * cy
# crop image
if self.crop_size[0] > 0:
cy = cy - self.crop_size[0]
image = image[:,self.crop_size[0]:-self.crop_size[0],:]
if self.crop_size[1] > 0:
cx = cx - self.crop_size[1]
image = image[:,:,self.crop_size[1]:-self.crop_size[1]]
intrinsic = torch.stack([fx, fy, cx, cy])
return tstamp, image, pose, intrinsic
class ImageStream(data.Dataset):
def __init__(self, datapath, intrinsics, rate=1, image_size=[384,512]):
rgb_list = osp.join(datapath, 'rgb.txt')
if os.path.isfile(rgb_list):
rgb_list = np.loadtxt(rgb_list, delimiter=' ', dtype=np.unicode_)
self.timestamps = rgb_list[:,0].astype(np.float)
self.images = [os.path.join(datapath, x) for x in rgb_list[:,1]]
self.images = self.images[::rate]
self.timestamps = self.timestamps[::rate]
else:
import glob
self.images = sorted(glob.glob(osp.join(datapath, '*.jpg'))) + sorted(glob.glob(osp.join(datapath, '*.png')))
self.images = self.images[::rate]
self.intrinsics = intrinsics
self.image_size = image_size
def __len__(self):
return len(self.images)
@staticmethod
def image_read(imfile):
return cv2.imread(imfile)
def __getitem__(self, index):
""" return training video """
image = self.__class__.image_read(self.images[index])
try:
tstamp = self.timestamps[index]
except:
tstamp = index
ht0, wd0 = image.shape[:2]
ht1, wd1 = self.image_size
intrinsics = torch.as_tensor(self.intrinsics)
intrinsics[0] *= wd1 / wd0
intrinsics[1] *= ht1 / ht0
intrinsics[2] *= wd1 / wd0
intrinsics[3] *= ht1 / ht0
# resize image
ikwargs = {'mode': 'bilinear', 'align_corners': True}
image = torch.from_numpy(image).float().permute(2, 0, 1)
image = F.interpolate(image[None], self.image_size, **ikwargs)[0]
return tstamp, image, intrinsics
class StereoStream(data.Dataset):
def __init__(self, datapath, intrinsics, rate=1, image_size=[384,512],
map_left=None, map_right=None, left_root='image_left', right_root='image_right'):
import glob
self.intrinsics = intrinsics
self.image_size = image_size
imgs = sorted(glob.glob(osp.join(datapath, left_root, '*.png')))[::rate]
self.images_l = []
self.images_r = []
self.tstamps = []
for img_l in imgs:
img_r = img_l.replace(left_root, right_root)
if os.path.isfile(img_r):
t = np.float(img_l.split('/')[-1].replace('.png', ''))
self.tstamps.append(t)
self.images_l += [ img_l ]
self.images_r += [ img_r ]
self.map_left = map_left
self.map_right = map_right
def __len__(self):
return len(self.images_l)
@staticmethod
def image_read(imfile, imap=None):
image = cv2.imread(imfile)
if imap is not None:
image = cv2.remap(image, imap[0], imap[1], interpolation=cv2.INTER_LINEAR)
return image
def __getitem__(self, index):
""" return training video """
tstamp = self.tstamps[index]
image_l = self.__class__.image_read(self.images_l[index], self.map_left)
image_r = self.__class__.image_read(self.images_r[index], self.map_right)
ht0, wd0 = image_l.shape[:2]
ht1, wd1 = self.image_size
intrinsics = torch.as_tensor(self.intrinsics)
intrinsics[0] *= wd1 / wd0
intrinsics[1] *= ht1 / ht0
intrinsics[2] *= wd1 / wd0
intrinsics[3] *= ht1 / ht0
image_l = torch.from_numpy(image_l).float().permute(2, 0, 1)
image_r = torch.from_numpy(image_r).float().permute(2, 0, 1)
# resize image
ikwargs = {'mode': 'bilinear', 'align_corners': True}
image_l = F.interpolate(image_l[None], self.image_size, **ikwargs)[0]
image_r = F.interpolate(image_r[None], self.image_size, **ikwargs)[0]
return tstamp, image_l, image_r, intrinsics
# class RGBDStream(data.Dataset):
# def __init__(self, datapath, intrinsics=None, rate=1, image_size=[384,512]):
# assoc_file = osp.join(datapath, 'associated.txt')
# assoc_list = np.loadtxt(assoc_file, delimiter=' ', dtype=np.unicode_)
# self.intrinsics = intrinsics
# self.image_size = image_size
# self.timestamps = assoc_list[:,0].astype(np.float)[::rate]
# self.images = [os.path.join(datapath, x) for x in assoc_list[:,1]][::rate]
# self.depths = [os.path.join(datapath, x) for x in assoc_list[:,3]][::rate]
# def __len__(self):
# return len(self.images)
# @staticmethod
# def image_read(imfile):
# return cv2.imread(imfile)
# @staticmethod
# def depth_read(depth_file):
# depth = cv2.imread(depth_file, cv2.IMREAD_ANYDEPTH)
# return depth.astype(np.float32) / 5000.0
# def __getitem__(self, index):
# """ return training video """
# tstamp = self.timestamps[index]
# image = self.__class__.image_read(self.images[index])
# depth = self.__class__.depth_read(self.depths[index])
# ht0, wd0 = image.shape[:2]
# ht1, wd1 = self.image_size
# intrinsics = torch.as_tensor(self.intrinsics)
# intrinsics[0] *= wd1 / wd0
# intrinsics[1] *= ht1 / ht0
# intrinsics[2] *= wd1 / wd0
# intrinsics[3] *= ht1 / ht0
# # resize image
# ikwargs = {'mode': 'bilinear', 'align_corners': True}
# image = torch.from_numpy(image).float().permute(2, 0, 1)
# image = F.interpolate(image[None], self.image_size, **ikwargs)[0]
# depth = torch.from_numpy(depth).float()[None,None]
# depth = F.interpolate(depth, self.image_size, mode='nearest').squeeze()
# return tstamp, image, depth, intrinsics
================================================
FILE: dbaf/data_readers/tartan.py
================================================
import numpy as np
import torch
import glob
import cv2
import os
import os.path as osp
from lietorch import SE3
from .base import RGBDDataset
from .stream import RGBDStream
cur_path = osp.dirname(osp.abspath(__file__))
test_split = osp.join(cur_path, 'tartan_test.txt')
test_split = open(test_split).read().split()
class TartanAir(RGBDDataset):
# scale depths to balance rot & trans
DEPTH_SCALE = 5.0
def __init__(self, mode='training', **kwargs):
self.mode = mode
self.n_frames = 2
super(TartanAir, self).__init__(name='TartanAir', **kwargs)
@staticmethod
def is_test_scene(scene):
# print(scene, any(x in scene for x in test_split))
return any(x in scene for x in test_split)
def _build_dataset(self):
from tqdm import tqdm
print("Building TartanAir dataset")
scene_info = {}
scenes = glob.glob(osp.join(self.root, '*/*/*/*'))
for scene in tqdm(sorted(scenes)):
images = sorted(glob.glob(osp.join(scene, 'image_left/*.png')))
depths = sorted(glob.glob(osp.join(scene, 'depth_left/*.npy')))
poses = np.loadtxt(osp.join(scene, 'pose_left.txt'), delimiter=' ')
poses = poses[:, [1, 2, 0, 4, 5, 3, 6]]
poses[:,:3] /= TartanAir.DEPTH_SCALE
intrinsics = [TartanAir.calib_read()] * len(images)
# graph of co-visible frames based on flow
graph = self.build_frame_graph(poses, depths, intrinsics)
scene = '/'.join(scene.split('/'))
scene_info[scene] = {'images': images, 'depths': depths,
'poses': poses, 'intrinsics': intrinsics, 'graph': graph}
return scene_info
@staticmethod
def calib_read():
return np.array([320.0, 320.0, 320.0, 240.0])
@staticmethod
def image_read(image_file):
return cv2.imread(image_file)
@staticmethod
def depth_read(depth_file):
depth = np.load(depth_file) / TartanAir.DEPTH_SCALE
depth[depth==np.nan] = 1.0
depth[depth==np.inf] = 1.0
return depth
class TartanAirStream(RGBDStream):
def __init__(self, datapath, **kwargs):
super(TartanAirStream, self).__init__(datapath=datapath, **kwargs)
def _build_dataset_index(self):
""" build list of images, poses, depths, and intrinsics """
self.root = 'datasets/TartanAir'
scene = osp.join(self.root, self.datapath)
image_glob = osp.join(scene, 'image_left/*.png')
images = sorted(glob.glob(image_glob))
poses = np.loadtxt(osp.join(scene, 'pose_left.txt'), delimiter=' ')
poses = poses[:, [1, 2, 0, 4, 5, 3, 6]]
poses = SE3(torch.as_tensor(poses))
poses = poses[[0]].inv() * poses
poses = poses.data.cpu().numpy()
intrinsic = self.calib_read(self.datapath)
intrinsics = np.tile(intrinsic[None], (len(images), 1))
self.images = images[::int(self.frame_rate)]
self.poses = poses[::int(self.frame_rate)]
self.intrinsics = intrinsics[::int(self.frame_rate)]
@staticmethod
def calib_read(datapath):
return np.array([320.0, 320.0, 320.0, 240.0])
@staticmethod
def image_read(image_file):
return cv2.imread(image_file)
class TartanAirTestStream(RGBDStream):
def __init__(self, datapath, **kwargs):
super(TartanAirTestStream, self).__init__(datapath=datapath, **kwargs)
def _build_dataset_index(self):
""" build list of images, poses, depths, and intrinsics """
self.root = 'datasets/mono'
image_glob = osp.join(self.root, self.datapath, '*.png')
images = sorted(glob.glob(image_glob))
poses = np.loadtxt(osp.join(self.root, 'mono_gt', self.datapath + '.txt'), delimiter=' ')
poses = poses[:, [1, 2, 0, 4, 5, 3, 6]]
poses = SE3(torch.as_tensor(poses))
poses = poses[[0]].inv() * poses
poses = poses.data.cpu().numpy()
intrinsic = self.calib_read(self.datapath)
intrinsics = np.tile(intrinsic[None], (len(images), 1))
self.images = images[::int(self.frame_rate)]
self.poses = poses[::int(self.frame_rate)]
self.intrinsics = intrinsics[::int(self.frame_rate)]
@staticmethod
def calib_read(datapath):
return np.array([320.0, 320.0, 320.0, 240.0])
@staticmethod
def image_read(image_file):
return cv2.imread(image_file)
================================================
FILE: dbaf/data_readers/tartan_test.txt
================================================
abandonedfactory/abandonedfactory/Easy/P011
abandonedfactory/abandonedfactory/Hard/P011
abandonedfactory_night/abandonedfactory_night/Easy/P013
abandonedfactory_night/abandonedfactory_night/Hard/P014
amusement/amusement/Easy/P008
amusement/amusement/Hard/P007
carwelding/carwelding/Easy/P007
endofworld/endofworld/Easy/P009
gascola/gascola/Easy/P008
gascola/gascola/Hard/P009
hospital/hospital/Easy/P036
hospital/hospital/Hard/P049
japanesealley/japanesealley/Easy/P007
japanesealley/japanesealley/Hard/P005
neighborhood/neighborhood/Easy/P021
neighborhood/neighborhood/Hard/P017
ocean/ocean/Easy/P013
ocean/ocean/Hard/P009
office2/office2/Easy/P011
office2/office2/Hard/P010
office/office/Hard/P007
oldtown/oldtown/Easy/P007
oldtown/oldtown/Hard/P008
seasidetown/seasidetown/Easy/P009
seasonsforest/seasonsforest/Easy/P011
seasonsforest/seasonsforest/Hard/P006
seasonsforest_winter/seasonsforest_winter/Easy/P009
seasonsforest_winter/seasonsforest_winter/Hard/P018
soulcity/soulcity/Easy/P012
soulcity/soulcity/Hard/P009
westerndesert/westerndesert/Easy/P013
westerndesert/westerndesert/Hard/P007
================================================
FILE: dbaf/dbaf.py
================================================
import torch
import lietorch
import numpy as np
from droid_net import DroidNet
from depth_video import DepthVideo
from motion_filter import MotionFilter
from dbaf_frontend import DBAFusionFrontend
from collections import OrderedDict
from torch.multiprocessing import Process
from lietorch import SE3
import geom.projective_ops as pops
import droid_backends
import pickle
class DBAFusion:
def __init__(self, args):
super(DBAFusion, self).__init__()
self.load_weights(args.weights) # load DroidNet weights
self.args = args
# store images, depth, poses, intrinsics (shared between processes)
self.video = DepthVideo(args.image_size, args.buffer, save_pkl = args.save_pkl, stereo=args.stereo, upsample=args.upsample)
# filter incoming frames so that there is enough motion
self.filterx = MotionFilter(self.net, self.video, thresh=args.filter_thresh)
# frontend process
self.frontend = DBAFusionFrontend(self.net, self.video, self.args)
self.pklpath = args.pklpath
self.upsample = args.upsample
def load_weights(self, weights):
""" load trained model weights """
print(weights)
self.net = DroidNet()
state_dict = OrderedDict([
(k.replace("module.", ""), v) for (k, v) in torch.load(weights).items()])
state_dict["update.weight.2.weight"] = state_dict["update.weight.2.weight"][:2]
state_dict["update.weight.2.bias"] = state_dict["update.weight.2.bias"][:2]
state_dict["update.delta.2.weight"] = state_dict["update.delta.2.weight"][:2]
state_dict["update.delta.2.bias"] = state_dict["update.delta.2.bias"][:2]
self.net.load_state_dict(state_dict)
self.net.to("cuda:0").eval()
def track(self, tstamp, image, depth=None, intrinsics=None):
""" main thread - update map """
with torch.no_grad():
# check there is enough motion
self.filterx.track(tstamp, image, depth, intrinsics)
# local bundle adjustment
self.frontend()
def terminate(self, stream=None):
""" terminate the visualization process, return poses [t, q] """
del self.frontend
def save_vis_easy(self):
mcameras = {}
mpoints = {}
mstamps = {}
with torch.no_grad():
dirty_index = torch.arange(0,self.video.count_save,device='cuda')
stamps= torch.index_select(self.video.tstamp_save, 0 ,dirty_index)
poses= torch.index_select( self.video.poses_save, 0 ,dirty_index)
disps= torch.index_select( self.video.disps_save, 0 ,dirty_index)
images = torch.index_select( self.video.images_save, 0 ,dirty_index)
Ps = SE3(poses).inv().matrix().cpu().numpy()
points = droid_backends.iproj(SE3(poses).inv().data, disps, self.video.intrinsics[0]).cpu()
thresh = 0.4 * torch.ones_like(disps.mean(dim=[1,2])) / 4.0 * (1.0 / torch.median(disps.mean(dim=[1,2])))
# thresh = 0.4 * torch.ones_like(disps.mean(dim=[1,2]))
count = droid_backends.depth_filter(
self.video.poses_save, self.video.disps_save, self.video.intrinsics[0], dirty_index, thresh)
count = count.cpu()
disps = disps.cpu()
if self.upsample:
disps_up= torch.index_select( self.video.disps_up_save, 0 ,dirty_index)
disps_up = disps_up.cpu()
masks = ((count >= 1) & (disps > .5*disps.mean(dim=[1,2], keepdim=True)))
for i in range(len(dirty_index)):
pose = Ps[i]
ix = dirty_index[i].item()
mcameras[ix] = pose
mask = masks[i].reshape(-1)
pts = points[i].reshape(-1, 3)[mask].cpu().numpy()
clr = images[i].reshape(-1, 3)[mask].cpu().numpy()
stamp = stamps[i].cpu()
if self.upsample:
mpoints[ix] = {'pts':pts,'clr':clr,'disp':disps[i].cpu().numpy(),'disps_up':disps_up[i].cpu().numpy(),'rgb':images[i].cpu().numpy()}
else:
mpoints[ix] = {'pts':pts,'clr':clr,'disp':disps[i].cpu().numpy(),'rgb':images[i].cpu().numpy()}
mstamps[ix] = stamp
ddict = {'points':mpoints,'cameras':mcameras,'stamps':mstamps}
f_save = open(self.pklpath, 'wb')
pickle.dump(ddict,f_save)
mcameras = {}
mpoints = {}
mstamps = {}
with torch.no_grad():
dirty_index = torch.arange(0,self.video.count_save,device='cuda')
stamps= torch.index_select(self.video.tstamp_save, 0 ,dirty_index)
poses= torch.index_select( self.video.poses_save, 0 ,dirty_index)
disps= torch.index_select( self.video.disps_save, 0 ,dirty_index)
images = torch.index_select( self.video.images_save, 0 ,dirty_index)
Ps = SE3(poses).inv().matrix().cpu().numpy()
points = droid_backends.iproj(SE3(poses).inv().data, disps, self.video.intrinsics[0]).cpu()
thresh = 0.4 * torch.ones_like(disps.mean(dim=[1,2]))
count = droid_backends.depth_filter(
self.video.poses_save, self.video.disps_save, self.video.intrinsics[0], dirty_index, thresh)
count = count.cpu()
disps = disps.cpu()
masks = ((count >= 0) & (disps > .5*disps.mean(dim=[1,2], keepdim=True)))
for i in range(len(dirty_index)):
pose = Ps[i]
ix = dirty_index[i].item()
mcameras[ix] = pose
mask = masks[i].reshape(-1)
pts = points[i].reshape(-1, 3)[mask].cpu().numpy()
clr = images[i].reshape(-1, 3)[mask].cpu().numpy()
stamp = stamps[i].cpu()
mpoints[ix] = {'pts':pts,'clr':clr,'disp':disps[i].cpu().numpy(),'rgb':images[i].cpu().numpy()}
mstamps[ix] = stamp
ddict = {'points':mpoints,'cameras':mcameras,'stamps':mstamps}
f_save = open(self.pklpath.split('.')[0] + '_raw.pkl', 'wb')
pickle.dump(ddict,f_save)
================================================
FILE: dbaf/dbaf_frontend.py
================================================
import torch
import torchvision
import numpy as np
from lietorch import SE3, SO3
from covisible_graph import CovisibleGraph
import matplotlib.pyplot as plt
import gtsam
import math
import bisect
from math import atan2, cos, sin
import geoFunc.trans as trans
from scipy.spatial.transform import Rotation
class DBAFusionFrontend:
def __init__(self, net, video, args):
self.video = video
self.update_op = net.update
self.graph = CovisibleGraph(video, net.update, args=args)
# local optimization window
self.t0 = 0
self.t1 = 0
# frontend variables
self.is_initialized = False
self.count = 0
self.warmup = args.warmup
self.vi_warmup = 12
if 'vi_warmup' in args: self.vi_warmup = args.vi_warmup
self.beta = args.beta
self.frontend_nms = args.frontend_nms
self.keyframe_thresh = args.keyframe_thresh
self.frontend_window = args.frontend_window
self.frontend_thresh = args.frontend_thresh
self.frontend_radius = args.frontend_radius
### DBAFusion
self.all_imu = None
self.cur_imu_ii = 0
self.is_init = False
self.all_gnss = None
self.all_odo = None
self.all_gt = None
self.all_gt_keys = None
self.all_stamp = None
self.cur_stamp_ii = 0
self.visual_only = args.visual_only
self.visual_only_init = False
self.translation_threshold = 0.0
self.active_window = args.active_window
self.high_freq_output = True
self.zupt = ('use_zupt' in args and args.use_zupt)
if not self.visual_only:
self.max_age = 25
self.iters1 = 2
self.iters2 = 1
else:
self.max_age = 25
self.iters1 = 4
self.iters2 = 2
# visualization/output
self.show_plot = args.show_plot
self.result_file = open(args.resultpath,'wt')
self.plt_pos = [[],[]] # X, Y
self.plt_pos_ref = [[],[]] # X, Y
self.plt_att = [[],[],[]] # pitch, roll, yaw
self.plt_bg = [[],[],[]] # X, Y, Z
self.plt_t = []
self.refTw = np.eye(4,4)
if self.show_plot:
plt.figure('monitor',figsize=[13,4])
plt.subplot(1,3,1); plt.gca().set_title('Trajectory')
plt.gca().set_aspect(1)
plt.subplot(1,3,2); plt.gca().set_title('Attitude Error/Attitude')
plt.subplot(1,3,3); plt.gca().set_title('Gyroscope Bias')
plt.ion()
plt.pause(0.1)
def get_pose_ref(self, tt:float):
tt_found = self.all_gt_keys[bisect.bisect(self.all_gt_keys,tt)]
return tt_found, self.all_gt[tt_found]
def __rollup(self, roll):
""" roll up window states to save memory """
self.t1 -= roll
self.count -= roll
self.video.counter.value -= roll
self.video.tstamp = torch.roll(self.video.tstamp ,-roll,0)
self.video.images = torch.roll(self.video.images ,-roll,0)
self.video.dirty = torch.roll(self.video.dirty ,-roll,0)
self.video.red = torch.roll(self.video.red ,-roll,0)
self.video.poses = torch.roll(self.video.poses ,-roll,0)
self.video.disps = torch.roll(self.video.disps ,-roll,0)
self.video.disps_sens = torch.roll(self.video.disps_sens,-roll,0)
self.video.disps_up = torch.roll(self.video.disps_up ,-roll,0)
self.video.intrinsics = torch.roll(self.video.intrinsics,-roll,0)
self.video.fmaps = torch.roll(self.video.fmaps ,-roll,0)
self.video.nets = torch.roll(self.video.nets ,-roll,0)
self.video.inps = torch.roll(self.video.inps ,-roll,0)
self.graph.ii -= roll
self.graph.jj -= roll
self.graph.ii_inac -= roll
self.graph.jj_inac -= roll
rm_inac_index = torch.logical_and(torch.greater_equal(self.graph.ii_inac,0),torch.greater_equal(self.graph.jj_inac,0))
self.graph.ii_inac = self.graph.ii_inac[rm_inac_index]
self.graph.jj_inac = self.graph.jj_inac[rm_inac_index]
self.graph.target_inac = self.graph.target_inac[:,rm_inac_index,:,:,:]
self.graph.weight_inac = self.graph.weight_inac[:,rm_inac_index,:,:,:] # need test
self.graph.ii_bad -= roll
self.graph.jj_bad -= roll
self.video.last_t0 -= roll
self.video.last_t1 -= roll
self.video.cur_ii -= roll
self.video.cur_jj -= roll
if self.video.imu_enabled:
graph_temp = gtsam.NonlinearFactorGraph()
for i in range(self.video.cur_graph.size()):
f = self.video.cur_graph.at(i)
graph_temp.push_back(f.rekey((np.array(f.keys())-roll).tolist()))
self.video.cur_graph = graph_temp
result_temp = gtsam.Values()
for i in self.video.cur_result.keys():
if gtsam.Symbol(i).chr() == ord('b'):
result_temp.insert(i-roll,self.video.cur_result.atConstantBias(i))
elif gtsam.Symbol(i).chr() == ord('v'):
result_temp.insert(i-roll,self.video.cur_result.atVector(i))
elif gtsam.Symbol(i).chr() == ord('x'):
result_temp.insert(i-roll,self.video.cur_result.atPose3(i))
else:
raise Exception()
self.video.cur_result = result_temp
self.video.marg_factor = self.video.marg_factor.rekey((np.array(self.video.marg_factor.keys())-roll).tolist())
self.video.state.timestamps = self.video.state.timestamps [roll:]
self.video.state.wTbs = self.video.state.wTbs [roll:]
self.video.state.vs = self.video.state.vs [roll:]
self.video.state.bs = self.video.state.bs [roll:]
self.video.state.preintegrations = self.video.state.preintegrations [roll:]
self.video.state.preintegrations_meas = self.video.state.preintegrations_meas [roll:]
self.video.state.gnss_valid = self.video.state.gnss_valid [roll:]
self.video.state.gnss_position = self.video.state.gnss_position [roll:]
self.video.state.odo_valid = self.video.state.odo_valid [roll:]
self.video.state.odo_vel = self.video.state.odo_vel [roll:]
def __update(self):
""" add edges, perform update """
self.count += 1
self.t1 += 1
if self.video.imu_enabled and (self.video.tstamp[self.t1-1] - self.video.vi_init_time > 5.0):
self.video.reinit = True
self.video.vi_init_time = 1e9
## new frame comes, append IMU
cur_t = float(self.video.tstamp[self.t1-1].detach().cpu())
self.video.logger.info('predict %f' %cur_t)
while self.all_imu[self.cur_imu_ii][0] < cur_t:
## high-frequency output
# predict the pose of skipped frames through IMU preintegration
if self.high_freq_output and self.video.imu_enabled:
if self.all_imu[self.cur_imu_ii][0] > float(self.all_stamp[self.cur_stamp_ii][0]):
self.video.state.append_imu_temp(float(self.all_stamp[self.cur_stamp_ii][0]),\
self.all_imu[self.cur_imu_ii][4:7],\
self.all_imu[self.cur_imu_ii][1:4]/180*math.pi,True)
if float(self.all_stamp[self.cur_stamp_ii][0]) > self.video.state.timestamps[-1] and\
math.fabs(cur_t - float(self.all_stamp[self.cur_stamp_ii][0]))>1e-3:
pose_temp = self.video.state.pose_temp
ppp = pose_temp.pose().translation()
qqq = Rotation.from_matrix(pose_temp.pose().rotation().matrix()).as_quat()
line = '%.6f %.6f %.6f %.6f %.6f %.6f %.6f %.6f'%(float(self.all_stamp[self.cur_stamp_ii][0]),ppp[0],ppp[1],ppp[2]\
,qqq[0],qqq[1],qqq[2],qqq[3])
if self.video.gnss_init_t1>0:
p = self.video.ten0 + np.matmul(trans.Cen(self.video.ten0), ppp)
line += ' %.6f %.6f %.6f'% (p[0],p[1],p[2])
self.result_file.writelines(line+'\n')
# self.result_file.flush()
self.cur_stamp_ii += 1
self.video.state.append_imu_temp(self.all_imu[self.cur_imu_ii][0],\
self.all_imu[self.cur_imu_ii][4:7],\
self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)
self.video.state.append_imu(self.all_imu[self.cur_imu_ii][0],\
self.all_imu[self.cur_imu_ii][4:7],\
self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)
self.cur_imu_ii += 1
self.video.state.append_imu(cur_t,\
self.all_imu[self.cur_imu_ii][4:7],\
self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)
self.video.state.append_img(cur_t)
## append GNSS
if len(self.all_gnss) > 0: gnss_found = bisect.bisect(self.all_gnss[:,0],cur_t - 1e-6)
else: gnss_found = -1
if gnss_found > 0 and self.all_gnss[gnss_found,0] - cur_t < 0.01 :
self.video.state.append_gnss(cur_t,self.all_gnss[gnss_found,1:4])
## append ZUPT
if self.zupt and self.video.state.preintegrations[self.t1-3].deltaTij() > 3.0:
if np.linalg.norm(self.video.state.vs[self.t1-2]) < 0.025:
self.video.state.append_odo(cur_t,np.array([.0,.0,.0]))
## append ODO
if len(self.all_odo) > 0: odo_found = bisect.bisect(self.all_odo[:,0],cur_t - 1e-6)
else: odo_found = -1
if odo_found > 0 and self.all_odo[odo_found,0] - cur_t < 0.01 :
self.video.state.append_odo(cur_t,self.all_odo[odo_found,1:4])
self.video.state.append_imu(self.all_imu[self.cur_imu_ii][0],\
self.all_imu[self.cur_imu_ii][4:7],\
self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)
self.cur_imu_ii += 1
## predict pose (<5 ms)
if self.video.imu_enabled:
Twc = (self.video.state.wTbs[-1] * self.video.Tbc).matrix()
TTT = torch.tensor(np.linalg.inv(Twc))
q = torch.tensor(Rotation.from_matrix(TTT[:3, :3]).as_quat())
t = TTT[:3,3]
self.video.poses[self.t1-1] = torch.cat([t,q])
self.video.logger.info('manage edges')
## manage edges (60 ms)
if self.graph.corr is not None:
if self.visual_only:
self.graph.rm_factors(torch.logical_and(self.graph.age > self.max_age,\
torch.logical_or(self.graph.ii < self.t1-self.active_window,self.graph.jj < self.t1-self.active_window)), store=True)
else:
self.graph.rm_factors(torch.logical_or(self.graph.age > self.max_age,\
torch.logical_or(self.graph.ii < self.t1-self.active_window,self.graph.jj < self.t1-self.active_window)), store=True)
self.graph.add_proximity_factors(self.t1-5, max(self.t1-self.frontend_window, 0),
rad=self.frontend_radius, nms=self.frontend_nms, thresh=self.frontend_thresh, beta=self.beta, remove=True)
self.video.logger.info('non-keyframes %d' % self.graph.ii.shape[0])
## non-keyframe update
self.video.disps[self.t1-1] = torch.where(self.video.disps_sens[self.t1-1] > 0,
self.video.disps_sens[self.t1-1], self.video.disps[self.t1-1])
for itr in range(self.iters1):
self.graph.update(None, None, use_inactive=True)
self.rollup = False
if self.t1 > 65:
self.__rollup(30)
print('rollup ',self.graph.ii)
self.rollup = True
self.video.logger.info('output')
## visualization/output
poses = SE3(self.video.poses)
d = self.video.distance([self.t1-3], [self.t1-2], beta=self.beta, bidirectional=True)
TTT = np.matmul(poses[self.t1-1].cpu().inv().matrix(),np.linalg.inv(self.video.Ti1c))
if self.video.imu_enabled or (self.visual_only and self.visual_only_init):
ppp = TTT[0:3,3]
qqq = Rotation.from_matrix(TTT[:3, :3]).as_quat()
line = '%.6f %.6f %.6f %.6f %.6f %.6f %.6f %.6f'%(cur_t,ppp[0],ppp[1],ppp[2]\
,qqq[0],qqq[1],qqq[2],qqq[3])
if self.video.gnss_init_t1>0:
p = self.video.ten0 + np.matmul(trans.Cen(self.video.ten0), ppp.numpy())
line += ' %.6f %.6f %.6f'% (p[0],p[1],p[2])
self.result_file.writelines(line+'\n')
self.result_file.flush()
TTTref = np.matmul(self.refTw,TTT)
ppp = TTTref[0:3,3]
if self.show_plot:
# if math.fabs(tt_found - cur_t) < 0.1: # for kitti and whu
self.plt_pos[0].append(ppp[0])
self.plt_pos[1].append(ppp[1])
a1 = np.array(trans.m2att(TTTref[0:3,0:3]) )* 57.3
if self.all_gt is not None:
tt_found,dd = self.get_pose_ref(cur_t -1e-3)
self.plt_pos_ref[0].append(dd['T'][0,3])
self.plt_pos_ref[1].append(dd['T'][1,3])
a2 = np.array(trans.m2att(dd['T'][0:3,0:3]) )* 57.3
a1 -= a2
self.plt_att[0].append(a1[0])
self.plt_att[1].append(a1[1])
self.plt_att[2].append(a1[2])
bg = self.video.state.bs[self.t1-1].gyroscope()
self.plt_bg[0].append(bg[0])
self.plt_bg[1].append(bg[1])
self.plt_bg[2].append(bg[2])
self.plt_t.append(cur_t)
if self.rollup:
plt.subplot(1,3,1)
plt.cla(); plt.gca().set_title('Trajectory')
plt.plot(self.plt_pos[0],self.plt_pos[1],marker='^')
plt.plot(self.plt_pos_ref[0],self.plt_pos_ref[1],marker='^')
plt.subplot(1,3,2)
plt.cla(); plt.gca().set_title('Attitude Error/Attitude')
plt.plot(self.plt_t,self.plt_att[0],c='r')
plt.plot(self.plt_t,self.plt_att[1],c='g')
plt.plot(self.plt_t,self.plt_att[2],c='b')
plt.ylim([-10,10])
plt.subplot(1,3,3)
plt.cla(); plt.gca().set_title('Gyroscope Bias')
plt.plot(self.plt_t,self.plt_bg[0],c='r')
plt.plot(self.plt_t,self.plt_bg[1],c='g')
plt.plot(self.plt_t,self.plt_bg[2],c='b')
plt.pause(0.1)
## keyframe update
self.video.logger.info('keyframes %d' % self.graph.ii.shape[0])
if self.t1 > 10:
cam_translation = torch.norm((poses[(self.t1-10):(self.t1-3)] * poses[self.t1-2].inv()[None]).translation()[:,0:3],dim=1)
else:
cam_translation = torch.norm((poses[(self.t1-6):(self.t1-3)] * poses[self.t1-2].inv()[None]).translation()[:,0:3],dim=1)
if (d.item() < self.keyframe_thresh or (self.video.imu_enabled and torch.sum(cam_translation < self.translation_threshold)>0)): # gnss
self.video.logger.info('remove new frame!!!!!!!!!!!!1')
self.graph.rm_keyframe(self.t1 - 2)
# merge preintegration[self.t1-2] and preintegration[self.t1-3]
for iii in range(len(self.video.state.preintegrations_meas[self.t1-2])):
dd = self.video.state.preintegrations_meas[self.t1-2][iii]
if dd[2] > 0:
self.video.state.preintegrations[self.t1-3].integrateMeasurement(dd[0],\
dd[1],\
dd[2])
self.video.state.preintegrations_meas[self.t1-3].append(dd)
self.video.state.preintegrations[self.t1-2] = self.video.state.preintegrations[self.t1-1]
self.video.state.preintegrations_meas[self.t1-2] = self.video.state.preintegrations_meas[self.t1-1]
self.video.state.preintegrations.pop()
self.video.state.preintegrations_meas.pop()
self.video.rm_new_gnss(self.t1-2)
self.video.state.wTbs[self.t1-2] = self.video.state.wTbs[self.t1-1]; self.video.state.wTbs.pop()
self.video.state.bs [self.t1-2] = self.video.state.bs [self.t1-1]; self.video.state.bs.pop()
self.video.state.vs [self.t1-2] = self.video.state.vs [self.t1-1]; self.video.state.vs .pop()
self.video.state.gnss_valid [self.t1-2] = self.video.state.gnss_valid [self.t1-1]; self.video.state.gnss_valid .pop()
self.video.state.gnss_position [self.t1-2] = self.video.state.gnss_position [self.t1-1]; self.video.state.gnss_position .pop()
self.video.state.odo_valid [self.t1-2] = self.video.state.odo_valid [self.t1-1]; self.video.state.odo_valid .pop()
self.video.state.odo_vel [self.t1-2] = self.video.state.odo_vel [self.t1-1]; self.video.state.odo_vel .pop()
with self.video.get_lock():
self.video.counter.value -= 1
self.t1 -= 1
else:
for itr in range(self.iters2):
# print('b%d' % itr)
self.graph.update(None, None, use_inactive=True)
## try initializing VI/GNSS
if self.t1 > self.vi_warmup and self.video.vi_init_t1 < 0:
self.init_VI()
if not self.visual_only:
for i in range(len(self.all_stamp)): # skip to next image
if float(self.all_stamp[i][0]) < cur_t + 1e-6: continue
else:
self.cur_stamp_ii = i
break
if self.video.imu_enabled and self.video.gnss_init_time <= 0.0 and len(self.all_gnss)>0:
self.init_GNSS()
## set pose for next itration
self.video.poses[self.t1] = self.video.poses[self.t1-1]
self.video.disps[self.t1] = self.video.disps[self.t1-1].mean() * 1.0
self.video.dirty[self.graph.ii.min():self.t1] = True
def init_IMU(self):
""" initialize IMU states """
cur_t = float(self.video.tstamp[self.t0].detach().cpu())
for i in range(len(self.all_imu)):
if self.all_imu[i][0] < cur_t - 1e-6: continue
else:
self.cur_imu_ii = i
break
for i in range(self.t0,self.t1):
tt = self.video.tstamp[i]
if i == self.t0:
self.video.state.init_first_state(cur_t,np.zeros(3),\
np.eye(3),\
np.zeros(3))
self.video.state.append_imu(self.all_imu[self.cur_imu_ii][0],\
self.all_imu[self.cur_imu_ii][4:7],\
self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)
self.cur_imu_ii += 1
self.is_init = True
else:
cur_t = float(self.video.tstamp[i].detach().cpu())
while self.all_imu[self.cur_imu_ii][0] < cur_t:
self.video.state.append_imu(self.all_imu[self.cur_imu_ii][0],\
self.all_imu[self.cur_imu_ii][4:7],\
self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)
self.cur_imu_ii += 1
self.video.state.append_imu(cur_t,\
self.all_imu[self.cur_imu_ii][4:7],\
self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)
self.video.state.append_img(cur_t)
if len(self.all_gnss) > 0: gnss_found = bisect.bisect(self.all_gnss[:,0],cur_t - 1e-6)
else: gnss_found = -1
if gnss_found > 0 and self.all_gnss[gnss_found,0] - cur_t < 0.01:
self.video.state.append_gnss(cur_t,self.all_gnss[gnss_found,1:4])
if len(self.all_odo) > 0: odo_found = bisect.bisect(self.all_odo[:,0],cur_t - 1e-6)
else: odo_found = -1
if odo_found > 0 and self.all_odo[odo_found,0] - cur_t < 0.01 :
self.video.state.append_odo(cur_t,self.all_odo[odo_found,1:4])
self.video.state.append_imu(self.all_imu[self.cur_imu_ii][0],\
self.all_imu[self.cur_imu_ii][4:7],\
self.all_imu[self.cur_imu_ii][1:4]/180*math.pi)
self.cur_imu_ii += 1
Twc = np.matmul(np.array([[1,0,0,0],\
[0,1,0,0],\
[0,0,1,0.02*i],\
[0,0,0,1]]),self.video.Ti1c) # perturb the camera poses, which benefits the robustness of initial BA
TTT = torch.tensor(np.linalg.inv(Twc))
q = torch.tensor(Rotation.from_matrix(TTT[:3, :3]).as_quat())
t = TTT[:3,3]
if not self.video.imu_enabled:
self.video.poses[i] = torch.cat([t,q])
def init_VI(self):
""" initialize the V-I system, referring to VIN-Fusion """
sum_g = np.zeros(3,dtype = np.float64)
ccount = 0
for i in range(self.t1 - 8 ,self.t1-1):
dt = self.video.state.preintegrations[i].deltaTij()
tmp_g = self.video.state.preintegrations[i].deltaVij()/dt
sum_g += tmp_g
ccount += 1
aver_g = sum_g * 1.0 / ccount
var_g = 0.0
for i in range(self.t1 - 8 ,self.t1-1):
dt = self.video.state.preintegrations[i].deltaTij()
tmp_g = self.video.state.preintegrations[i].deltaVij()/dt
var_g += np.linalg.norm(tmp_g - aver_g)**2
var_g =math.sqrt(var_g/ccount)
if var_g < 0.25:
print("IMU excitation not enough!",var_g)
else:
poses = SE3(self.video.poses)
self.plt_pos = [[],[]]
self.plt_pos_ref = [[],[]]
for i in range(0,self.t1):
ppp = np.matmul(poses[i].cpu().inv().matrix(),np.linalg.inv(self.video.Ti1c))[0:3,3]
self.plt_pos[0].append(ppp[0])
self.plt_pos[1].append(ppp[1])
if self.all_gt is not None:
tt_found,dd = self.get_pose_ref(self.video.tstamp[i]-1e-3)
self.plt_pos_ref[0].append(dd['T'][0,3])
self.plt_pos_ref[1].append(dd['T'][1,3])
if self.show_plot:
plt.subplot(1,3,1)
plt.cla(); plt.gca().set_title('Trajectory')
plt.plot(self.plt_pos[0],self.plt_pos[1],marker='^')
plt.plot(self.plt_pos_ref[0],self.plt_pos_ref[1],marker='^')
plt.pause(0.1)
if not self.visual_only:
self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= True)
self.graph.update(None, None, use_inactive=True)
self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= False)
self.graph.update(None, None, use_inactive=True)
self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= False)
self.video.imu_enabled = True
else:
self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= True)
self.graph.update(None, None, use_inactive=True)
self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= False)
self.graph.update(None, None, use_inactive=True)
self.VisualIMUAlignment(self.t1 - 8 ,self.t1, ignore_lever= False)
self.visual_only_init = True
self.video.set_prior(self.video.last_t0,self.t1)
self.plt_pos = [[],[]]
self.plt_pos_ref = [[],[]]
for i in range(0,self.t1):
TTT = self.video.state.wTbs[i].matrix()
ppp = TTT[0:3,3]
qqq = Rotation.from_matrix(TTT[:3, :3]).as_quat()
self.result_file.writelines('%.6f %.6f %.6f %.6f %.6f %.6f %.6f %.6f\n'%(self.video.tstamp[i],ppp[0],ppp[1],ppp[2]\
,qqq[0],qqq[1],qqq[2],qqq[3]))
TTTref = np.matmul(self.refTw,TTT) # for visualization
ppp = TTTref[0:3,3]
qqq = Rotation.from_matrix(TTTref[:3, :3]).as_quat()
self.plt_pos[0].append(ppp[0])
self.plt_pos[1].append(ppp[1])
if self.all_gt is not None:
tt_found,dd = self.get_pose_ref(self.video.tstamp[i]-1e-3)
self.plt_pos_ref[0].append(dd['T'][0,3])
self.plt_pos_ref[1].append(dd['T'][1,3])
if self.show_plot:
plt.subplot(1,3,1)
plt.cla(); plt.gca().set_title('Trajectory')
plt.plot(self.plt_pos[0],self.plt_pos[1],marker='^')
plt.plot(self.plt_pos_ref[0],self.plt_pos_ref[1],marker='^')
plt.pause(0.1)
for itr in range(1):
self.graph.update(None, None, use_inactive=True)
def init_GNSS(self):
""" initialize the GNSS for geo-referencing fusion """
ten0 = np.array([self.all_gt[self.all_gt_keys[0]]['X0'],\
self.all_gt[self.all_gt_keys[0]]['Y0'],\
self.all_gt[self.all_gt_keys[0]]['Z0']])
self.video.ten0 = ten0
tn0 = []; tw =[]
for i in range(len(self.video.state.wTbs) - 10,len(self.video.state.wTbs)):
if self.video.state.gnss_valid[i]:
# if not is_ref_set:
# ten0 = self.video.sgraph.gnss_position[i]
# is_ref_set = True
teg = self.video.state.gnss_position[i]
print(self.video.ten0)
print(self.video.state.gnss_position[i])
tn0g = np.matmul(trans.Cen(self.video.ten0).T,(self.video.state.gnss_position[i] - self.video.ten0))
twb = self.video.state.wTbs[i].translation()
tn0.append(tn0g)
tw.append(twb)
if len(tn0) > 1:
tn0 = np.array(tn0)
tw = np.array(tw)
bl = np.linalg.norm(tn0[-1] - tn0[0])
print('GNSS Alignment Baseline: %.5f' % bl)
if bl < 10.0:
print('Baseline too short!!')
return
heading_w = math.atan2(tw[-1,1]-tw[0,1],tw[-1,0]-tw[0,0])
heading_n0 = math.atan2(tn0[-1,1]-tn0[0,1],tn0[-1,0]-tn0[0,0])
s_w = np.linalg.norm(tw[-1] - tw[0])
s_n0 = np.linalg.norm(tn0[-1] - tn0[0])
s = s_n0 / s_w
Rn0w = trans.att2m(np.array([.0,.0,-heading_w + heading_n0]))
tn0w = tn0 - np.matmul(Rn0w,tw.T * s).T
poses = SE3(self.video.poses)
wTcs = poses.inv().matrix().cpu().numpy()
wTbs = np.matmul(wTcs,self.video.Tbc.inverse().matrix())
wTbs[:,0:3,3] = np.matmul(Rn0w,(wTbs[:,0:3,3]*s).T).T + tn0w[0]
wTbs[:,0:3,0:3] = np.matmul(Rn0w, (wTbs[:,0:3,0:3]).T).T
self.refTw = np.eye(4,4)
for i in range(0,self.t1):
self.video.state.wTbs[i] = gtsam.Pose3(wTbs[i])
self.video.state.vs[i] *= s
wTcs = np.matmul(wTbs,self.video.Tbc.matrix())
for i in range(0,self.t1):
TTT = np.linalg.inv(wTcs[i])
q = torch.tensor(Rotation.from_matrix(TTT[:3, :3]).as_quat())
t = torch.tensor(TTT[:3,3])
self.video.poses[i] = torch.cat([t,q])
self.video.disps[i] /= s
self.video.gnss_init_t1 = self.t1
self.video.gnss_init_time = self.video.tstamp[self.t1-1]
self.video.set_prior(self.video.last_t0,self.t1)
self.plt_pos = [[],[]]
self.plt_pos_ref = [[],[]]
for i in range(0,self.t1):
TTT = self.video.state.wTbs[i].matrix()
ppp = TTT[0:3,3]
qqq = Rotation.from_matrix(TTT[:3, :3]).as_quat()
self.result_file.writelines('%.6f %.6f %.6f %.6f %.6f %.6f %.6f %.6f\n'%(self.video.tstamp[i],ppp[0],ppp[1],ppp[2]\
,qqq[0],qqq[1],qqq[2],qqq[3]))
TTTref = np.matmul(self.refTw,TTT) # for visualization
ppp = TTTref[0:3,3]
qqq = Rotation.from_matrix(TTTref[:3, :3]).as_quat()
self.plt_pos[0].append(ppp[0])
self.plt_pos[1].append(ppp[1])
if self.all_gt is not None:
tt_found,dd = self.get_pose_ref(self.video.tstamp[i]-1e-3)
self.plt_pos_ref[0].append(dd['T'][0,3])
self.plt_pos_ref[1].append(dd['T'][1,3])
if self.show_plot:
plt.subplot(1,3,1)
plt.cla(); plt.gca().set_title('Trajectory')
plt.plot(self.plt_pos[0],self.plt_pos[1],marker='^')
plt.plot(self.plt_pos_ref[0],self.plt_pos_ref[1],marker='^')
plt.pause(0.1)
for itr in range(1):
self.graph.update(None, None, use_inactive=True)
print('GNSS initialized!!!!')
def VisualIMUAlignment(self, t0, t1, ignore_lever, disable_scale = False):
poses = SE3(self.video.poses)
wTcs = poses.inv().matrix().cpu().numpy()
if not ignore_lever:
wTbs = np.matmul(wTcs,self.video.Tbc.inverse().matrix())
else:
T_tmp = self.video.Tbc.inverse().matrix()
T_tmp[0:3,3] = 0.0
wTbs = np.matmul(wTcs,T_tmp)
cost = 0.0
# solveGyroscopeBias
A = np.zeros([3,3])
b = np.zeros(3)
H1 =np.zeros([15,6], order='F', dtype=np.float64)
H2 =np.zeros([15,3], order='F', dtype=np.float64)
H3 =np.zeros([15,6], order='F', dtype=np.float64)
H4 =np.zeros([15,3], order='F', dtype=np.float64)
H5 =np.zeros([15,6], order='F', dtype=np.float64) # navstate wrt. bias
H6 =np.zeros([15,6], order='F', dtype=np.float64)
for i in range(t0,t1-1):
pose_i = gtsam.Pose3(wTbs[i])
pose_j = gtsam.Pose3(wTbs[i+1])
Rij = np.matmul(pose_i.rotation().matrix().T,pose_j.rotation().matrix())
imu_factor = gtsam.gtsam.CombinedImuFactor(0,1,2,3,4,5,self.video.state.preintegrations[i])
err = imu_factor.evaluateErrorCustom(pose_i,self.video.state.vs[i],\
pose_j,self.video.state.vs[i+1],\
self.video.state.bs[i],self.video.state.bs[i+1],\
H1,H2,H3,H4,H5,H6)
tmp_A = H5[0:3,3:6]
tmp_b = err[0:3]
cost += np.dot(tmp_b,tmp_b)
A += np.matmul(tmp_A.T,tmp_A)
b += np.matmul(tmp_A.T,tmp_b)
bg = -np.matmul(np.linalg.inv(A),b)
for i in range(0,t1-1):
pim = gtsam.PreintegratedCombinedMeasurements(self.video.state.params,\
gtsam.imuBias.ConstantBias(np.array([.0,.0,.0]),bg))
for iii in range(len(self.video.state.preintegrations_meas[i])):
dd = self.video.state.preintegrations_meas[i][iii]
if dd[2] > 0: pim.integrateMeasurement(dd[0],dd[1],dd[2])
self.video.state.preintegrations[i] = pim
self.video.state.bs[i] = gtsam.imuBias.ConstantBias(np.array([.0,.0,.0]),bg)
print('bg: ',bg)
# linearAlignment
all_frame_count = t1 - t0
n_state = all_frame_count * 3 + 3 + 1
A = np.zeros([n_state,n_state])
b = np.zeros(n_state)
i_count = 0
for i in range(t0,t1-1):
pose_i = gtsam.Pose3(wTbs[i])
pose_j = gtsam.Pose3(wTbs[i+1])
R_i = pose_i.rotation().matrix()
t_i = pose_i.translation()
R_j = pose_j.rotation().matrix()
t_j = pose_j.translation()
pim = self.video.state.preintegrations[i]
tic = self.video.Tbc.translation()
tmp_A = np.zeros([6,10])
tmp_b = np.zeros(6)
dt = pim.deltaTij()
tmp_A[0:3,0:3] = -dt * np.eye(3,3)
tmp_A[0:3,6:9] = R_i.T * dt * dt / 2
tmp_A[0:3,9] = np.matmul(R_i.T, t_j-t_i) / 100.0
tmp_b[0:3] = pim.deltaPij()
tmp_A[3:6,0:3] = -np.eye(3,3)
tmp_A[3:6,3:6] = np.matmul(R_i.T, R_j)
tmp_A[3:6,6:9] = R_i.T * dt
tmp_b[3:6] = pim.deltaVij()
r_A = np.matmul(tmp_A.T,tmp_A)
r_b = np.matmul(tmp_A.T,tmp_b)
A[i_count*3:i_count*3+6,i_count*3:i_count*3+6] += r_A[0:6,0:6]
b[i_count*3:i_count*3+6] += r_b[0:6]
A[-4:,-4:] += r_A[-4:,-4:]
b[-4:] += r_b[-4:]
A[i_count*3:i_count*3+6,n_state-4:] += r_A[0:6,-4:]
A[n_state-4:,i_count*3:i_count*3+6] += r_A[-4:,0:6]
i_count += 1
A = A * 1000.0
b = b * 1000.0
x = np.matmul(np.linalg.inv(A),b)
s = x[n_state-1] / 100.0
g = x[-4:-1]
# RefineGravity
g0 = g / np.linalg.norm(g) * 9.81
lx = np.zeros(3)
ly = np.zeros(3)
n_state = all_frame_count * 3 + 2 + 1
A = np.zeros([n_state,n_state])
b = np.zeros(n_state)
for k in range(4):
aa = g / np.linalg.norm(g)
tmp = np.array([.0,.0,1.0])
bb = (tmp - np.dot(aa,tmp) * aa)
bb /= np.linalg.norm(bb)
cc = np.cross(aa,bb)
bc = np.zeros([3,2])
bc[0:3,0] = bb
bc[0:3,1] = cc
lxly = bc
i_count = 0
for i in range(t0,t1-1):
pose_i = gtsam.Pose3(wTbs[i])
pose_j = gtsam.Pose3(wTbs[i+1])
R_i = pose_i.rotation().matrix()
t_i = pose_i.translation()
R_j = pose_j.rotation().matrix()
t_j = pose_j.translation()
tmp_A = np.zeros([6,9])
tmp_b = np.zeros(6)
pim = self.video.state.preintegrations[i]
dt = pim.deltaTij()
tmp_A[0:3,0:3] = -dt *np.eye(3,3)
tmp_A[0:3,6:8] = np.matmul(R_i.T,lxly) * dt * dt /2
tmp_A[0:3,8] = np.matmul(R_i.T,t_j - t_i) / 100.0
tmp_b[0:3] = pim.deltaPij() - np.matmul(R_i.T,g0) * dt * dt / 2
tmp_A[3:6,0:3] = -np.eye(3)
tmp_A[3:6,3:6] = np.matmul(R_i.T,R_j)
tmp_A[3:6,6:8] = np.matmul(R_i.T,lxly) * dt
tmp_b[3:6] = pim.deltaVij() - np.matmul(R_i.T,g0) * dt
r_A = np.matmul(tmp_A.T,tmp_A)
r_b = np.matmul(tmp_A.T,tmp_b)
A[i_count*3:i_count*3+6,i_count*3:i_count*3+6] += r_A[0:6,0:6]
b[i_count*3:i_count*3+6] += r_b[0:6]
A[-3:,-3:] += r_A[-3:,-3:]
b[-3:] += r_b[-3:]
A[i_count*3:i_count*3+6,n_state-3:] += r_A[0:6,-3:]
A[n_state-3:,i_count*3:i_count*3+6] += r_A[-3:,0:6]
i_count += 1
A = A * 1000.0
b = b * 1000.0
x = np.matmul(np.linalg.inv(A),b)
dg = x[-3:-1]
g0 = g0 + np.matmul(lxly,dg)
g0 = g0 / np.linalg.norm(g0) * 9.81
s = x[-1] / 100.0
print(s,g0,x)
if disable_scale:
s = 1.0
print('g,s:',g,s)
if math.fabs(np.linalg.norm(g) - 9.81) < 0.5 and s > 0:
print('V-I successfully initialized!')
# visualInitialAlign
wTbs[:,0:3,3] *= s # !!!!!!!!!!!!!!!!!!!!!!!!
for i in range(0, t1-t0):
self.video.state.vs[i+t0] = np.matmul(wTbs[i+t0,0:3,0:3],x[i*3:i*3+3])
# g2R
ng1 = g0/ np.linalg.norm(g0)
ng2 = np.array([0,0,1.0])
R0 = trans.FromTwoVectors(ng1,ng2)
yaw = trans.R2ypr(R0)[0]
R0 = np.matmul(trans.ypr2R(np.array([-yaw,0,0])),R0)
# align for visualization
ppp = np.matmul(R0,wTbs[t1-1,0:3,3])
RRR = np.matmul(R0,wTbs[t1-1,0:3,0:3])
if self.all_gt is not None: # align the initial poses for visualization
tt_found,dd = self.get_pose_ref(self.video.tstamp[t1-1]-1e-3)
self.refTw = np.matmul(dd['T'],np.linalg.inv(wTbs[t1-1]))
self.refTw[0:3,0:3] = trans.att2m([0,0,trans.m2att(self.refTw[0:3,0:3])[2]])
g = np.matmul(R0,g0)
for i in range(0,t1):
wTbs[i,0:3,3] = np.matmul(R0,wTbs[i,0:3,3])
wTbs[i,0:3,0:3] = np.matmul(R0,wTbs[i,0:3,0:3])
self.video.state.vs[i] = np.matmul(R0, self.video.state.vs[i])
self.video.state.wTbs[i] = gtsam.Pose3(wTbs[i])
self.video.vi_init_t1 = t1
self.video.vi_init_time = self.video.tstamp[t1-1]
if not ignore_lever:
wTcs = np.matmul(wTbs,self.video.Tbc.matrix())
else:
T_tmp = self.video.Tbc.matrix()
T_tmp[0:3,3] = 0.0
wTcs = np.matmul(wTbs,T_tmp)
for i in range(0,t1):
TTT = np.linalg.inv(wTcs[i])
q = torch.tensor(Rotation.from_matrix(TTT[:3, :3]).as_quat())
t = torch.tensor(TTT[:3,3])
self.video.poses[i] = torch.cat([t,q])
self.video.disps[i] /= s
def __initialize(self):
""" initialize the SLAM system """
self.t0 = 0
self.t1 = self.video.counter.value
self.graph.add_neighborhood_factors(self.t0, self.t1, r=3)
self.init_IMU()
self.graph.video.imu_enabled = False
for itr in range(8):
self.graph.update(1, use_inactive=True)
self.graph.add_proximity_factors(0, 0, rad=2, nms=2, thresh=self.frontend_thresh, remove=False)
for itr in range(8):
self.graph.update(1, use_inactive=True)
self.graph.video.imu_enabled = False
for itr in range(8):
self.graph.update(1, use_inactive=True)
# torch.concat([self.graph.ii[None],self.graph.jj[None]]).T
# self.video.normalize()
self.video.poses[self.t1] = self.video.poses[self.t1-1].clone()
self.video.disps[self.t1] = self.video.disps[self.t1-4:self.t1].mean()
# initialization complete
self.is_initialized = True
with self.video.get_lock():
self.video.ready.value = 1
self.video.dirty[:self.t1] = True
self.graph.rm_factors(self.graph.ii < self.warmup-4, store=True)
def __call__(self):
""" main update """
# do initialization
if not self.is_initialized and self.video.counter.value == self.warmup:
self.__initialize()
# do update
elif self.is_initialized and self.t1 < self.video.counter.value:
self.__update()
================================================
FILE: dbaf/depth_video.py
================================================
import numpy as np
import torch
import lietorch
import droid_backends
from torch.multiprocessing import Process, Queue, Lock, Value
from droid_net import cvx_upsample
import geom.projective_ops as pops
from multi_sensor import MultiSensorState
import gtsam
from gtsam.symbol_shorthand import B, V, X
from scipy.spatial.transform import Rotation
import copy
import logging
import geoFunc.trans as trans
from lietorch import SE3
def BA2GTSAM(H: np.ndarray, v: np.ndarray, Tbc: gtsam.Pose3):
A = -Tbc.inverse().AdjointMap()
# A = -np.eye(6,6)
A = np.concatenate([A[3:6,:],A[0:3,:]],axis=0)
ss = H.shape[0]//6
J = np.zeros_like(H)
for i in range(ss):
J[(i*6):(i*6+6),(i*6):(i*6+6)] = A
JT = J.T
return np.matmul(np.matmul(JT,H),J),np.matmul(JT,v)
def CustomHessianFactor(values: gtsam.Values, H: np.ndarray, v: np.ndarray):
info_expand = np.zeros([H.shape[0]+1,H.shape[1]+1])
info_expand[0:-1,0:-1] = H
info_expand[0:-1,-1] = v
info_expand[-1,-1] = 0.0 # This is meaningless.
h_f = gtsam.HessianFactor(values.keys(),[6]*len(values.keys()),info_expand)
l_c = gtsam.LinearContainerFactor(h_f,values)
return l_c
class DepthVideo:
def __init__(self, image_size=[480, 640], buffer=1024, save_pkl = False, stereo=False, upsample = False, device="cuda:0"):
# current keyframe count
self.counter = Value('i', 0)
self.ready = Value('i', 0)
self.ht = ht = image_size[0]
self.wd = wd = image_size[1]
### state attributes ###
self.tstamp = torch.zeros(buffer, device="cuda", dtype=torch.float64).share_memory_()
self.images = torch.zeros(buffer, 3, ht, wd, device="cuda", dtype=torch.uint8)
self.dirty = torch.zeros(buffer, device="cuda", dtype=torch.bool).share_memory_()
self.red = torch.zeros(buffer, device="cuda", dtype=torch.bool).share_memory_()
self.poses = torch.zeros(buffer, 7, device="cuda", dtype=torch.float).share_memory_()
self.disps = torch.ones(buffer, ht//8, wd//8, device="cuda", dtype=torch.float).share_memory_()
self.disps_sens = torch.zeros(buffer, ht//8, wd//8, device="cuda", dtype=torch.float).share_memory_()
self.disps_up = torch.zeros(buffer, ht, wd, device="cuda", dtype=torch.float).share_memory_()
self.intrinsics = torch.zeros(buffer, 4, device="cuda", dtype=torch.float).share_memory_()
self.stereo = stereo
c = 1 if not self.stereo else 2
### feature attributes ###
self.fmaps = torch.zeros(buffer, c, 128, ht//8, wd//8, dtype=torch.half, device="cuda").share_memory_()
self.nets = torch.zeros(buffer, 128, ht//8, wd//8, dtype=torch.half, device="cuda").share_memory_()
self.inps = torch.zeros(buffer, 128, ht//8, wd//8, dtype=torch.half, device="cuda").share_memory_()
# initialize poses to identity transformation
self.poses[:] = torch.as_tensor([0, 0, 0, 0, 0, 0, 1], dtype=torch.float, device="cuda")
### DBAFusion
# for .pkl saving
self.disps_save = torch.ones(5000, ht//8, wd//8, device="cuda", dtype=torch.float)
self.poses_save = torch.ones(5000, 7, device="cuda", dtype=torch.float)
self.tstamp_save = torch.zeros(5000, device="cuda", dtype=torch.float64)
self.images_save = torch.zeros(5000, ht//8, wd//8, 3, device="cuda", dtype=torch.float)
if upsample:
self.disps_up_save = torch.zeros(5000, ht, wd, device="cuda", dtype=torch.float).share_memory_()
self.count_save = 0
self.save_pkl = save_pkl
self.upsample_flag = upsample
self.state = MultiSensorState()
self.last_t0 = 0
self.last_t1 = 0
self.cur_graph = None
self.cur_result = None
self.marg_factor = None
self.prior_factor = []
self.prior_factor_map = {}
self.cur_ii = None
self.cur_jj = None
self.cur_target = None
self.cur_weight = None
self.cur_eta = None
self.imu_enabled = False
self.ignore_imu = False
self.xyz_ref = []
# extrinsics, need to be set in the main .py
self.Ti1c = None # shape = (4,4)
self.Tbc = None # gtsam.Pose3
self.tbg = None # shape = (3)
self.reinit = False
self.vi_init_t1 = -1
self.vi_init_time = 0.0
self.gnss_init_t1 = -1
self.gnss_init_time = 0.0
self.ten0 = None
self.init_pose_sigma =np.array([0.1, 0.1, 0.0001, 0.0001,0.0001,0.0001])
self.init_bias_sigma =np.array([1.0,1.0,1.0, 0.1, 0.1, 0.1])
self.logger = logging.getLogger('dba_fusion')
self.logger.setLevel(logging.DEBUG)
fh = logging.FileHandler('dba_fusion.log')
formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
fh.setFormatter(formatter)
# add the handlers to the logger
self.logger.addHandler(fh)
self.logger.info('Start logging!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!')
def get_lock(self):
return self.counter.get_lock()
def __item_setter(self, index, item):
if isinstance(index, int) and index >= self.counter.value:
self.counter.value = index + 1
elif isinstance(index, torch.Tensor) and index.max().item() > self.counter.value:
self.counter.value = index.max().item() + 1
self.tstamp[index] = item[0]
self.images[index] = item[1]
if item[2] is not None:
self.poses[index] = item[2]
if item[3] is not None:
self.disps[index] = item[3]
if item[4] is not None:
depth = item[4][3::8,3::8]
self.disps_sens[index] = torch.where(depth>0, 1.0/depth, depth)
if item[5] is not None:
self.intrinsics[index] = item[5]
if len(item) > 6:
self.fmaps[index] = item[6]
if len(item) > 7:
self.nets[index] = item[7]
if len(item) > 8:
self.inps[index] = item[8]
def __setitem__(self, index, item):
with self.get_lock():
self.__item_setter(index, item)
def __getitem__(self, index):
""" index the depth video """
with self.get_lock():
# support negative indexing
if isinstance(index, int) and index < 0:
index = self.counter.value + index
item = (
self.poses[index],
self.disps[index],
self.intrinsics[index],
self.fmaps[index],
self.nets[index],
self.inps[index])
return item
def append(self, *item):
with self.get_lock():
self.__item_setter(self.counter.value, item)
### geometric operations ###
@staticmethod
def format_indicies(ii, jj):
""" to device, long, {-1} """
if not isinstance(ii, torch.Tensor):
ii = torch.as_tensor(ii)
if not isinstance(jj, torch.Tensor):
jj = torch.as_tensor(jj)
ii = ii.to(device="cuda", dtype=torch.long).reshape(-1)
jj = jj.to(device="cuda", dtype=torch.long).reshape(-1)
return ii, jj
def upsample(self, ix, mask):
""" upsample disparity """
disps_up = cvx_upsample(self.disps[ix].unsqueeze(-1), mask)
self.disps_up[ix] = disps_up.squeeze()
def normalize(self):
""" normalize depth and poses """
with self.get_lock():
s = self.disps[:self.counter.value].mean()
self.disps[:self.counter.value] /= s
self.poses[:self.counter.value,:3] *= s
self.dirty[:self.counter.value] = True
def reproject(self, ii, jj):
""" project points from ii -> jj """
ii, jj = DepthVideo.format_indicies(ii, jj)
Gs = lietorch.SE3(self.poses[None])
coords, valid_mask = \
pops.projective_transform(Gs, self.disps[None], self.intrinsics[None], ii, jj)
return coords, valid_mask
def reproject_comp(self, ii, jj, xyz_comp):
ii, jj = DepthVideo.format_indicies(ii,jj)
Gs = lietorch.SE3(self.poses[None])
coords, valid_mask = \
pops.projective_transform_comp(Gs, self.disps[None], self.intrinsics[None], ii, jj, xyz_comp)
return coords, valid_mask
def distance(self, ii=None, jj=None, beta=0.3, bidirectional=True):
""" frame distance metric """
return_matrix = False
if ii is None:
return_matrix = True
N = self.counter.value
ii, jj = torch.meshgrid(torch.arange(N), torch.arange(N))
ii, jj = DepthVideo.format_indicies(ii, jj)
if bidirectional:
poses = self.poses[:self.counter.value].clone()
d1 = droid_backends.frame_distance(
poses, self.disps, self.intrinsics[0], ii, jj, beta)
d2 = droid_backends.frame_distance(
poses, self.disps, self.intrinsics[0], jj, ii, beta)
d = .5 * (d1 + d2)
else:
d = droid_backends.frame_distance(
self.poses, self.disps, self.intrinsics[0], ii, jj, beta)
if return_matrix:
return d.reshape(N, N)
return d
def rm_new_gnss(self, t1):
if (self.gnss_init_t1> 0 and self.state.gnss_valid[t1]) or self.state.odo_valid[t1]:
graph_temp = gtsam.NonlinearFactorGraph()
linear_point = self.marg_factor.linearizationPoint()
graph_temp.push_back(self.marg_factor)
if self.state.gnss_valid[t1]:
T1 = self.state.wTbs[t1]
T0 = self.state.wTbs[t1-1]
p = np.matmul(trans.Cen(self.ten0).T, self.state.gnss_position[t1] - self.ten0)
n0pbg = self.state.wTbs[t1].rotation().rotate(self.tbg)
p = p - n0pbg
p = p - T1.translation() + T0.translation()
if not linear_point.exists(X(t1-1)):
linear_point.insert(X(t1-1), self.cur_result.atPose3(X(t1-1)))
gnss_factor = gtsam.GPSFactor(X(t1-1), p,\
gtsam.noiseModel.Robust.Create(\
gtsam.noiseModel.mEstimator.Cauchy(0.08),\
gtsam.noiseModel.Diagonal.Sigmas(np.array([1.0,1.0,5.0]))))
graph_temp.push_back(gnss_factor)
if self.state.odo_valid[t1]:
v1 = np.matmul(self.state.wTbs[t1].rotation().matrix().T, self.state.vs[t1])
v0 = np.matmul(self.state.wTbs[t1-1].rotation().matrix().T, self.state.vs[t1-1])
v = self.state.odo_vel[t1] - v1 + v0
if not linear_point.exists(X(t1-1)):
linear_point.insert(X(t1-1), self.cur_result.atPose3(X(t1-1)))
if not linear_point.exists(V(t1-1)):
linear_point.insert(V(t1-1), self.cur_result.atVector(V(t1-1)))
odo_factor = gtsam.VelFactor(X(t1-1),V(t1-1),v,gtsam.noiseModel.Diagonal.Sigmas(np.array([2.0,2.0,2.0])))
graph_temp.push_back(odo_factor)
h_factor = graph_temp.linearizeToHessianFactor(linear_point)
self.marg_factor = gtsam.LinearContainerFactor(h_factor,linear_point)
def set_prior(self, t0, t1):
for i in range(t0,t0+2):
self.prior_factor_map[i] = []
init_pose_sigma = self.init_pose_sigma
if len(self.init_pose_sigma.shape) > 1:
init_pose_sigma = self.init_pose_sigma[i-t0]
self.prior_factor_map[i].append(gtsam.PriorFactorPose3(X(i),\
self.state.wTbs[i], \
gtsam.noiseModel.Diagonal.Sigmas(init_pose_sigma)))
if not self.ignore_imu:
self.prior_factor_map[i].append(gtsam.PriorFactorConstantBias(B(i),\
self.state.bs[i], \
gtsam.noiseModel.Diagonal.Sigmas(self.init_bias_sigma)))
self.last_t0 = t0
self.last_t1 = t1
def ba(self, target, weight, eta, ii, jj, t0=1, t1=None, itrs=2, lm=1e-4, ep=0.1, motion_only=False):
""" dense bundle adjustment (DBA) """
with self.get_lock():
if t1 is None:
t1 = max(ii.max().item(), jj.max().item()) + 1
# 1) visual-only BA
# 2) multi-sensor BA
if not self.imu_enabled:
droid_backends.ba(self.poses, self.disps, self.intrinsics[0], self.disps_sens,
target, weight, eta, ii, jj, t0, t1, itrs, lm, ep, motion_only)
for i in range(self.last_t0, min(ii.min().item(), jj.min().item())):
if self.save_pkl:
# save marginalized results
self.tstamp_save[self.count_save] = self.tstamp[i].clone()
self.disps_save[self.count_save] = self.disps[i].clone()
self.poses_save[self.count_save] = self.poses[i].clone()
if self.upsample_flag:
self.disps_up_save[self.count_save] = self.disps_up[i].clone()
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"?
self.count_save += 1
self.last_t0 = min(ii.min().item(), jj.min().item())
self.last_t1 = t1
else:
t0 = min(ii.min().item(), jj.min().item())
""" marginalization """
if self.last_t1!=t1 or self.last_t0 != t0:
if self.last_t0 > t0:
t0 = self.last_t0
elif self.last_t0 == t0:
t0 = self.last_t0
else:
marg_paras = []
# Construct a temporary factor graph (related to the old states) to obtain the marginalization information
graph = gtsam.NonlinearFactorGraph()
marg_idx = torch.logical_and(torch.greater_equal(self.cur_ii,self.last_t0),\
torch.less(self.cur_ii,t0))
marg_idx2 = torch.logical_and(torch.less(self.cur_ii,self.last_t1-2),\
torch.less(self.cur_jj,self.last_t1-2))
marg_idx = torch.logical_and(marg_idx,marg_idx2)
marg_ii = self.cur_ii[marg_idx]
marg_jj = self.cur_jj[marg_idx]
marg_t0 = self.last_t0
marg_t1 = t0 + 1
if len(marg_ii) > 0:
marg_t0 = self.last_t0
marg_t1 = torch.max(marg_jj).item()+1
marg_result = gtsam.Values()
for i in range(self.last_t0,marg_t1):
if i < t0:
marg_paras.append(X(i))
if self.save_pkl:
# save marginalized results
self.tstamp_save[self.count_save] = self.tstamp[i].clone()
self.disps_save[self.count_save] = self.disps[i].clone()
self.poses_save[self.count_save] = self.poses[i].clone()
if self.upsample_flag:
self.disps_up_save[self.count_save] = self.disps_up[i].clone()
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"?
self.count_save += 1
marg_result.insert(X(i), self.cur_result.atPose3(X(i)))
marg_target = self.cur_target[marg_idx]
marg_weight = self.cur_weight[marg_idx]
marg_eta = self.cur_eta[0:marg_t1-marg_t0]
bacore = droid_backends.BACore()
bacore.init(self.poses, self.disps, self.intrinsics[0], torch.zeros_like(self.disps_sens),
marg_target, marg_weight, marg_eta, marg_ii, marg_jj, marg_t0, marg_t1, itrs, lm, ep, motion_only)
H = torch.zeros([(marg_t1-marg_t0)*6,(marg_t1-marg_t0)*6],dtype=torch.float64,device='cpu')
v = torch.zeros([(marg_t1-marg_t0)*6],dtype=torch.float64,device='cpu')
bacore.hessian(H,v)
for i in range(6): H[i,i] += 0.00025 # for stability
# Hg,vg = BA2GTSAM(H,v,self.Tbc)
Hgg = gtsam.BA2GTSAM(H,v,self.Tbc)
Hg = Hgg[0:(marg_t1-marg_t0)*6,0:(marg_t1-marg_t0)*6]
vg = Hgg[0:(marg_t1-marg_t0)*6, (marg_t1-marg_t0)*6]
vis_factor = CustomHessianFactor(marg_result,Hg,vg)
graph.push_back(vis_factor)
for i in range(self.last_t0,marg_t1):
if i < t0:
if X(i) not in marg_paras:
marg_paras.append(X(i))
if not self.ignore_imu:
marg_paras.append(V(i))
marg_paras.append(B(i))
graph.push_back(gtsam.gtsam.CombinedImuFactor(\
X(i),V(i),X(i+1),V(i+1),B(i),B(i+1),\
self.state.preintegrations[i]))
if self.gnss_init_t1 > 0:
if self.state.gnss_valid[i]:
p = np.matmul(trans.Cen(self.ten0).T, self.state.gnss_position[i] - self.ten0)
n0pbg = self.state.wTbs[i].rotation().rotate(self.tbg)
p = p - n0pbg
gnss_factor = gtsam.GPSFactor(X(i), p,\
gtsam.noiseModel.Robust.Create(\
gtsam.noiseModel.mEstimator.Cauchy(0.08),\
gtsam.noiseModel.Diagonal.Sigmas(np.array([1.0,1.0,5.0]))))
graph.push_back(gnss_factor)
if self.state.odo_valid[i]:
vb = self.state.odo_vel[i]
odo_factor = gtsam.VelFactor(X(i),V(i),vb,gtsam.noiseModel.Diagonal.Sigmas(np.array([2.0,2.0,2.0])))
graph.push_back(odo_factor)
keys = self.prior_factor_map.keys()
for i in sorted(keys):
if i < t0:
for iii in range(len(self.prior_factor_map[i])):
graph.push_back(self.prior_factor_map[i][iii])
del self.prior_factor_map[i]
if not self.marg_factor == None:
graph.push_back(self.marg_factor)
self.marg_factor = gtsam.marginalizeOut(graph,self.cur_result,marg_paras)
# covariance inflation of IMU biases
if self.reinit == True:
all_keys = self.marg_factor.keys()
for i in range(len(all_keys)):
if all_keys[i] == B(t0):
all_keys[i] = B(0)
graph = gtsam.NonlinearFactorGraph()
graph.push_back(self.marg_factor.rekey(all_keys))
b_l = gtsam.BetweenFactorConstantBias(B(0),B(t0),gtsam.imuBias.ConstantBias(np.array([.0,.0,.0]),np.array([.0,.0,.0])),\
gtsam.noiseModel.Diagonal.Sigmas(self.init_bias_sigma))
graph.push_back(b_l)
result_tmp = self.marg_factor.linearizationPoint()
result_tmp.insert(B(0),result_tmp.atConstantBias(B(t0)))
self.marg_factor = gtsam.marginalizeOut(graph,result_tmp,[B(0)])
self.reinit = False
self.last_t0 = t0
self.last_t1 = t1
""" optimization """
H = torch.zeros([(t1-t0)*6,(t1-t0)*6],dtype=torch.float64,device='cpu')
v = torch.zeros([(t1-t0)*6],dtype=torch.float64,device='cpu')
dx = torch.zeros([(t1-t0)*6],dtype=torch.float64,device='cpu')
bacore = droid_backends.BACore()
active_index = torch.logical_and(ii>=t0,jj>=t0)
self.cur_ii = ii[active_index]
self.cur_jj = jj[active_index]
self.cur_target = target[active_index]
self.cur_weight = weight[active_index]
self.cur_eta = eta[(t0-ii.min().item()):]
bacore.init(self.poses, self.disps, self.intrinsics[0], self.disps_sens,
self.cur_target, self.cur_weight, self.cur_eta, self.cur_ii, self.cur_jj, t0, t1, itrs, lm, ep, motion_only)
self.cur_graph = gtsam.NonlinearFactorGraph()
params = gtsam.LevenbergMarquardtParams()#;params.setMaxIterations(1)
# imu factor
if not self.ignore_imu:
for i in range(t0,t1):
if i > t0:
imu_factor = gtsam.gtsam.CombinedImuFactor(\
X(i-1),V(i-1),X(i),V(i),B(i-1),B(i),\
self.state.preintegrations[i-1])
self.cur_graph.add(imu_factor)
# prior factor
keys = self.prior_factor_map.keys()
for i in sorted(keys):
if i >= t0 and i < t1:
for iii in range(len(self.prior_factor_map[i])):
self.cur_graph.push_back(self.prior_factor_map[i][iii])
# marginalization factor
if self.marg_factor is not None:
self.cur_graph.push_back(self.marg_factor)
# GNSS factor
if self.gnss_init_t1 > 0:
for i in range(t0,t1):
if self.state.gnss_valid[i]:
p = np.matmul(trans.Cen(self.ten0).T, self.state.gnss_position[i] - self.ten0)
n0pbg = self.state.wTbs[i].rotation().rotate(self.tbg)
p = p - n0pbg
gnss_factor = gtsam.GPSFactor(X(i), p,\
gtsam.noiseModel.Robust.Create(\
gtsam.noiseModel.mEstimator.Cauchy(0.08),\
gtsam.noiseModel.Diagonal.Sigmas(np.array([1.0,1.0,5.0]))))
self.cur_graph.push_back(gnss_factor)
# Odo factor
for i in range(t0,t1):
if self.state.odo_valid[i]:
vb = self.state.odo_vel[i]
odo_factor = gtsam.VelFactor(X(i),V(i),vb,gtsam.noiseModel.Diagonal.Sigmas(np.array([2.0,2.0,2.0])))
self.cur_graph.push_back(odo_factor)
""" multi-sensor DBA iterations """
for iter in range(2):
if iter > 0:
self.cur_graph.resize(self.cur_graph.size()-1)
bacore.hessian(H,v) # camera frame
Hgg = gtsam.BA2GTSAM(H,v,self.Tbc)
Hg = Hgg[0:(t1-t0)*6,0:(t1-t0)*6]
vg = Hgg[0:(t1-t0)*6,(t1-t0)*6]
initial = gtsam.Values()
for i in range(t0,t1):
initial.insert(X(i), self.state.wTbs[i]) # the indice need to be handled
initial_vis = copy.deepcopy(initial)
vis_factor = CustomHessianFactor(initial_vis,Hg,vg)
self.cur_graph.push_back(vis_factor)
if not self.ignore_imu:
for i in range(t0,t1):
initial.insert(B(i),self.state.bs[i])
initial.insert(V(i),self.state.vs[i])
optimizer = gtsam.LevenbergMarquardtOptimizer(self.cur_graph, initial, params)
self.cur_result = optimizer.optimize()
# retraction and depth update
for i in range(t0,t1):
p0 = initial.atPose3(X(i))
p1 = self.cur_result.atPose3(X(i))
xi = gtsam.Pose3.Logmap(p0.inverse()*p1)
dx[(i-t0)*6:(i-t0)*6+6] = torch.tensor(xi)
if not self.ignore_imu:
self.state.bs[i] = self.cur_result.atConstantBias(B(i))
self.state.vs[i] = self.cur_result.atVector(V(i))
self.state.wTbs[i] = self.cur_result.atPose3(X(i))
dx = torch.tensor(gtsam.GTSAM2BA(dx,self.Tbc))
dx_dz = bacore.retract(dx)
del bacore
self.disps.clamp_(min=0.001)
================================================
FILE: dbaf/droid_net.py
================================================
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
from collections import OrderedDict
from modules.extractor import BasicEncoder
from modules.corr import CorrBlock
from modules.gru import ConvGRU
from modules.clipping import GradientClip
from lietorch import SE3
from geom.ba import BA
import geom.projective_ops as pops
from geom.graph_utils import graph_to_edge_list, keyframe_indicies
from torch_scatter import scatter_mean
import time
def cvx_upsample(data, mask):
""" upsample pixel-wise transformation field """
batch, ht, wd, dim = data.shape
data = data.permute(0, 3, 1, 2)
mask = mask.view(batch, 1, 9, 8, 8, ht, wd)
mask = torch.softmax(mask, dim=2)
up_data = F.unfold(data, [3,3], padding=1)
up_data = up_data.view(batch, dim, 9, 1, 1, ht, wd)
up_data = torch.sum(mask * up_data, dim=2)
up_data = up_data.permute(0, 4, 2, 5, 3, 1)
up_data = up_data.reshape(batch, 8*ht, 8*wd, dim)
return up_data
def upsample_disp(disp, mask):
batch, num, ht, wd = disp.shape
disp = disp.view(batch*num, ht, wd, 1)
mask = mask.view(batch*num, -1, ht, wd)
return cvx_upsample(disp, mask).view(batch, num, 8*ht, 8*wd)
class GraphAgg(nn.Module):
def __init__(self):
super(GraphAgg, self).__init__()
self.conv1 = nn.Conv2d(128, 128, 3, padding=1)
self.conv2 = nn.Conv2d(128, 128, 3, padding=1)
self.relu = nn.ReLU(inplace=True)
self.eta = nn.Sequential(
nn.Conv2d(128, 1, 3, padding=1),
GradientClip(),
nn.Softplus())
self.upmask = nn.Sequential(
nn.Conv2d(128, 8*8*9, 1, padding=0))
def forward(self, net, ii):
batch, num, ch, ht, wd = net.shape
net = net.view(batch*num, ch, ht, wd)
_, ix = torch.unique(ii, return_inverse=True)
net = self.relu(self.conv1(net))
net = net.view(batch, num, 128, ht, wd)
net = scatter_mean(net, ix, dim=1)
net = net.view(-1, 128, ht, wd)
net = self.relu(self.conv2(net))
eta = self.eta(net).view(batch, -1, ht, wd)
upmask = self.upmask(net).view(batch, -1, 8*8*9, ht, wd)
return .01 * eta, upmask
class UpdateModule(nn.Module):
def __init__(self):
super(UpdateModule, self).__init__()
cor_planes = 4 * (2*3 + 1)**2
self.corr_encoder = nn.Sequential(
nn.Conv2d(cor_planes, 128, 1, padding=0),
nn.ReLU(inplace=True),
nn.Conv2d(128, 128, 3, padding=1),
nn.ReLU(inplace=True))
self.flow_encoder = nn.Sequential(
nn.Conv2d(4, 128, 7, padding=3),
nn.ReLU(inplace=True),
nn.Conv2d(128, 64, 3, padding=1),
nn.ReLU(inplace=True))
self.weight = nn.Sequential(
nn.Conv2d(128, 128, 3, padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(128, 2, 3, padding=1),
GradientClip(),
nn.Sigmoid())
self.delta = nn.Sequential(
nn.Conv2d(128, 128, 3, padding=1),
nn.ReLU(inplace=True),
nn.Conv2d(128, 2, 3, padding=1),
GradientClip())
self.gru = ConvGRU(128, 128+128+64)
self.agg = GraphAgg()
def forward(self, net, inp, corr, flow=None, ii=None, jj=None, upsample = False):
""" RaftSLAM update operator """
batch, num, ch, ht, wd = net.shape
if flow is None:
flow = torch.zeros(batch, num, 4, ht, wd, device=net.device)
output_dim = (batch, num, -1, ht, wd)
net = net.view(batch*num, -1, ht, wd)
inp = inp.view(batch*num, -1, ht, wd)
corr = corr.view(batch*num, -1, ht, wd)
flow = flow.view(batch*num, -1, ht, wd)
corr = self.corr_encoder(corr)
flow = self.flow_encoder(flow)
net = self.gru(net, inp, corr, flow)
### update variables ###
delta = self.delta(net).view(*output_dim)
weight = self.weight(net).view(*output_dim)
delta = delta.permute(0,1,3,4,2)[...,:2].contiguous()
weight = weight.permute(0,1,3,4,2)[...,:2].contiguous()
net = net.view(*output_dim)
if ii is not None:
### ATTENTION!!!! ###
# We found this useless for VIO performance, thus disable it to save computation.
# Feel free to re-enable it.
if upsample:
eta, upmask = self.agg(net, ii.to(net.device))
return net, delta, weight, eta, upmask
else:
return net, delta, weight, None, None
else:
return net, delta, weight
class DroidNet(nn.Module):
def __init__(self):
super(DroidNet, self).__init__()
self.fnet = BasicEncoder(output_dim=128, norm_fn='instance')
self.cnet = BasicEncoder(output_dim=256, norm_fn='none')
self.update = UpdateModule()
def extract_features(self, images):
""" run feeature extraction networks """
# normalize images
images = images[:, :, [2,1,0]] / 255.0
mean = torch.as_tensor([0.485, 0.456, 0.406], device=images.device)
std = torch.as_tensor([0.229, 0.224, 0.225], device=images.device)
images = images.sub_(mean[:, None, None]).div_(std[:, None, None])
fmaps = self.fnet(images)
net = self.cnet(images)
net, inp = net.split([128,128], dim=2)
net = torch.tanh(net)
inp = torch.relu(inp)
return fmaps, net, inp
def forward(self, Gs, images, disps, intrinsics, graph=None, num_steps=12, fixedp=2):
""" Estimates SE3 or Sim3 between pair of frames """
u = keyframe_indicies(graph)
ii, jj, kk = graph_to_edge_list(graph)
ii = ii.to(device=images.device, dtype=torch.long)
jj = jj.to(device=images.device, dtype=torch.long)
fmaps, net, inp = self.extract_features(images)
net, inp = net[:,ii], inp[:,ii]
corr_fn = CorrBlock(fmaps[:,ii], fmaps[:,jj], num_levels=4, radius=3)
ht, wd = images.shape[-2:]
coords0 = pops.coords_grid(ht//8, wd//8, device=images.device)
coords1, _ = pops.projective_transform(Gs, disps, intrinsics, ii, jj)
target = coords1.clone()
Gs_list, disp_list, residual_list = [], [], []
for step in range(num_steps):
Gs = Gs.detach()
disps = disps.detach()
coords1 = coords1.detach()
target = target.detach()
# extract motion features
corr = corr_fn(coords1)
resd = target - coords1
flow = coords1 - coords0
motion = torch.cat([flow, resd], dim=-1)
motion = motion.permute(0,1,4,2,3).clamp(-64.0, 64.0)
net, delta, weight, eta, upmask = \
self.update(net, inp, corr, motion, ii, jj)
target = coords1 + delta
for i in range(2):
Gs, disps = BA(target, weight, eta, Gs, disps, intrinsics, ii, jj, fixedp=2)
coords1, valid_mask = pops.projective_transform(Gs, disps, intrinsics, ii, jj)
residual = (target - coords1)
Gs_list.append(Gs)
disp_list.append(upsample_disp(disps, upmask))
residual_list.append(valid_mask * residual)
return Gs_list, disp_list, residual_list
================================================
FILE: dbaf/geoFunc/__init__.py
================================================
================================================
FILE: dbaf/geoFunc/const_value.py
================================================
import math
pi=math.pi
a = 6378137
finv = 298.257223563
================================================
FILE: dbaf/geoFunc/trans.py
================================================
import math
from math import atan2, sin, cos
from . import const_value
import numpy as np
from scipy.spatial.transform import Rotation
def cart2geod(Xinput):
X=Xinput[0]
Y=Xinput[1]
Z=Xinput[2]
tolsq = 1e-10
maxit = 10
rtd = 180/const_value.pi
esq = (2-1/const_value.finv) / const_value.finv
oneesq = 1-esq
P=math.sqrt(X*X+Y*Y)
if P > 1e-20:
dlambda = math.atan2(Y,X) *rtd
else:
dlambda = 0
if dlambda <0:
dlambda = dlambda +360
r=math.sqrt(P*P+Z*Z)
if r>1e-20:
sinphi=Z/r
else :
sinphi=0
dphi = math.asin(sinphi)
if r<1e-20:
h=0
return
h = r-const_value.a*(1-sinphi*sinphi/const_value.finv)
for i in range(maxit):
sinphi = math.sin(dphi)
cosphi = math.cos(dphi)
N_phi = const_value.a/math.sqrt(1-esq*sinphi*sinphi)
dP =P -(N_phi+h)*cosphi
dZ=Z-(N_phi*oneesq+h)*sinphi
h=h+(sinphi*dZ+cosphi*dP)
dphi =dphi+(cosphi*dZ - sinphi*dP)/(N_phi+h)
if dP*dP + dZ*dZ= 0) & (jj >= 0) & (ii < n) & (jj < m)
return scatter_sum(A[:,v], ii[v]*m + jj[v], dim=1, dim_size=n*m)
def safe_scatter_add_vec(b, ii, n):
v = (ii >= 0) & (ii < n)
return scatter_sum(b[:,v], ii[v], dim=1, dim_size=n)
# apply retraction operator to inv-depth maps
def disp_retr(disps, dz, ii):
ii = ii.to(device=dz.device)
return disps + scatter_sum(dz, ii, dim=1, dim_size=disps.shape[1])
# apply retraction operator to poses
def pose_retr(poses, dx, ii):
ii = ii.to(device=dx.device)
return poses.retr(scatter_sum(dx, ii, dim=1, dim_size=poses.shape[1]))
def BA(target, weight, eta, poses, disps, intrinsics, ii, jj, fixedp=1, rig=1):
""" Full Bundle Adjustment """
B, P, ht, wd = disps.shape
N = ii.shape[0]
D = poses.manifold_dim
### 1: commpute jacobians and residuals ###
coords, valid, (Ji, Jj, Jz) = pops.projective_transform(
poses, disps, intrinsics, ii, jj, jacobian=True)
r = (target - coords).view(B, N, -1, 1)
w = .001 * (valid * weight).view(B, N, -1, 1)
### 2: construct linear system ###
Ji = Ji.reshape(B, N, -1, D)
Jj = Jj.reshape(B, N, -1, D)
wJiT = (w * Ji).transpose(2,3)
wJjT = (w * Jj).transpose(2,3)
Jz = Jz.reshape(B, N, ht*wd, -1)
Hii = torch.matmul(wJiT, Ji) # [B,num,6,6]
Hij = torch.matmul(wJiT, Jj)
Hji = torch.matmul(wJjT, Ji)
Hjj = torch.matmul(wJjT, Jj)
vi = torch.matmul(wJiT, r).squeeze(-1) # [B,num,6]
vj = torch.matmul(wJjT, r).squeeze(-1)
Ei = (wJiT.view(B,N,D,ht*wd,-1) * Jz[:,:,None]).sum(dim=-1) # [B,num,6,ht*wd]
Ej = (wJjT.view(B,N,D,ht*wd,-1) * Jz[:,:,None]).sum(dim=-1) # [B,num,6,ht*wd]
w = w.view(B, N, ht*wd, -1)
r = r.view(B, N, ht*wd, -1)
wk = torch.sum(w*r*Jz, dim=-1)
Ck = torch.sum(w*Jz*Jz, dim=-1) # [B,num,ht*wd]
kx, kk = torch.unique(ii, return_inverse=True)
M = kx.shape[0]
# only optimize keyframe poses
P = P // rig - fixedp
ii = ii // rig - fixedp
jj = jj // rig - fixedp
H = safe_scatter_add_mat(Hii, ii, ii, P, P) + \
safe_scatter_add_mat(Hij, ii, jj, P, P) + \
safe_scatter_add_mat(Hji, jj, ii, P, P) + \
safe_scatter_add_mat(Hjj, jj, jj, P, P)
E = safe_scatter_add_mat(Ei, ii, kk, P, M) + \
safe_scatter_add_mat(Ej, jj, kk, P, M)
v = safe_scatter_add_vec(vi, ii, P) + \
safe_scatter_add_vec(vj, jj, P)
C = safe_scatter_add_vec(Ck, kk, M)
w = safe_scatter_add_vec(wk, kk, M)
C = C + eta.view(*C.shape) + 1e-7
H = H.view(B, P, P, D, D)
E = E.view(B, P, M, D, ht*wd)
### 3: solve the system ###
dx, dz = schur_solve(H, E, C, v, w)
### 4: apply retraction ###
poses = pose_retr(poses, dx, torch.arange(P) + fixedp)
disps = disp_retr(disps, dz.view(B,-1,ht,wd), kx)
disps = torch.where(disps > 10, torch.zeros_like(disps), disps)
disps = disps.clamp(min=0.0)
return poses, disps
def MoBA(target, weight, eta, poses, disps, intrinsics, ii, jj, fixedp=1, rig=1):
""" Motion only bundle adjustment """
B, P, ht, wd = disps.shape
N = ii.shape[0]
D = poses.manifold_dim
### 1: commpute jacobians and residuals ###
coords, valid, (Ji, Jj, Jz) = pops.projective_transform(
poses, disps, intrinsics, ii, jj, jacobian=True)
r = (target - coords).view(B, N, -1, 1)
w = .001 * (valid * weight).view(B, N, -1, 1)
### 2: construct linear system ###
Ji = Ji.reshape(B, N, -1, D)
Jj = Jj.reshape(B, N, -1, D)
wJiT = (w * Ji).transpose(2,3)
wJjT = (w * Jj).transpose(2,3)
Hii = torch.matmul(wJiT, Ji)
Hij = torch.matmul(wJiT, Jj)
Hji = torch.matmul(wJjT, Ji)
Hjj = torch.matmul(wJjT, Jj)
vi = torch.matmul(wJiT, r).squeeze(-1)
vj = torch.matmul(wJjT, r).squeeze(-1)
# only optimize keyframe poses
P = P // rig - fixedp
ii = ii // rig - fixedp
jj = jj // rig - fixedp
H = safe_scatter_add_mat(Hii, ii, ii, P, P) + \
safe_scatter_add_mat(Hij, ii, jj, P, P) + \
safe_scatter_add_mat(Hji, jj, ii, P, P) + \
safe_scatter_add_mat(Hjj, jj, jj, P, P)
v = safe_scatter_add_vec(vi, ii, P) + \
safe_scatter_add_vec(vj, jj, P)
H = H.view(B, P, P, D, D)
### 3: solve the system ###
dx = block_solve(H, v)
### 4: apply retraction ###
poses = pose_retr(poses, dx, torch.arange(P) + fixedp)
return poses
================================================
FILE: dbaf/geom/chol.py
================================================
import torch
import torch.nn.functional as F
import geom.projective_ops as pops
class CholeskySolver(torch.autograd.Function):
@staticmethod
def forward(ctx, H, b):
# don't crash training if cholesky decomp fails
try:
U = torch.linalg.cholesky(H)
xs = torch.cholesky_solve(b, U)
ctx.save_for_backward(U, xs)
ctx.failed = False
except Exception as e:
print(e)
ctx.failed = True
xs = torch.zeros_like(b)
return xs
@staticmethod
def backward(ctx, grad_x):
if ctx.failed:
return None, None
U, xs = ctx.saved_tensors
dz = torch.cholesky_solve(grad_x, U)
dH = -torch.matmul(xs, dz.transpose(-1,-2))
return dH, dz
def block_solve(H, b, ep=0.1, lm=0.0001):
""" solve normal equations """
B, N, _, D, _ = H.shape
I = torch.eye(D).to(H.device)
H = H + (ep + lm*H) * I
H = H.permute(0,1,3,2,4)
H = H.reshape(B, N*D, N*D)
b = b.reshape(B, N*D, 1)
x = CholeskySolver.apply(H,b)
return x.reshape(B, N, D)
def schur_solve(H, E, C, v, w, ep=0.1, lm=0.0001, sless=False):
""" solve using shur complement """
B, P, M, D, HW = E.shape
H = H.permute(0,1,3,2,4).reshape(B, P*D, P*D)
E = E.permute(0,1,3,2,4).reshape(B, P*D, M*HW)
Q = (1.0 / C).view(B, M*HW, 1)
# damping
I = torch.eye(P*D).to(H.device)
H = H + (ep + lm*H) * I
v = v.reshape(B, P*D, 1)
w = w.reshape(B, M*HW, 1)
Et = E.transpose(1,2)
S = H - torch.matmul(E, Q*Et)
v = v - torch.matmul(E, Q*w)
dx = CholeskySolver.apply(S, v)
if sless:
return dx.reshape(B, P, D)
dz = Q * (w - Et @ dx)
dx = dx.reshape(B, P, D)
dz = dz.reshape(B, M, HW)
return dx, dz
================================================
FILE: dbaf/geom/graph_utils.py
================================================
import torch
import numpy as np
from collections import OrderedDict
import lietorch
from data_readers.rgbd_utils import compute_distance_matrix_flow, compute_distance_matrix_flow2
def graph_to_edge_list(graph):
ii, jj, kk = [], [], []
for s, u in enumerate(graph):
for v in graph[u]:
ii.append(u)
jj.append(v)
kk.append(s)
ii = torch.as_tensor(ii)
jj = torch.as_tensor(jj)
kk = torch.as_tensor(kk)
return ii, jj, kk
def keyframe_indicies(graph):
return torch.as_tensor([u for u in graph])
def meshgrid(m, n, device='cuda'):
ii, jj = torch.meshgrid(torch.arange(m), torch.arange(n))
return ii.reshape(-1).to(device), jj.reshape(-1).to(device)
def neighbourhood_graph(n, r):
ii, jj = meshgrid(n, n)
d = (ii - jj).abs()
keep = (d >= 1) & (d <= r)
return ii[keep], jj[keep]
def build_frame_graph(poses, disps, intrinsics, num=16, thresh=24.0, r=2):
""" construct a frame graph between co-visible frames """
N = poses.shape[1]
poses = poses[0].cpu().numpy()
disps = disps[0][:,3::8,3::8].cpu().numpy()
intrinsics = intrinsics[0].cpu().numpy() / 8.0
d = compute_distance_matrix_flow(poses, disps, intrinsics)
count = 0
graph = OrderedDict()
for i in range(N):
graph[i] = []
d[i,i] = np.inf
for j in range(i-r, i+r+1):
if 0 <= j < N and i != j:
graph[i].append(j)
d[i,j] = np.inf
count += 1
while count < num:
ix = np.argmin(d)
i, j = ix // N, ix % N
if d[i,j] < thresh:
graph[i].append(j)
d[i,j] = np.inf
count += 1
else:
break
return graph
def build_frame_graph_v2(poses, disps, intrinsics, num=16, thresh=24.0, r=2):
""" construct a frame graph between co-visible frames """
N = poses.shape[1]
# poses = poses[0].cpu().numpy()
# disps = disps[0].cpu().numpy()
# intrinsics = intrinsics[0].cpu().numpy()
d = compute_distance_matrix_flow2(poses, disps, intrinsics)
# import matplotlib.pyplot as plt
# plt.imshow(d)
# plt.show()
count = 0
graph = OrderedDict()
for i in range(N):
graph[i] = []
d[i,i] = np.inf
for j in range(i-r, i+r+1):
if 0 <= j < N and i != j:
graph[i].append(j)
d[i,j] = np.inf
count += 1
while 1:
ix = np.argmin(d)
i, j = ix // N, ix % N
if d[i,j] < thresh:
graph[i].append(j)
for i1 in range(i-1, i+2):
for j1 in range(j-1, j+2):
if 0 <= i1 < N and 0 <= j1 < N:
d[i1, j1] = np.inf
count += 1
else:
break
return graph
================================================
FILE: dbaf/geom/losses.py
================================================
from collections import OrderedDict
import numpy as np
import torch
from lietorch import SO3, SE3, Sim3
from .graph_utils import graph_to_edge_list
from .projective_ops import projective_transform
def pose_metrics(dE):
""" Translation/Rotation/Scaling metrics from Sim3 """
t, q, s = dE.data.split([3, 4, 1], -1)
ang = SO3(q).log().norm(dim=-1)
# convert radians to degrees
r_err = (180 / np.pi) * ang
t_err = t.norm(dim=-1)
s_err = (s - 1.0).abs()
return r_err, t_err, s_err
def fit_scale(Ps, Gs):
b = Ps.shape[0]
t1 = Ps.data[...,:3].detach().reshape(b, -1)
t2 = Gs.data[...,:3].detach().reshape(b, -1)
s = (t1*t2).sum(-1) / ((t2*t2).sum(-1) + 1e-8)
return s
def geodesic_loss(Ps, Gs, graph, gamma=0.9, do_scale=True):
""" Loss function for training network """
# relative pose
ii, jj, kk = graph_to_edge_list(graph)
dP = Ps[:,jj] * Ps[:,ii].inv()
n = len(Gs)
geodesic_loss = 0.0
for i in range(n):
w = gamma ** (n - i - 1)
dG = Gs[i][:,jj] * Gs[i][:,ii].inv()
if do_scale:
s = fit_scale(dP, dG)
dG = dG.scale(s[:,None])
# pose error
d = (dG * dP.inv()).log()
if isinstance(dG, SE3):
tau, phi = d.split([3,3], dim=-1)
geodesic_loss += w * (
tau.norm(dim=-1).mean() +
phi.norm(dim=-1).mean())
elif isinstance(dG, Sim3):
tau, phi, sig = d.split([3,3,1], dim=-1)
geodesic_loss += w * (
tau.norm(dim=-1).mean() +
phi.norm(dim=-1).mean() +
0.05 * sig.norm(dim=-1).mean())
dE = Sim3(dG * dP.inv()).detach()
r_err, t_err, s_err = pose_metrics(dE)
metrics = {
'rot_error': r_err.mean().item(),
'tr_error': t_err.mean().item(),
'bad_rot': (r_err < .1).float().mean().item(),
'bad_tr': (t_err < .01).float().mean().item(),
}
return geodesic_loss, metrics
def residual_loss(residuals, gamma=0.9):
""" loss on system residuals """
residual_loss = 0.0
n = len(residuals)
for i in range(n):
w = gamma ** (n - i - 1)
residual_loss += w * residuals[i].abs().mean()
return residual_loss, {'residual': residual_loss.item()}
def flow_loss(Ps, disps, poses_est, disps_est, intrinsics, graph, gamma=0.9):
""" optical flow loss """
N = Ps.shape[1]
graph = OrderedDict()
for i in range(N):
graph[i] = [j for j in range(N) if abs(i-j)==1]
ii, jj, kk = graph_to_edge_list(graph)
coords0, val0 = projective_transform(Ps, disps, intrinsics, ii, jj)
val0 = val0 * (disps[:,ii] > 0).float().unsqueeze(dim=-1)
n = len(poses_est)
flow_loss = 0.0
for i in range(n):
w = gamma ** (n - i - 1)
coords1, val1 = projective_transform(poses_est[i], disps_est[i], intrinsics, ii, jj)
v = (val0 * val1).squeeze(dim=-1)
epe = v * (coords1 - coords0).norm(dim=-1)
flow_loss += w * epe.mean()
epe = epe.reshape(-1)[v.reshape(-1) > 0.5]
metrics = {
'f_error': epe.mean().item(),
'1px': (epe<1.0).float().mean().item(),
}
return flow_loss, metrics
================================================
FILE: dbaf/geom/projective_ops.py
================================================
import torch
import torch.nn.functional as F
from lietorch import SE3, Sim3
MIN_DEPTH = 0.2
def extract_intrinsics(intrinsics):
return intrinsics[...,None,None,:].unbind(dim=-1)
def coords_grid(ht, wd, **kwargs):
y, x = torch.meshgrid(
torch.arange(ht).to(**kwargs).float(),
torch.arange(wd).to(**kwargs).float())
return torch.stack([x, y], dim=-1)
def iproj(disps, intrinsics, jacobian=False):
""" pinhole camera inverse projection """
ht, wd = disps.shape[2:]
fx, fy, cx, cy = extract_intrinsics(intrinsics)
y, x = torch.meshgrid(
torch.arange(ht).to(disps.device).float(),
torch.arange(wd).to(disps.device).float())
i = torch.ones_like(disps)
X = (x - cx) / fx
Y = (y - cy) / fy
pts = torch.stack([X, Y, i, disps], dim=-1)
if jacobian:
J = torch.zeros_like(pts)
J[...,-1] = 1.0
return pts, J
return pts, None
def proj(Xs, intrinsics, jacobian=False, return_depth=False):
""" pinhole camera projection """
fx, fy, cx, cy = extract_intrinsics(intrinsics)
X, Y, Z, D = Xs.unbind(dim=-1)
Z = torch.where(Z < 0.5*MIN_DEPTH, torch.ones_like(Z), Z)
d = 1.0 / Z
x = fx * (X * d) + cx
y = fy * (Y * d) + cy
if return_depth:
coords = torch.stack([x, y, D*d], dim=-1)
else:
coords = torch.stack([x, y], dim=-1)
if jacobian:
B, N, H, W = d.shape
o = torch.zeros_like(d)
proj_jac = torch.stack([
fx*d, o, -fx*X*d*d, o,
o, fy*d, -fy*Y*d*d, o,
# o, o, -D*d*d, d,
], dim=-1).view(B, N, H, W, 2, 4)
return coords, proj_jac
return coords, None
def actp(Gij, X0, jacobian=False):
""" action on point cloud """
X1 = Gij[:,:,None,None] * X0
if jacobian:
X, Y, Z, d = X1.unbind(dim=-1)
o = torch.zeros_like(d)
B, N, H, W = d.shape
if isinstance(Gij, SE3):
Ja = torch.stack([
d, o, o, o, Z, -Y,
o, d, o, -Z, o, X,
o, o, d, Y, -X, o,
o, o, o, o, o, o,
], dim=-1).view(B, N, H, W, 4, 6)
elif isinstance(Gij, Sim3):
Ja = torch.stack([
d, o, o, o, Z, -Y, X,
o, d, o, -Z, o, X, Y,
o, o, d, Y, -X, o, Z,
o, o, o, o, o, o, o
], dim=-1).view(B, N, H, W, 4, 7)
return X1, Ja
return X1, None
def projective_transform(poses, depths, intrinsics, ii, jj, jacobian=False, return_depth=False):
""" map points from ii->jj """
# inverse project (pinhole)
X0, Jz = iproj(depths[:,ii], intrinsics[:,ii], jacobian=jacobian)
# transform
Gij = poses[:,jj] * poses[:,ii].inv()
Gij.data[:,ii==jj] = torch.as_tensor([-0.1, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], device="cuda")
X1, Ja = actp(Gij, X0, jacobian=jacobian) # 4*6
# project (pinhole)
x1, Jp = proj(X1, intrinsics[:,jj], jacobian=jacobian, return_depth=return_depth) # 2*4
# exclude points too close to camera
valid = ((X1[...,2] > MIN_DEPTH) & (X0[...,2] > MIN_DEPTH)).float()
valid = valid.unsqueeze(-1)
if jacobian:
# Ji transforms according to dual adjoint
Jj = torch.matmul(Jp, Ja) # 2*6
Ji = -Gij[:,:,None,None,None].adjT(Jj) # 2*6 * 6*6
Jz = Gij[:,:,None,None] * Jz
Jz = torch.matmul(Jp, Jz.unsqueeze(-1))
return x1, valid, (Ji, Jj, Jz)
return x1, valid
def projective_transform_comp(poses, depths, intrinsics, ii, jj, xyz_comp, jacobian=False, return_depth=False):
""" map points from ii->jj """
# inverse project (pinhole)
X0, Jz = iproj(depths[:,ii], intrinsics[:,ii], jacobian=jacobian)
# transform
Gij = poses[:,jj] * poses[:,ii].inv()
Gij.data[:,ii==jj] = torch.as_tensor([-0.1, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0], device="cuda")
X1, Ja = actp(Gij, X0, jacobian=jacobian)
X1 = X1 + xyz_comp # compensate the object motion
# project (pinhole)
x1, Jp = proj(X1, intrinsics[:,jj], jacobian=jacobian, return_depth=return_depth)
# exclude points too close to camera
valid = ((X1[...,2] > MIN_DEPTH) & (X0[...,2] > MIN_DEPTH)).float()
valid = valid.unsqueeze(-1)
if jacobian:
# Ji transforms according to dual adjoint
Jj = torch.matmul(Jp, Ja)
Ji = -Gij[:,:,None,None,None].adjT(Jj)
Jz = Gij[:,:,None,None] * Jz
Jz = torch.matmul(Jp, Jz.unsqueeze(-1))
return x1, valid, (Ji, Jj, Jz)
return x1, valid
def induced_flow(poses, disps, intrinsics, ii, jj):
""" optical flow induced by camera motion """
ht, wd = disps.shape[2:]
y, x = torch.meshgrid(
torch.arange(ht).to(disps.device).float(),
torch.arange(wd).to(disps.device).float())
coords0 = torch.stack([x, y], dim=-1)
coords1, valid = projective_transform(poses, disps, intrinsics, ii, jj, False)
return coords1[...,:2] - coords0, valid
================================================
FILE: dbaf/modules/__init__.py
================================================
================================================
FILE: dbaf/modules/clipping.py
================================================
import torch
import torch.nn as nn
import torch.nn.functional as F
GRAD_CLIP = .01
class GradClip(torch.autograd.Function):
@staticmethod
def forward(ctx, x):
return x
@staticmethod
def backward(ctx, grad_x):
o = torch.zeros_like(grad_x)
grad_x = torch.where(grad_x.abs()>GRAD_CLIP, o, grad_x)
grad_x = torch.where(torch.isnan(grad_x), o, grad_x)
return grad_x
class GradientClip(nn.Module):
def __init__(self):
super(GradientClip, self).__init__()
def forward(self, x):
return GradClip.apply(x)
================================================
FILE: dbaf/modules/corr.py
================================================
import torch
import torch.nn.functional as F
import droid_backends
class CorrSampler(torch.autograd.Function):
@staticmethod
def forward(ctx, volume, coords, radius):
ctx.save_for_backward(volume,coords)
ctx.radius = radius
corr, = droid_backends.corr_index_forward(volume, coords, radius)
return corr
@staticmethod
def backward(ctx, grad_output):
volume, coords = ctx.saved_tensors
grad_output = grad_output.contiguous()
grad_volume, = droid_backends.corr_index_backward(volume, coords, grad_output, ctx.radius)
return grad_volume, None, None
class CorrBlock:
def __init__(self, fmap1, fmap2, num_levels=4, radius=3):
self.num_levels = num_levels
self.radius = radius
self.corr_pyramid = []
# all pairs correlation
corr = CorrBlock.corr(fmap1, fmap2)
batch, num, h1, w1, h2, w2 = corr.shape
corr = corr.reshape(batch*num*h1*w1, 1, h2, w2)
for i in range(self.num_levels):
self.corr_pyramid.append(
corr.view(batch*num, h1, w1, h2//2**i, w2//2**i))
corr = F.avg_pool2d(corr, 2, stride=2)
def __call__(self, coords):
out_pyramid = []
batch, num, ht, wd, _ = coords.shape
coords = coords.permute(0,1,4,2,3)
coords = coords.contiguous().view(batch*num, 2, ht, wd)
for i in range(self.num_levels):
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
out_pyramid.append(corr.view(batch, num, -1, ht, wd))
return torch.cat(out_pyramid, dim=2)
def cat(self, other):
for i in range(self.num_levels):
self.corr_pyramid[i] = torch.cat([self.corr_pyramid[i], other.corr_pyramid[i]], 0)
return self
def __getitem__(self, index):
for i in range(self.num_levels):
self.corr_pyramid[i] = self.corr_pyramid[i][index]
return self
@staticmethod
def corr(fmap1, fmap2):
""" all-pairs correlation """
batch, num, dim, ht, wd = fmap1.shape
fmap1 = fmap1.reshape(batch*num, dim, ht*wd) / 4.0
fmap2 = fmap2.reshape(batch*num, dim, ht*wd) / 4.0
corr = torch.matmul(fmap1.transpose(1,2), fmap2)
return corr.view(batch, num, ht, wd, ht, wd)
class CorrLayer(torch.autograd.Function):
@staticmethod
def forward(ctx, fmap1, fmap2, coords, r):
ctx.r = r
ctx.save_for_backward(fmap1, fmap2, coords)
corr, = droid_backends.altcorr_forward(fmap1, fmap2, coords, ctx.r)
return corr
@staticmethod
def backward(ctx, grad_corr):
fmap1, fmap2, coords = ctx.saved_tensors
grad_corr = grad_corr.contiguous()
fmap1_grad, fmap2_grad, coords_grad = \
droid_backends.altcorr_backward(fmap1, fmap2, coords, grad_corr, ctx.r)
return fmap1_grad, fmap2_grad, coords_grad, None
class AltCorrBlock:
def __init__(self, fmaps, num_levels=4, radius=3):
self.num_levels = num_levels
self.radius = radius
B, N, C, H, W = fmaps.shape
fmaps = fmaps.view(B*N, C, H, W) / 4.0
self.pyramid = []
for i in range(self.num_levels):
sz = (B, N, H//2**i, W//2**i, C)
fmap_lvl = fmaps.permute(0, 2, 3, 1).contiguous()
self.pyramid.append(fmap_lvl.view(*sz))
fmaps = F.avg_pool2d(fmaps, 2, stride=2)
def corr_fn(self, coords, ii, jj):
B, N, H, W, S, _ = coords.shape
coords = coords.permute(0, 1, 4, 2, 3, 5)
corr_list = []
for i in range(self.num_levels):
r = self.radius
fmap1_i = self.pyramid[0][:, ii]
fmap2_i = self.pyramid[i][:, jj]
coords_i = (coords / 2**i).reshape(B*N, S, H, W, 2).contiguous()
fmap1_i = fmap1_i.reshape((B*N,) + fmap1_i.shape[2:])
fmap2_i = fmap2_i.reshape((B*N,) + fmap2_i.shape[2:])
corr = CorrLayer.apply(fmap1_i.float(), fmap2_i.float(), coords_i, self.radius)
corr = corr.view(B, N, S, -1, H, W).permute(0, 1, 3, 4, 5, 2)
corr_list.append(corr)
corr = torch.cat(corr_list, dim=2)
return corr
def __call__(self, coords, ii, jj):
squeeze_output = False
if len(coords.shape) == 5:
coords = coords.unsqueeze(dim=-2)
squeeze_output = True
corr = self.corr_fn(coords, ii, jj)
if squeeze_output:
corr = corr.squeeze(dim=-1)
return corr.contiguous()
================================================
FILE: dbaf/modules/extractor.py
================================================
import torch
import torch.nn as nn
import torch.nn.functional as F
class ResidualBlock(nn.Module):
def __init__(self, in_planes, planes, norm_fn='group', stride=1):
super(ResidualBlock, self).__init__()
self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, padding=1, stride=stride)
self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1)
self.relu = nn.ReLU(inplace=True)
num_groups = planes // 8
if norm_fn == 'group':
self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
if not stride == 1:
self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
elif norm_fn == 'batch':
self.norm1 = nn.BatchNorm2d(planes)
self.norm2 = nn.BatchNorm2d(planes)
if not stride == 1:
self.norm3 = nn.BatchNorm2d(planes)
elif norm_fn == 'instance':
self.norm1 = nn.InstanceNorm2d(planes)
self.norm2 = nn.InstanceNorm2d(planes)
if not stride == 1:
self.norm3 = nn.InstanceNorm2d(planes)
elif norm_fn == 'none':
self.norm1 = nn.Sequential()
self.norm2 = nn.Sequential()
if not stride == 1:
self.norm3 = nn.Sequential()
if stride == 1:
self.downsample = None
else:
self.downsample = nn.Sequential(
nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm3)
def forward(self, x):
y = x
y = self.relu(self.norm1(self.conv1(y)))
y = self.relu(self.norm2(self.conv2(y)))
if self.downsample is not None:
x = self.downsample(x)
return self.relu(x+y)
class BottleneckBlock(nn.Module):
def __init__(self, in_planes, planes, norm_fn='group', stride=1):
super(BottleneckBlock, self).__init__()
self.conv1 = nn.Conv2d(in_planes, planes//4, kernel_size=1, padding=0)
self.conv2 = nn.Conv2d(planes//4, planes//4, kernel_size=3, padding=1, stride=stride)
self.conv3 = nn.Conv2d(planes//4, planes, kernel_size=1, padding=0)
self.relu = nn.ReLU(inplace=True)
num_groups = planes // 8
if norm_fn == 'group':
self.norm1 = nn.GroupNorm(num_groups=num_groups, num_channels=planes//4)
self.norm2 = nn.GroupNorm(num_groups=num_groups, num_channels=planes//4)
self.norm3 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
if not stride == 1:
self.norm4 = nn.GroupNorm(num_groups=num_groups, num_channels=planes)
elif norm_fn == 'batch':
self.norm1 = nn.BatchNorm2d(planes//4)
self.norm2 = nn.BatchNorm2d(planes//4)
self.norm3 = nn.BatchNorm2d(planes)
if not stride == 1:
self.norm4 = nn.BatchNorm2d(planes)
elif norm_fn == 'instance':
self.norm1 = nn.InstanceNorm2d(planes//4)
self.norm2 = nn.InstanceNorm2d(planes//4)
self.norm3 = nn.InstanceNorm2d(planes)
if not stride == 1:
self.norm4 = nn.InstanceNorm2d(planes)
elif norm_fn == 'none':
self.norm1 = nn.Sequential()
self.norm2 = nn.Sequential()
self.norm3 = nn.Sequential()
if not stride == 1:
self.norm4 = nn.Sequential()
if stride == 1:
self.downsample = None
else:
self.downsample = nn.Sequential(
nn.Conv2d(in_planes, planes, kernel_size=1, stride=stride), self.norm4)
def forward(self, x):
y = x
y = self.relu(self.norm1(self.conv1(y)))
y = self.relu(self.norm2(self.conv2(y)))
y = self.relu(self.norm3(self.conv3(y)))
if self.downsample is not None:
x = self.downsample(x)
return self.relu(x+y)
DIM=32
class BasicEncoder(nn.Module):
def __init__(self, output_dim=128, norm_fn='batch', dropout=0.0, multidim=False):
super(BasicEncoder, self).__init__()
self.norm_fn = norm_fn
self.multidim = multidim
if self.norm_fn == 'group':
self.norm1 = nn.GroupNorm(num_groups=8, num_channels=DIM)
elif self.norm_fn == 'batch':
self.norm1 = nn.BatchNorm2d(DIM)
elif self.norm_fn == 'instance':
self.norm1 = nn.InstanceNorm2d(DIM)
elif self.norm_fn == 'none':
self.norm1 = nn.Sequential()
self.conv1 = nn.Conv2d(3, DIM, kernel_size=7, stride=2, padding=3)
self.relu1 = nn.ReLU(inplace=True)
self.in_planes = DIM
self.layer1 = self._make_layer(DIM, stride=1)
self.layer2 = self._make_layer(2*DIM, stride=2)
self.layer3 = self._make_layer(4*DIM, stride=2)
# output convolution
self.conv2 = nn.Conv2d(4*DIM, output_dim, kernel_size=1)
if self.multidim:
self.layer4 = self._make_layer(256, stride=2)
self.layer5 = self._make_layer(512, stride=2)
self.in_planes = 256
self.layer6 = self._make_layer(256, stride=1)
self.in_planes = 128
self.layer7 = self._make_layer(128, stride=1)
self.up1 = nn.Conv2d(512, 256, 1)
self.up2 = nn.Conv2d(256, 128, 1)
self.conv3 = nn.Conv2d(128, output_dim, kernel_size=1)
if dropout > 0:
self.dropout = nn.Dropout2d(p=dropout)
else:
self.dropout = None
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, (nn.BatchNorm2d, nn.InstanceNorm2d, nn.GroupNorm)):
if m.weight is not None:
nn.init.constant_(m.weight, 1)
if m.bias is not None:
nn.init.constant_(m.bias, 0)
def _make_layer(self, dim, stride=1):
layer1 = ResidualBlock(self.in_planes, dim, self.norm_fn, stride=stride)
layer2 = ResidualBlock(dim, dim, self.norm_fn, stride=1)
layers = (layer1, layer2)
self.in_planes = dim
return nn.Sequential(*layers)
def forward(self, x):
b, n, c1, h1, w1 = x.shape
x = x.view(b*n, c1, h1, w1)
x = self.conv1(x)
x = self.norm1(x)
x = self.relu1(x)
x = self.layer1(x)
x = self.layer2(x)
x = self.layer3(x)
x = self.conv2(x)
_, c2, h2, w2 = x.shape
return x.view(b, n, c2, h2, w2)
================================================
FILE: dbaf/modules/gru.py
================================================
import torch
import torch.nn as nn
class ConvGRU(nn.Module):
def __init__(self, h_planes=128, i_planes=128):
super(ConvGRU, self).__init__()
self.do_checkpoint = False
self.convz = nn.Conv2d(h_planes+i_planes, h_planes, 3, padding=1)
self.convr = nn.Conv2d(h_planes+i_planes, h_planes, 3, padding=1)
self.convq = nn.Conv2d(h_planes+i_planes, h_planes, 3, padding=1)
self.w = nn.Conv2d(h_planes, h_planes, 1, padding=0)
self.convz_glo = nn.Conv2d(h_planes, h_planes, 1, padding=0)
self.convr_glo = nn.Conv2d(h_planes, h_planes, 1, padding=0)
self.convq_glo = nn.Conv2d(h_planes, h_planes, 1, padding=0)
def forward(self, net, *inputs):
inp = torch.cat(inputs, dim=1)
net_inp = torch.cat([net, inp], dim=1) #[1,128+128+128+64,H//8,W//8]
b, c, h, w = net.shape
glo = torch.sigmoid(self.w(net)) * net
glo = glo.view(b, c, h*w).mean(-1).view(b, c, 1, 1) # global context
z = torch.sigmoid(self.convz(net_inp) + self.convz_glo(glo))
r = torch.sigmoid(self.convr(net_inp) + self.convr_glo(glo))
q = torch.tanh(self.convq(torch.cat([r*net, inp], dim=1)) + self.convq_glo(glo))
net = (1-z) * net + z * q
return net
================================================
FILE: dbaf/motion_filter.py
================================================
import cv2
import torch
import lietorch
from collections import OrderedDict
from droid_net import DroidNet
import geom.projective_ops as pops
from modules.corr import CorrBlock
import numpy as np
class MotionFilter:
""" This class is used to filter incoming frames and extract features """
def __init__(self, net, video, thresh=2.5, device="cuda:0"):
# split net modules
self.cnet = net.cnet
self.fnet = net.fnet
self.update = net.update
self.video = video
self.thresh = thresh
self.device = device
self.count = 0
# mean, std for image normalization
self.MEAN = torch.as_tensor([0.485, 0.456, 0.406], device=self.device)[:, None, None]
self.STDV = torch.as_tensor([0.229, 0.224, 0.225], device=self.device)[:, None, None]
@torch.cuda.amp.autocast(enabled=True)
def __context_encoder(self, image):
""" context features """
net, inp = self.cnet(image).split([128,128], dim=2)
return net.tanh().squeeze(0), inp.relu().squeeze(0)
@torch.cuda.amp.autocast(enabled=True)
def context_encoder(self, image):
""" context features """
net, inp = self.cnet(image).split([128,128], dim=2)
return net.tanh().squeeze(0), inp.relu().squeeze(0)
@torch.cuda.amp.autocast(enabled=True)
def __feature_encoder(self, image):
""" features for correlation volume """
return self.fnet(image).squeeze(0)
@torch.cuda.amp.autocast(enabled=True)
def feature_encoder(self, image):
""" features for correlation volume """
return self.fnet(image).squeeze(0)
@torch.cuda.amp.autocast(enabled=True)
@torch.no_grad()
def track(self, tstamp, image, depth=None, intrinsics=None):
""" main update operation - run on every frame in video """
Id = lietorch.SE3.Identity(1,).data.squeeze()
ht = image.shape[-2] // 8
wd = image.shape[-1] // 8
# normalize images
inputs = image[None, :, [2,1,0]].to(self.device) / 255.0
inputs = inputs.sub_(self.MEAN).div_(self.STDV)
# extract features
gmap = self.__feature_encoder(inputs) #当前帧的特征, fnet
### always add first frame to the depth video ###
if self.video.counter.value == 0:
net, inp = self.__context_encoder(inputs[:,[0]])
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]
self.video.append(tstamp, image[0], Id, 1.0, depth, intrinsics / 8.0, gmap, net[0,0], inp[0,0])
### only add new frame if there is enough motion ###
else:
# index correlation volume
coords0 = pops.coords_grid(ht, wd, device=self.device)[None,None]
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)，
# approximate flow magnitude using 1 update iteration
_, delta, weight = self.update(self.net[None], self.inp[None], corr)
# check motion magnitue / add new frame to video
if delta.norm(dim=-1).mean().item() > self.thresh:
self.count = 0
net, inp = self.__context_encoder(inputs[:,[0]])
self.net, self.inp, self.fmap = net, inp, gmap
self.video.append(tstamp, image[0], None, None, depth, intrinsics / 8.0, gmap, net[0], inp[0])
else:
self.count += 1
================================================
FILE: dbaf/multi_sensor.py
================================================
import numpy as np
import gtsam
import math
GRAVITY = 9.807
class MultiSensorState:
def __init__(self):
self.cur_t = 0.0
""" IMU-centered states """
self.timestamps = [] # timestamps (len == N)
self.wTbs = [] # poses (len == N)
self.vs = [] # vels (len == N)
self.bs = [] # biases (len == N)
self.preintegrations = [] # preintegrations (len == N)
self.preintegrations_meas = [] # raw IMU data (len == N)
self.preintegration_temp = None # used for high-frequency prediction
self.pose_temp = None # used for high-frequency prediction
self.gnss_valid = [] # GNSS flags (len == N)
self.gnss_position = [] # GNSS pos (len == N)
self.odo_valid = [] # Odo flags (len == N)
self.odo_vel = [] # Odo vel (len == N)
self.marg_factor = None
self.set_imu_params()
def set_imu_params(self, noise = None):
# default
accel_noise_sigma = 0.0
gyro_noise_sigma = 0.0
accel_bias_rw_sigma = 0.0
gyro_bias_rw_sigma = 0.0
if noise != None:
accel_noise_sigma = noise[0]
gyro_noise_sigma = noise[1]
accel_bias_rw_sigma = noise[2]
gyro_bias_rw_sigma = noise[3]
measured_acc_cov = np.eye(3,3) * math.pow(accel_noise_sigma,2)
measured_omega_cov = np.eye(3,3) * math.pow(gyro_noise_sigma,2)
integration_error_cov = np.eye(3,3) * 0e-8
bias_acc_cov = np.eye(3,3) * math.pow(accel_bias_rw_sigma,2)
bias_omega_cov = np.eye(3,3) * math.pow(gyro_bias_rw_sigma,2)
bias_acc_omega_init = np.eye(6,6) * 0e-5
params = gtsam.PreintegrationCombinedParams.MakeSharedU(GRAVITY)
params.setAccelerometerCovariance(measured_acc_cov)
params.setIntegrationCovariance(integration_error_cov)
params.setGyroscopeCovariance(measured_omega_cov)
params.setBiasAccCovariance(bias_acc_cov)
params.setBiasOmegaCovariance(bias_omega_cov)
params.setBiasAccOmegaInit(bias_acc_omega_init)
self.params = params
params_loose = gtsam.PreintegrationCombinedParams.MakeSharedU(GRAVITY)
params_loose.setAccelerometerCovariance(measured_acc_cov* 100)
params_loose.setIntegrationCovariance(integration_error_cov)
params_loose.setGyroscopeCovariance(measured_omega_cov * 100)
params_loose.setBiasAccCovariance(bias_acc_cov)
params_loose.setBiasOmegaCovariance(bias_omega_cov)
params_loose.setBiasAccOmegaInit(bias_acc_omega_init)
self.params_loose = params_loose
def init_first_state(self,t,pos,R,vel):
self.timestamps.append(t)
self.wTbs.append(gtsam.Pose3(gtsam.Rot3(R), gtsam.Point3(pos)))
self.vs.append(vel)
self.bs.append(gtsam.imuBias.ConstantBias(np.array([.0,.0,.0]),np.array([.0,.0,.0])))
self.preintegrations.append(gtsam.PreintegratedCombinedMeasurements(self.params,self.bs[-1]))
self.preintegrations_meas.append([])
self.preintegration_temp = gtsam.PreintegratedCombinedMeasurements(self.params,self.bs[-1])
self.gnss_valid.append(False)
self.gnss_position.append(np.array([.0,.0,.0]))
self.odo_valid.append(False)
self.odo_vel.append(np.array([.0,.0,.0]))
self.cur_t = t
def append_imu(self, t, measuredAcc, measuredOmega):
if t - self.cur_t > 0:
if t-self.cur_t > 0.025: # IMU gap found, loose the IMU factor
new_preintegration = gtsam.PreintegratedCombinedMeasurements(self.params_loose,self.bs[-1])
for iii in range(len(self.preintegrations_meas[-1])):
dd = self.preintegrations_meas[-1][iii]
if dd[2] > 0:
new_preintegration.integrateMeasurement(dd[0],dd[1],dd[2])
self.preintegrations[-1] = new_preintegration
self.preintegrations[-1].integrateMeasurement(\
measuredAcc, measuredOmega, t - self.cur_t)
if t - self.cur_t < 0:
raise Exception("may not happen")
self.preintegrations_meas[-1].append([measuredAcc, measuredOmega, t - self.cur_t, t])
# print('append_imu: ',measuredAcc,measuredOmega,t - self.cur_t,t)
self.last_measuredAcc = measuredAcc
self.last_measuredOmega = measuredOmega
self.cur_t = t
def append_imu_temp(self, t, measuredAcc, measuredOmega, predict_pose = False):
if t - self.cur_t > 0:
self.preintegration_temp.integrateMeasurement(\
measuredAcc, measuredOmega, t - self.cur_t)
if predict_pose:
prev_state = gtsam.gtsam.NavState(self.wTbs[-1],self.vs[-1])
prev_bias = self.bs[-1]
self.pose_temp = self.preintegration_temp.predict(prev_state, prev_bias)
def append_img(self, t):
self.cur_t = t
prev_state = gtsam.gtsam.NavState(self.wTbs[-1],self.vs[-1])
prev_bias = self.bs[-1]
prop_state = self.preintegrations[-1].predict(prev_state, prev_bias)
if self.preintegrations[-1].deltaTij()>1.0:
prop_state = gtsam.gtsam.NavState(self.wTbs[-1],self.vs[-1])
self.timestamps.append(t)
self.wTbs.append(prop_state.pose())
self.vs.append(prop_state.velocity())
self.bs.append(prev_bias)
self.gnss_valid.append(False)
self.gnss_position.append(np.array([.0,.0,.0]))
self.odo_valid.append(False)
self.odo_vel.append(np.array([.0,.0,.0]))
self.preintegrations.append(\
gtsam.PreintegratedCombinedMeasurements(self.params,self.bs[-1]))
self.preintegrations_meas.append([])
self.preintegration_temp = gtsam.PreintegratedCombinedMeasurements(self.params,self.bs[-1])
# ugly implementation
# this should be called after append_img()
def append_gnss(self,t,pos):
if math.fabs(self.cur_t - t) > 0.01:
print('Skip GNSS data due to unsynchronization!!')
else:
self.gnss_valid[-1] = True
self.gnss_position[-1] = pos
def append_odo(self,t,vel):
if math.fabs(self.cur_t - t) > 0.01:
print('Skip ODO data due to unsynchronization!!')
else:
self.odo_valid[-1] = True
self.odo_vel[-1] = vel
def predict(self):
prev_state = gtsam.gtsam.NavState(self.wTbs[-1],self.vs[-1])
prev_bias = self.bs[-1]
self.preintegrations[-1].predict(prev_state,prev_bias)
================================================
FILE: demo_vio_kitti360.py
================================================
import sys
sys.path.append('dbaf')
sys.path.append('dbaf/geoFunc')
from tqdm import tqdm
import numpy as np
import torch
import cv2
import os
import argparse
from dbaf import DBAFusion
import h5py
import pickle
import re
import math
import gtsam
import quaternion
def show_image(image):
image = image.permute(1, 2, 0).cpu().numpy()
cv2.imshow('image', image / 255.0)
cv2.waitKey(1)
def image_stream(imagedir, imagestamp, enable_h5, h5path, calib, stride):
""" image generator """
calib = np.loadtxt(calib, delimiter=" ")
fx, fy, cx, cy = calib[:4]
K = np.eye(3)
K[0,0] = fx
K[0,2] = cx
K[1,1] = fy
K[1,2] = cy
if not enable_h5:
image_list = sorted(os.listdir(imagedir))[::stride]
image_stamps = np.loadtxt(imagestamp,str)
image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))
for t, imfile in enumerate(image_list):
image = cv2.imread(os.path.join(imagedir, imfile))
if len(calib) > 4:
image = cv2.undistort(image, K, calib[4:])
tt = float(image_dict[imfile])
h0, w0, _ = image.shape
h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))
w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))
image = cv2.resize(image, (w1, h1))
image = image[:h1-h1%8, :w1-w1%8]
image = torch.as_tensor(image).permute(2, 0, 1)
intrinsics = torch.as_tensor([fx, fy, cx, cy])
intrinsics[0::2] *= (w1 / w0)
intrinsics[1::2] *= (h1 / h0)
yield tt, image[None], intrinsics
else:
ccount = 0
h5_f = h5py.File(h5path,'r')
all_keys = sorted(list(h5_f.keys()))
for key in all_keys:
ccount += 1
yield pickle.loads(np.array(h5_f[key]))
if __name__ == '__main__':
print(torch.cuda.device_count())
print(torch.cuda.is_available())
print(torch.cuda.current_device())
parser = argparse.ArgumentParser()
parser.add_argument("--imagedir", type=str, help="path to image directory")
parser.add_argument("--imagestamp", type=str, help="")
parser.add_argument("--imupath", type=str, help="")
parser.add_argument("--gtpath", type=str, help="")
parser.add_argument("--enable_h5", action="store_true", help="")
parser.add_argument("--h5path", type=str, help="")
parser.add_argument("--resultpath", type=str, default="result.txt", help="")
parser.add_argument("--calib", type=str, help="path to calibration file")
parser.add_argument("--t0", default=0, type=int, help="starting frame")
parser.add_argument("--stride", default=3, type=int, help="frame stride")
parser.add_argument("--weights", default="droid.pth")
parser.add_argument("--buffer", type=int, default=80)
parser.add_argument("--image_size", default=[240, 320])
parser.add_argument("--max_factors", type=int, default=48, help="maximum active edges (which determines the GPU memory usage)")
parser.add_argument("--beta", type=float, default=0.3, help="weight for translation / rotation components of flow")
parser.add_argument("--filter_thresh", type=float, default=2.4, help="how much motion before considering new keyframe")
parser.add_argument("--warmup", type=int, default=8, help="number of warmup frames")
parser.add_argument("--keyframe_thresh", type=float, default=4.0, help="threshold to create a new keyframe")
parser.add_argument("--frontend_thresh", type=float, default=16.0, help="add edges between frames whithin this distance")
parser.add_argument("--frontend_window", type=int, default=25, help="frontend optimization window")
parser.add_argument("--active_window", type=int, default=8, help="maximum frames involved in DBA")
parser.add_argument("--inac_range", type=int, default=3, help="maximum inactive frames (whose flow wouldn't be updated) involved in DBA")
parser.add_argument("--frontend_radius", type=int, default=2, help="force edges between frames within radius")
parser.add_argument("--frontend_nms", type=int, default=1, help="non-maximal supression of edges")
parser.add_argument("--backend_thresh", type=float, default=22.0)
parser.add_argument("--backend_radius", type=int, default=2)
parser.add_argument("--backend_nms", type=int, default=3)
parser.add_argument("--upsample", action="store_true")
parser.add_argument("--visual_only", type=int,default=0, help="wheter to disbale the IMU")
parser.add_argument("--far_threshold", type=float, default=0.02, help="far pixels would be downweighted (unit: m^-1)")
parser.add_argument("--translation_threshold", type=float, default=0.2, help="avoid the insertion of too close keyframes (unit: m)")
parser.add_argument("--mask_threshold", type=float, default=-1, help="downweight too close edges (unit: m)")
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)")
parser.add_argument("--save_pkl", action="store_true")
parser.add_argument("--pklpath", default="result.pkl", help="path to saved reconstruction")
parser.add_argument("--show_plot", action="store_true", help="plot the image/trajectory during running")
args = parser.parse_args()
args.skip_edge = eval(args.skip_edge)
args.stereo = False
dbaf = None
torch.multiprocessing.set_start_method('spawn')
""" Load reference trajectory (for visualization) """
all_gt ={}
try:
fp = open(args.gtpath,'rt')
while True:
line = fp.readline().strip()
if line == '':break
if line[0] == '#' : continue
line = re.sub('\s\s+',' ',line)
elem = line.split(' ')
sod = float(elem[0])
if sod not in all_gt.keys():
all_gt[sod] ={}
R = quaternion.as_rotation_matrix(quaternion.from_float_array([float(elem[7]),\
float(elem[4]),\
float(elem[5]),\
float(elem[6])]))
TTT = np.eye(4,4)
TTT[0:3,0:3] = R
TTT[0:3,3] = np.array([ float(elem[1]), float(elem[2]), float(elem[3])])
all_gt[sod]['T'] = TTT
all_gt_keys =sorted(all_gt.keys())
fp.close()
except:
pass
""" Load IMU data """
all_imu = np.loadtxt(args.imupath)
all_odo = []
all_gnss = []
tstamps = []
""" Load images """
try:
for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp, args.enable_h5,\
args.h5path, args.calib, args.stride)):
if args.show_plot:
show_image(image[0])
if dbaf is None:
args.image_size = [image.shape[2], image.shape[3]]
dbaf = DBAFusion(args)
all_imu[:,0] -= 0.04 # IMU-camera time offset
dbaf.frontend.all_imu = all_imu
dbaf.frontend.all_gnss = all_gnss
dbaf.frontend.all_odo = all_odo
dbaf.frontend.all_stamp = np.loadtxt(args.imagestamp,str)
dbaf.frontend.all_stamp = dbaf.frontend.all_stamp[:,0].astype(np.float64)[None].transpose(1,0)
if len(all_gt) > 0:
dbaf.frontend.all_gt = all_gt
dbaf.frontend.all_gt_keys = all_gt_keys
# IMU-Camera Extrinsics
dbaf.video.Ti1c = np.array(
[0.99944133,-0.00228419,-0.03334389,-0.03734697,
0.03268308,-0.14183394,0.98935078,1.75837780,
-0.00698916,-0.98988784,-0.14168005,0.59911765,
0.00000000,0.00000000,0.00000000,1.00000000]).reshape([4,4])
dbaf.video.Tbc = gtsam.Pose3(dbaf.video.Ti1c)
# IMU parameters
dbaf.video.state.set_imu_params([ 0.0003924 * 25,0.000205689024915 * 25, 0.004905 * 10, 0.000001454441043 * 500])
dbaf.video.init_pose_sigma = np.array([1.0, 1.0, 0.0001, 1.0, 1.0, 1.0])
dbaf.video.init_bias_sigma = np.array([0.1, 0.1, 0.1, 0.1, 0.1, 0.1])
dbaf.frontend.translation_threshold = args.translation_threshold
dbaf.frontend.graph.mask_threshold = args.mask_threshold
dbaf.track(t, image, intrinsics=intrinsics)
dbaf.save_vis_easy()
except Exception as err:
print(err)
dbaf.save_vis_easy()
dbaf.terminate()
================================================
FILE: demo_vio_subt.py
================================================
import sys
sys.path.append('dbaf')
from tqdm import tqdm
import numpy as np
import torch
import cv2
import os
import argparse
from dbaf import DBAFusion
import h5py
import pickle
import re
import math
import quaternion
import gtsam
def show_image(image):
image = image.permute(1, 2, 0).cpu().numpy()
cv2.imshow('image', image / 255.0)
cv2.waitKey(1)
def image_stream(imagedir, imagestamp, enable_h5, h5path, calib, stride):
""" image generator """
calib = np.loadtxt(calib, delimiter=" ")
fx, fy, cx, cy = calib[:4]
K = np.eye(3)
K[0,0] = fx
K[0,2] = cx
K[1,1] = fy
K[1,2] = cy
Kn = np.eye(3)
Kn[0,0] = fx*0.35
Kn[0,2] = cx
Kn[1,1] = fy*0.35
Kn[1,2] = cy
D = calib[5:]
xi = np.array([calib[4]])
if not enable_h5:
image_list = [f for f in os.listdir(imagedir) if f.endswith('.png')]
image_list = sorted(image_list,key = lambda x: int(x.split('.')[0]))[2000::stride]
image_stamps = np.loadtxt(imagestamp,str,delimiter=',')[2000::stride]
for ii in range(len(image_list)):
image = cv2.imread(os.path.join(imagedir, image_list[ii]))
if len(calib) > 4:
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)
image = cv2.remap(image, m1, m2, interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_CONSTANT)
tt = float(image_stamps[ii]) /1e9
h0, w0, _ = image.shape
h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))
w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))
image = cv2.resize(image, (w1, h1))
image = image[:h1-h1%8, :w1-w1%8]
image = torch.as_tensor(image).permute(2, 0, 1)
intrinsics = torch.as_tensor([fx*0.35, fy*0.35, cx, cy ])
intrinsics[0::2] *= (w1 / w0)
intrinsics[1::2] *= (h1 / h0)
yield tt, image[None], intrinsics
else:
ccount = 0
h5_f = h5py.File(h5path,'r')
all_keys = sorted(list(h5_f.keys()))
for key in all_keys:
ccount += 1
yield pickle.loads(np.array(h5_f[key]))
if __name__ == '__main__':
print(torch.cuda.device_count())
print(torch.cuda.is_available())
print(torch.cuda.current_device())
parser = argparse.ArgumentParser()
parser.add_argument("--imagedir", type=str, help="path to image directory")
parser.add_argument("--imagestamp", type=str, help="")
parser.add_argument("--imupath", type=str, help="")
parser.add_argument("--gtpath", type=str, help="")
parser.add_argument("--enable_h5", action="store_true", help="")
parser.add_argument("--h5path", type=str, help="")
parser.add_argument("--resultpath", type=str, default="result.txt", help="")
parser.add_argument("--calib", type=str, help="path to calibration file")
parser.add_argument("--t0", default=0, type=int, help="starting frame")
parser.add_argument("--stride", default=3, type=int, help="frame stride")
parser.add_argument("--weights", default="droid.pth")
parser.add_argument("--buffer", type=int, default=80)
parser.add_argument("--image_size", default=[240, 320])
parser.add_argument("--max_factors", type=int, default=48, help="maximum active edges (which determines the GPU memory usage)")
parser.add_argument("--beta", type=float, default=0.3, help="weight for translation / rotation components of flow")
parser.add_argument("--filter_thresh", type=float, default=2.4, help="how much motion before considering new keyframe")
parser.add_argument("--warmup", type=int, default=8, help="number of warmup frames")
parser.add_argument("--keyframe_thresh", type=float, default=4.0, help="threshold to create a new keyframe")
parser.add_argument("--frontend_thresh", type=float, default=16.0, help="add edges between frames whithin this distance")
parser.add_argument("--frontend_window", type=int, default=25, help="frontend optimization window")
parser.add_argument("--active_window", type=int, default=8, help="maximum frames involved in DBA")
parser.add_argument("--inac_range", type=int, default=3, help="maximum inactive frames (whose flow wouldn't be updated) involved in DBA")
parser.add_argument("--frontend_radius", type=int, default=2, help="force edges between frames within radius")
parser.add_argument("--frontend_nms", type=int, default=1, help="non-maximal supression of edges")
parser.add_argument("--backend_thresh", type=float, default=22.0)
parser.add_argument("--backend_radius", type=int, default=2)
parser.add_argument("--backend_nms", type=int, default=3)
parser.add_argument("--upsample", action="store_true")
parser.add_argument("--visual_only", type=int,default=0, help="wheter to disbale the IMU")
parser.add_argument("--far_threshold", type=float, default=0.02, help="far pixels would be downweighted (unit: m^-1)")
parser.add_argument("--translation_threshold", type=float, default=0.2, help="avoid the insertion of too close keyframes (unit: m)")
parser.add_argument("--mask_threshold", type=float, default=-1, help="downweight too close edges (unit: m)")
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)")
parser.add_argument("--save_pkl", action="store_true")
parser.add_argument("--pklpath", default="result.pkl", help="path to saved reconstruction")
parser.add_argument("--show_plot", action="store_true", help="plot the trajectory during running")
args = parser.parse_args()
args.skip_edge = eval(args.skip_edge)
args.stereo = False
dbaf = None
torch.multiprocessing.set_start_method('spawn')
""" Load reference trajectory (for visualization) """
all_gt ={}
try:
fp = open(args.gtpath,'rt')
while True:
line = fp.readline().strip()
if line == '':break
if line[0] == '#' : continue
line = re.sub('\s\s+',' ',line)
elem = line.split(',')
sod = float(elem[0])/1e9
if sod not in all_gt.keys():
all_gt[sod] ={}
R = quaternion.as_rotation_matrix(quaternion.from_float_array([float(elem[4]),\
float(elem[5]),\
float(elem[6]),\
float(elem[7])]))
TTT = np.eye(4,4)
TTT[0:3,0:3] = R
TTT[0:3,3] = np.array([ float(elem[1]), float(elem[2]), float(elem[3])])
all_gt[sod]['T'] = TTT
all_gt_keys =sorted(all_gt.keys())
fp.close()
except:
pass
""" Load IMU data """
all_imu = np.loadtxt(args.imupath,delimiter=',',comments='#',skiprows=1)
all_imu[:,0] /= 1e9
# all_imu[:,1:4] *= 180/math.pi
all_imu_new = np.zeros_like(all_imu)
all_imu_new[:,0] = all_imu[:,0]
all_imu_new[:,1:4] = all_imu[:,5:8] * 180/math.pi
all_imu_new[:,4:7] = all_imu[:,8:11]
all_imu_new = all_imu_new[:,:7]
tstamps = []
""" Load images """
for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp, args.enable_h5,\
args.h5path, args.calib, args.stride)):
if args.show_plot:
show_image(image[0])
if dbaf is None:
args.image_size = [image.shape[2], image.shape[3]]
dbaf = DBAFusion(args)
dbaf.frontend.all_imu = all_imu_new
dbaf.frontend.all_gnss = []
dbaf.frontend.all_odo = []
dbaf.frontend.all_stamp = np.loadtxt(args.imagestamp,str,delimiter=',')
dbaf.frontend.all_stamp = dbaf.frontend.all_stamp.astype(np.float64)[2000::args.stride][None].transpose(1,0)/1e9
if len(all_gt) > 0:
dbaf.frontend.all_gt = all_gt
dbaf.frontend.all_gt_keys = all_gt_keys
# IMU-Camera Extrinsics
dbaf.video.Ti1c = np.array(
[-0.04279531, -0.00237969, 0.99908103, 0.19499356,
-0.99880330, -0.02359508, -0.04283961, 0.04340662,
0.02367534, -0.99971877, -0.00136708, -0.01782382,
0.00000000, 0.00000000, 0.00000000, 1.00000000]).reshape([4,4])
dbaf.video.Tbc = gtsam.Pose3(dbaf.video.Ti1c)
# IMU parameters
dbaf.video.state.set_imu_params((np.array([ 0.0003924 * 25,0.000205689024915 * 25, 0.004905 * 10, 0.000001454441043 * 500])*1.0).tolist())
dbaf.video.init_pose_sigma = np.array([1, 1, 0.0001, 0.0001,0.0001,0.0001])
dbaf.video.init_bias_sigma = np.array([1.0,1.0,1.0,1.0,1.0,1.0])
dbaf.frontend.translation_threshold = args.translation_threshold
dbaf.frontend.graph.mask_threshold = args.mask_threshold
dbaf.track(t, image, intrinsics=intrinsics)
if args.save_pkl:
dbaf.save_vis_easy()
dbaf.terminate()
================================================
FILE: demo_vio_tumvi.py
================================================
import sys
sys.path.append('dbaf')
from tqdm import tqdm
import numpy as np
import torch
import cv2
import os
import argparse
from dbaf import DBAFusion
import h5py
import pickle
import re
import math
import quaternion
import gtsam
def show_image(image):
image = image.permute(1, 2, 0).cpu().numpy()
cv2.imshow('image', image / 255.0)
cv2.waitKey(1)
def image_stream(imagedir, imagestamp, enable_h5, h5path, calib, stride):
""" image generator """
calib = np.loadtxt(calib, delimiter=" ")
fx, fy, cx, cy = calib[:4]
K = np.eye(3)
K[0,0] = fx
K[0,2] = cx
K[1,1] = fy
K[1,2] = cy
Kn = np.eye(3)
Kn[0,0] = fx
Kn[0,2] = cx
Kn[1,1] = fy
Kn[1,2] = cy
if not enable_h5:
image_list = sorted(os.listdir(imagedir))[::stride]
image_stamps = np.loadtxt(imagestamp,str,delimiter=',')
image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))
for t, imfile in enumerate(image_list):
image = cv2.imread(os.path.join(imagedir, imfile))
if len(calib) > 4:
m1, m2 = cv2.fisheye.initUndistortRectifyMap(K,calib[4:],np.eye(3),Kn,(512,512),cv2.CV_32FC1)
image = cv2.remap(image, m1, m2, interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_CONSTANT)
tt = float(image_dict[imfile]) /1e9
h0, w0, _ = image.shape
h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))
w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))
image = cv2.resize(image, (w1, h1))
image = image[:h1-h1%8, :w1-w1%8]
image = torch.as_tensor(image).permute(2, 0, 1)
intrinsics = torch.as_tensor([fx, fy, cx, cy ])
intrinsics[0::2] *= (w1 / w0)
intrinsics[1::2] *= (h1 / h0)
yield tt, image[None], intrinsics
else:
ccount = 0
h5_f = h5py.File(h5path,'r')
all_keys = sorted(list(h5_f.keys()))
for key in all_keys:
ccount += 1
yield pickle.loads(np.array(h5_f[key]))
if __name__ == '__main__':
print(torch.cuda.device_count())
print(torch.cuda.is_available())
print(torch.cuda.current_device())
parser = argparse.ArgumentParser()
parser.add_argument("--imagedir", type=str, help="path to image directory")
parser.add_argument("--imagestamp", type=str, help="")
parser.add_argument("--imupath", type=str, help="")
parser.add_argument("--gtpath", type=str, help="")
parser.add_argument("--enable_h5", action="store_true", help="")
parser.add_argument("--h5path", type=str, help="")
parser.add_argument("--resultpath", type=str, default="result.txt", help="")
parser.add_argument("--calib", type=str, help="path to calibration file")
parser.add_argument("--t0", default=0, type=int, help="starting frame")
parser.add_argument("--stride", default=3, type=int, help="frame stride")
parser.add_argument("--weights", default="droid.pth")
parser.add_argument("--buffer", type=int, default=80)
parser.add_argument("--image_size", default=[240, 320])
parser.add_argument("--max_factors", type=int, default=48, help="maximum active edges (which determines the GPU memory usage)")
parser.add_argument("--beta", type=float, default=0.3, help="weight for translation / rotation components of flow")
parser.add_argument("--filter_thresh", type=float, default=2.4, help="how much motion before considering new keyframe")
parser.add_argument("--warmup", type=int, default=8, help="number of warmup frames")
parser.add_argument("--keyframe_thresh", type=float, default=3.0, help="threshold to create a new keyframe")
parser.add_argument("--frontend_thresh", type=float, default=16.0, help="add edges between frames whithin this distance")
parser.add_argument("--frontend_window", type=int, default=25, help="frontend optimization window")
parser.add_argument("--active_window", type=int, default=8, help="maximum frames involved in DBA")
parser.add_argument("--inac_range", type=int, default=3, help="maximum inactive frames (whose flow wouldn't be updated) involved in DBA")
parser.add_argument("--frontend_radius", type=int, default=2, help="force edges between frames within radius")
parser.add_argument("--frontend_nms", type=int, default=1, help="non-maximal supression of edges")
parser.add_argument("--backend_thresh", type=float, default=22.0)
parser.add_argument("--backend_radius", type=int, default=2)
parser.add_argument("--backend_nms", type=int, default=3)
parser.add_argument("--upsample", action="store_true")
parser.add_argument("--visual_only", type=int,default=0, help="wheter to disbale the IMU")
parser.add_argument("--far_threshold", type=float, default=0.02, help="far pixels would be downweighted (unit: m^-1)")
parser.add_argument("--translation_threshold", type=float, default=0.2, help="avoid the insertion of too close keyframes (unit: m)")
parser.add_argument("--mask_threshold", type=float, default=-1, help="downweight too close edges (unit: m)")
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)")
parser.add_argument("--save_pkl", action="store_true")
parser.add_argument("--pklpath", default="result.pkl", help="path to saved reconstruction")
parser.add_argument("--show_plot", action="store_true", help="plot the trajectory during running")
args = parser.parse_args()
args.skip_edge = eval(args.skip_edge)
args.stereo = False
dbaf = None
torch.multiprocessing.set_start_method('spawn')
""" Load reference trajectory (for visualization) """
all_gt ={}
try:
fp = open(args.gtpath,'rt')
while True:
line = fp.readline().strip()
if line == '':break
if line[0] == '#' : continue
line = re.sub('\s\s+',' ',line)
elem = line.split(',')
sod = float(elem[0])/1e9
if sod not in all_gt.keys():
all_gt[sod] ={}
R = quaternion.as_rotation_matrix(quaternion.from_float_array([float(elem[4]),\
float(elem[5]),\
float(elem[6]),\
float(elem[7])]))
TTT = np.eye(4,4)
TTT[0:3,0:3] = R
TTT[0:3,3] = np.array([ float(elem[1]), float(elem[2]), float(elem[3])])
all_gt[sod]['T'] = TTT
all_gt_keys =sorted(all_gt.keys())
fp.close()
except:
pass
""" Load IMU data """
all_imu = np.loadtxt(args.imupath,delimiter=',')
all_imu[:,0] /= 1e9
all_imu[:,1:4] *= 180/math.pi
tstamps = []
""" Load images """
clahe = cv2.createCLAHE(2.0,tileGridSize=(8, 8))
for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp, args.enable_h5,\
args.h5path, args.calib, args.stride)):
mm = clahe.apply(image[0][0].numpy())
image[0] = torch.tensor(mm[None].repeat(3,0))
if args.show_plot:
show_image(image[0])
if dbaf is None:
args.image_size = [image.shape[2], image.shape[3]]
dbaf = DBAFusion(args)
dbaf.frontend.all_imu = all_imu
dbaf.frontend.all_gnss = []
dbaf.frontend.all_odo = []
dbaf.frontend.all_stamp = np.loadtxt(args.imagestamp,str,delimiter=',')
dbaf.frontend.all_stamp = dbaf.frontend.all_stamp[:,0].astype(np.float64)[None].transpose(1,0)/1e9
if len(all_gt) > 0:
dbaf.frontend.all_gt = all_gt
dbaf.frontend.all_gt_keys = all_gt_keys
# IMU-Camera Extrinsics
dbaf.video.Ti1c = np.array(
[-0.9995250378696743, 0.029615343885863205, -0.008522328211654736, 0.04727988224914392,
0.0075019185074052044, -0.03439736061393144, -0.9993800792498829, -0.047443232143367084,
-0.02989013031643309, -0.998969345370175, 0.03415885127385616, -0.0681999605066297,
0.0, 0.0, 0.0, 1.0]).reshape([4,4])
dbaf.video.Ti1c = np.linalg.inv(dbaf.video.Ti1c)
dbaf.video.Tbc = gtsam.Pose3(dbaf.video.Ti1c)
# IMU parameters
dbaf.video.state.set_imu_params((np.array([ 0.0003924 * 25,0.000205689024915 * 25, 0.004905 * 10, 0.000001454441043 * 5000])*1.0).tolist())
dbaf.video.init_pose_sigma = np.array([0.1, 0.1, 0.0001, 0.0001,0.0001,0.0001])
dbaf.video.init_bias_sigma = np.array([1.0,1.0,1.0, 1.0,1.0,1.0])
dbaf.frontend.translation_threshold = args.translation_threshold
dbaf.frontend.graph.mask_threshold = args.mask_threshold
dbaf.track(t, image, intrinsics=intrinsics)
if args.save_pkl:
dbaf.save_vis_easy()
dbaf.terminate()
================================================
FILE: demo_vio_whu.py
================================================
import sys
sys.path.append('dbaf')
sys.path.append('dbaf/geoFunc')
from tqdm import tqdm
import numpy as np
import torch
import cv2
import os
import argparse
from dbaf import DBAFusion
import h5py
import pickle
import re
import math
import gtsam
import geoFunc.trans as trans
def show_image(image):
image = image.permute(1, 2, 0).cpu().numpy()
cv2.imshow('image', image / 255.0)
cv2.waitKey(1)
def image_stream(imagedir, imagestamp, enable_h5, h5path, calib, stride):
""" image generator """
calib = np.loadtxt(calib, delimiter=" ")
fx, fy, cx, cy = calib[:4]
K = np.eye(3)
K[0,0] = fx
K[0,2] = cx
K[1,1] = fy
K[1,2] = cy
if not enable_h5:
image_stamps = np.loadtxt(imagestamp,str,delimiter=',')
image_dict = dict(zip(image_stamps[:,1],image_stamps[:,0]))
image_list = list(image_dict)
ccount = 0
for t, imfile in enumerate(image_list):
tt = float(image_dict[imfile])
if int(tt*10)%2 == 1: continue
ccount += 1
image = cv2.imread(os.path.join(imagedir, imfile))
if len(calib) > 4:
image = cv2.undistort(image, K, calib[4:])
h0, w0, _ = image.shape
h1 = int(h0 * np.sqrt((384 * 512) / (h0 * w0)))
w1 = int(w0 * np.sqrt((384 * 512) / (h0 * w0)))
image = cv2.resize(image, (w1, h1))
image = image[:h1-h1%8, :w1-w1%8]
image = torch.as_tensor(image).permute(2, 0, 1)
intrinsics = torch.as_tensor([fx, fy, cx, cy])
intrinsics[0::2] *= (w1 / w0)
intrinsics[1::2] *= (h1 / h0)
yield tt, image[None], intrinsics
else:
raise Exception()
if __name__ == '__main__':
print(torch.cuda.device_count())
print(torch.cuda.is_available())
print(torch.cuda.current_device())
parser = argparse.ArgumentParser()
parser.add_argument("--imagedir", type=str, help="path to image directory")
parser.add_argument("--imagestamp", type=str, help="")
parser.add_argument("--imupath", type=str, help="")
parser.add_argument("--gtpath", type=str, help="")
parser.add_argument("--enable_h5", action="store_true", help="")
parser.add_argument("--h5path", type=str, help="")
parser.add_argument("--resultpath", type=str, help="")
parser.add_argument("--calib", type=str, help="path to calibration file")
parser.add_argument("--t0", default=0, type=int, help="starting frame")
parser.add_argument("--stride", default=3, type=int, help="frame stride")
parser.add_argument("--weights", default="droid.pth")
parser.add_argument("--buffer", type=int, default=80)
parser.add_argument("--image_size", default=[240, 320])
parser.add_argument("--max_factors", type=int, default=48, help="maximum active edges (which determines the GPU memory usage)")
parser.add_argument("--beta", type=float, default=0.3, help="weight for translation / rotation components of flow")
parser.add_argument("--filter_thresh", type=float, default=0.00, help="how much motion before considering new keyframe")
parser.add_argument("--warmup", type=int, default=8, help="number of warmup frames")
parser.add_argument("--vi_warmup", type=int, default=15, help="")
parser.add_argument("--keyframe_thresh", type=float, default=3.0, help="threshold to create a new keyframe")
parser.add_argument("--frontend_thresh", type=float, default=16.0, help="add edges between frames whithin this distance")
parser.add_argument("--frontend_window", type=int, default=25, help="frontend optimization window")
parser.add_argument("--active_window", type=int, default=8)
parser.add_argument("--inac_range", type=int, default=3)
parser.add_argument("--frontend_radius", type=int, default=2, help="force edges between frames within radius")
parser.add_argument("--frontend_nms", type=int, default=1, help="non-maximal supression of edges")
parser.add_argument("--backend_thresh", type=float, default=22.0)
parser.add_argument("--backend_radius", type=int, default=2)
parser.add_argument("--backend_nms", type=int, default=3)
parser.add_argument("--upsample", action="store_true")
parser.add_argument("--visual_only", type=int,default=0, help="wheter to disbale the IMU")
parser.add_argument("--far_threshold", type=float, default=0.02, help="far pixels would be downweighted (unit: m^-1)")
parser.add_argument("--translation_threshold", type=float, default=0.2, help="avoid the insertion of too close keyframes (unit: m)")
parser.add_argument("--mask_threshold", type=float, default=-1, help="downweight too close edges (unit: m)")
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)")
parser.add_argument("--save_pkl", action="store_true")
parser.add_argument("--pklpath", default="result.pkl", help="path to saved reconstruction")
parser.add_argument("--show_plot", action="store_true", help="plot the image/trajectory during running")
parser.add_argument("--use_gnss", action="store_true")
parser.add_argument("--gnsspath", type=str, help="")
parser.add_argument("--use_odo", action="store_true")
parser.add_argument("--odopath", type=str, help="")
parser.add_argument("--use_zupt", action="store_true")
args = parser.parse_args()
args.skip_edge = eval(args.skip_edge)
args.stereo = False
dbaf = None
all_gt ={}
Ti0i1 =np.array([[ 9.99902524e-01, 1.39619889e-02, -7.31054713e-05, 1.00000000e-02],
[-1.39621803e-02, 9.99888818e-01, -5.23545345e-03, -2.05000000e-01],
[ 0.00000000e+00, 5.23596383e-03, 9.99986292e-01, -5.00000000e-02],
[ 0.00000000e+00, 0.00000000e+00, 0.00000000e+00, 1.00000000e+00]])
Ten0 = None
is_ref_set = False
fp = open(args.gtpath,'rt')
while True:
line = fp.readline().strip()
if line == '':break
if line[0] == '#' :continue
line = re.sub('\s\s+',' ',line)
elem = line.split(' ')
sod = float(elem[1])
if sod not in all_gt.keys():
all_gt[sod] ={}
all_gt[sod]['X0'] = float(elem[2])
all_gt[sod]['Y0'] = float(elem[3])
all_gt[sod]['Z0'] = float(elem[4])
all_gt[sod]['VX0'] = float(elem[15])
all_gt[sod]['VY0'] = float(elem[16])
all_gt[sod]['VZ0'] = float(elem[17])
all_gt[sod]['ATTX0']= float(elem[25])
all_gt[sod]['ATTY0']= float(elem[26])
all_gt[sod]['ATTZ0']= -float(elem[24])
Ren = trans.Cen([all_gt[sod]['X0'],all_gt[sod]['Y0'],all_gt[sod]['Z0']])
ani0 = [all_gt[sod]['ATTX0']/180*math.pi,\
all_gt[sod]['ATTY0']/180*math.pi,\
all_gt[sod]['ATTZ0']/180*math.pi]
Rni0 = trans.att2m(ani0)
Rei0 = np.matmul(Ren,Rni0)
tei0 = np.array([all_gt[sod]['X0'],all_gt[sod]['Y0'],all_gt[sod]['Z0']])
Tei0 = np.eye(4,4)
Tei0[0:3,0:3] = Rei0
Tei0[0:3,3] = tei0
if not is_ref_set:
is_ref_set = True
Ten0 = np.eye(4,4)
Ten0[0:3,0:3] = trans.Cen(tei0)
Ten0[0:3,3] = tei0
Tn0i0 = np.matmul(np.linalg.inv(Ten0),Tei0)
Tn0i1 = np.matmul(Tn0i0,Ti0i1)
all_gt[sod]['T'] = Tn0i1
all_gt_keys =sorted(all_gt.keys())
fp.close()
all_imu = np.loadtxt(args.imupath,delimiter=' ')
if args.use_gnss and os.path.isfile(args.gnsspath):
fix_map = {b'Fixed':1.0,b'Float':0.0}
all_gnss = np.genfromtxt(args.gnsspath,converters={16: lambda x: fix_map[x]})
else:
all_gnss = []
if args.use_odo and os.path.isfile(args.odopath):
all_odo = np.genfromtxt(args.odopath)
all_odo = all_odo[np.fabs(all_odo[:,0] - np.round(all_odo[:,0]))<0.001]
np.random.seed(12345)
all_odo[:,1:] += np.random.randn(all_odo.shape[0],3)*0.05
else:
all_odo = []
tstamps = []
# try:
for (t, image, intrinsics) in tqdm(image_stream(args.imagedir, args.imagestamp, args.enable_h5,\
args.h5path, args.calib, args.stride)):
if args.show_plot:
show_image(image[0])
if dbaf is None:
args.image_size = [image.shape[2], image.shape[3]]
dbaf = DBAFusion(args)
dbaf.frontend.all_imu = all_imu
dbaf.frontend.all_stamp = np.loadtxt(args.imagestamp,str,delimiter=',')
dbaf.frontend.all_gnss = all_gnss
dbaf.frontend.all_odo = all_odo
if len(all_gt) > 0:
dbaf.frontend.all_gt = all_gt
dbaf.frontend.all_gt_keys = all_gt_keys
dbaf.video.Ti1c = np.array(
[0.99988370,-0.00563944,-0.01418468,-0.15590000,
0.01424932,0.01159187,0.99983149,0.63466000,
-0.00547407,-0.99991712,0.01167088,0.04605000,
0.00000000,0.00000000,0.00000000,1.00000000]).reshape([4,4])
dbaf.video.tbg = np.array([-0.0125, -0.26, 0.2091])
dbaf.video.Tbc = gtsam.Pose3(dbaf.video.Ti1c)
dbaf.video.state.set_imu_params([ 0.0003924 * 25,0.000205689024915 * 25, 0.004905 * 10, 0.000001454441043 * 25])
if args.use_gnss:
dbaf.video.init_pose_sigma = np.array([1.0, 1.0, 10.0,10.0,10.0,10.0])
else:
dbaf.video.init_pose_sigma = np.array([[0.1, 0.1, 0.0001, 0.0001,0.0001,0.0001],
[1.0, 1.0, 0.0001, 10.0, 10.0, 10.0]])
dbaf.video.init_bias_sigma = np.array([1.0,1.0,1.0, 0.1, 0.1, 0.1])
dbaf.frontend.translation_threshold = args.translation_threshold
dbaf.frontend.graph.mask_threshold = args.mask_threshold
dbaf.track(t, image, intrinsics=intrinsics)
dbaf.save_vis_easy()
dbaf.terminate()
================================================
FILE: evaluation_scripts/batch_tumvi.py
================================================
import subprocess
for seq in ['magistrale1','magistrale2','magistrale3','magistrale4','magistrale5','magistrale6',\
'outdoors1','outdoors2','outdoors3','outdoors4','outdoors5','outdoors6','outdoors7','outdoors8']:
p = subprocess.Popen('python ./evaluation_scripts/evaluate_tumvi.py --batch --seq=%s | grep rmse' % seq, shell = True)
p.wait()
================================================
FILE: evaluation_scripts/evaluate_kitti.py
================================================
import argparse
import logging
import typing
import numpy as np
import evo.common_ape_rpe as common
from evo.core import lie_algebra, sync, metrics
from evo.core.result import Result
from evo.core.trajectory import PosePath3D, PoseTrajectory3D
from evo.tools import file_interface, log
from evo.tools.settings import SETTINGS
import matplotlib.pyplot as plt
import copy
from scipy.spatial.transform import Rotation
import bisect
import math
import time
logger = logging.getLogger(__name__)
SEP = "-" * 80 # separator line
def ape(traj_ref: PosePath3D, traj_est: PosePath3D,
pose_relation: metrics.PoseRelation, align: bool = False,
correct_scale: bool = False, n_to_align: int = -1,
align_origin: bool = False, ref_name: str = "reference",
est_name: str = "estimate",
change_unit: typing.Optional[metrics.Unit] = None) -> Result:
if n_to_align >0 :
print('>>>>> only use the starting segment')
n_to_align = np.where((np.array(traj_ref.timestamps)[1:]-np.array(traj_ref.timestamps)[0:-1])>100)[0][0]-1
# Align the trajectories.
only_scale = correct_scale and not align
alignment_transformation = None
if align or correct_scale:
logger.debug(SEP)
alignment_transformation = lie_algebra.sim3(
*traj_est.align(traj_ref, correct_scale, only_scale, n=n_to_align))
elif align_origin:
logger.debug(SEP)
alignment_transformation = traj_est.align_origin(traj_ref)
# Calculate APE.
logger.debug(SEP)
data = (traj_ref, traj_est)
ape_metric = metrics.APE(pose_relation)
ape_metric.process_data(data)
if change_unit:
ape_metric.change_unit(change_unit)
title = str(ape_metric)
if align and not correct_scale:
title += "\n(with SE(3) Umeyama alignment)"
elif align and correct_scale:
title += "\n(with Sim(3) Umeyama alignment)"
elif only_scale:
title += "\n(scale corrected)"
elif align_origin:
title += "\n(with origin alignment)"
else:
title += "\n(not aligned)"
if (align or correct_scale) and n_to_align != -1:
title += " (aligned poses: {})".format(n_to_align)
ape_result = ape_metric.get_result(ref_name, est_name)
ape_result.info["title"] = title
logger.debug(SEP)
logger.info(ape_result.pretty_str())
ape_result.add_trajectory(ref_name, traj_ref)
ape_result.add_trajectory(est_name, traj_est)
if isinstance(traj_est, PoseTrajectory3D):
seconds_from_start = np.array(
[t - traj_est.timestamps[0] for t in traj_est.timestamps])
ape_result.add_np_array("seconds_from_start", seconds_from_start)
ape_result.add_np_array("timestamps", traj_est.timestamps)
ape_result.add_np_array("distances_from_start", traj_ref.distances)
ape_result.add_np_array("distances", traj_est.distances)
if alignment_transformation is not None:
ape_result.add_np_array("alignment_transformation_sim3",
alignment_transformation)
return ape_result
if __name__ == '__main__':
color_list = [[0,0,1],[1,0.6,1],[1,0,0]]
plt.figure('1',figsize=[6,6])
parser = argparse.ArgumentParser()
parser.add_argument('--seq', type=str, help='seq',default='0010')
args = parser.parse_args()
args.subcommand = 'tum'
seq = args.seq
args.ref_file = '/home/zhouyuxuan/data/2013_05_28_drive_%s_sync/gt_local.txt' % seq
args.pose_relation = 'trans_part'
args.align = True
args.correct_scale = False
args.n_to_align = 1
args.align_origin = False
args.plot_mode = 'xyz'
args.plot_x_dimension = 'seconds'
args.plot_colormap_min = None
args.plot_colormap_max = None
args.plot_colormap_max_percentile = None
args.ros_map_yaml = None
args.plot = True
args.est_files = ['results/result_%s.txt' %seq]
label_list = ['DBA-Fusion (M)']
args.save_plot = False
args.serialize_plot = False
for iii in range(len(args.est_files)):
args.est_file = args.est_files[iii]
if args.est_file.find('visual') != -1:
args.correct_scale = True
else:
args.correct_scale = False
traj_ref, traj_est, ref_name, est_name = common.load_trajectories(args)
traj_ref_sel, traj_est_sel = sync.associate_trajectories(
traj_ref, traj_est, 0.01,0.0,
first_name=ref_name, snd_name=est_name)
args.n_to_align = -1
pose_relation = common.get_pose_relation(args)
result = ape(traj_ref=traj_ref_sel, traj_est=traj_est_sel,
pose_relation=pose_relation, align=args.align,
correct_scale=args.correct_scale, n_to_align=args.n_to_align,
align_origin=args.align_origin, ref_name=ref_name,
est_name=est_name)
traj_est_sel = copy.deepcopy(result.trajectories[est_name])
T01 = result.np_arrays['alignment_transformation_sim3']
print(T01)
result = ape(traj_ref=traj_ref_sel, traj_est=traj_est_sel,
pose_relation=pose_relation, align=args.align,
correct_scale=False, n_to_align=-1,
align_origin=args.align_origin, ref_name=ref_name,
est_name=est_name)
print(result)
traj_est.transform(T01)
if iii == 0:
x_series=[]
y_series=[]
z_series=[]
for i in range(len(traj_ref.poses_se3)):
TTT = traj_ref.poses_se3[i]
x_series.append(TTT[0,3])
y_series.append(TTT[1,3])
z_series.append(TTT[2,3])
plt.plot(x_series,y_series,c=[0,0,0],linestyle = '--')
x_series=[]
y_series=[]
z_series=[]
for i in range(len(traj_est.poses_se3)):
TTT = traj_est.poses_se3[i]
x_series.append(TTT[0,3])
y_series.append(TTT[1,3])
z_series.append(TTT[2,3])
ppp = TTT[0:3,3]
qqq = Rotation.from_matrix(TTT[:3, :3]/np.power(np.linalg.det(TTT[:3, :3]),1.0/3)).as_quat()
plt.plot(x_series,y_series,c=color_list[iii],label = label_list[iii])
# t_series=[]
# x_series=[]
# y_series=[]
# z_series=[]
# for i in range(len(traj_ref_sel.timestamps)):
# T0 = traj_ref_sel.poses_se3[i]
# T1 = traj_est_sel.poses_se3[i]
# T01 = np.matmul(np.linalg.inv(T0),T1)
# att = Rotation.from_matrix(T01[0:3,0:3]).as_rotvec()
# t_series.append(traj_ref_sel.timestamps[i])
# x_series.append(att[0])
# y_series.append(att[1])
# z_series.append(att[2])
# plt.figure()
# plt.plot(t_series,x_series)
# plt.plot(t_series,y_series)
# plt.plot(t_series,z_series)
# plt.show()
print('Evaluating relative pose error ...')
subtraj_length = [100,200,300,400,500,600,700,800]
max_dist_difH=1
rel_trans_error_dist = []
rel_att_error_dist = []
for i in range(8):
subsection_index=[]
max_dist_diff=0.2*subtraj_length[i]
traj_len = len(traj_ref_sel.timestamps)
for j in range(traj_len-2):
k = bisect.bisect(traj_ref_sel.distances,traj_ref_sel.distances[j]+subtraj_length[i]-max_dist_difH)
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:
subsection_index.append([j,k])
print("The trajectory at %dm have %d matching points... " %(subtraj_length[i],len(subsection_index)))
rel_tran_errors = []
rel_att_errors = []
for ii in subsection_index:
T_gt_1 =traj_ref_sel.poses_se3[ii[0]]
T_gt_2 =traj_ref_sel.poses_se3[ii[1]]
T_est_1 =traj_est_sel.poses_se3[ii[0]]
T_est_2 =traj_est_sel.poses_se3[ii[1]]
T_gt_12=np.matmul(np.linalg.inv(T_gt_1),T_gt_2)
T_est_12=np.matmul(np.linalg.inv(T_est_1),T_est_2)
T_error=np.matmul(np.linalg.inv(T_gt_12),T_est_12)
rel_tran_error = np.linalg.norm(T_error[0:3,3])
rel_att_error = np.linalg.norm(Rotation.from_matrix(T_error[0:3,0:3]).as_rotvec())
rel_tran_errors.append(rel_tran_error/subtraj_length[i]*100)
rel_att_errors.append(rel_att_error/subtraj_length[i]*100/math.pi*180)
rel_trans_error_dist.append(np.mean(np.array(rel_tran_errors)))
rel_att_error_dist.append(np.mean(np.array(rel_att_errors)))
print('Relative Translation Error: %f%%' % np.mean(np.array(rel_trans_error_dist)))
print('Relative Rotation Error: %f deg / 100 m' % np.mean(np.array(rel_att_error_dist)))
plt.show()
================================================
FILE: evaluation_scripts/evaluate_tumvi.py
================================================
import argparse
import logging
import typing
import numpy as np
import evo.common_ape_rpe as common
from evo.core import lie_algebra, sync, metrics
from evo.core.result import Result
from evo.core.trajectory import PosePath3D, PoseTrajectory3D
from evo.tools import file_interface, log
from evo.tools.settings import SETTINGS
import matplotlib.pyplot as plt
import matplotlib
import copy
import os
matplotlib.rcParams['mathtext.fontset'] = 'custom'
matplotlib.rcParams['mathtext.rm'] = 'Times New Roman'
matplotlib.rcParams['mathtext.it'] = 'Times New Roman:italic'
matplotlib.rcParams['mathtext.bf'] = 'Times New Roman:bold'
matplotlib.rcParams['font.family'] = 'Arial'
font0={'family':'Arial',
'style':'normal',
'weight':'bold',
'color':'black',
'size':6
}
font1={'family':'Arial',
'style':'normal',
'weight':'bold',
'color':'black',
'size':8
}
logger = logging.getLogger(__name__)
SEP = "-" * 80 # separator line
def ape(traj_ref: PosePath3D, traj_est: PosePath3D,
pose_relation: metrics.PoseRelation, align: bool = False,
correct_scale: bool = False, n_to_align: int = -1,
align_origin: bool = False, ref_name: str = "reference",
est_name: str = "estimate",
change_unit: typing.Optional[metrics.Unit] = None) -> Result:
if n_to_align >0 :
print('>>>>> only use the starting segment')
n_to_align = np.where((np.array(traj_ref.timestamps)[1:]-np.array(traj_ref.timestamps)[0:-1])>100)[0][0]-1
# Align the trajectories.
only_scale = correct_scale and not align
alignment_transformation = None
if align or correct_scale:
logger.debug(SEP)
alignment_transformation = lie_algebra.sim3(
*traj_est.align(traj_ref, correct_scale, only_scale, n=n_to_align))
elif align_origin:
logger.debug(SEP)
alignment_transformation = traj_est.align_origin(traj_ref)
# Calculate APE.
logger.debug(SEP)
data = (traj_ref, traj_est)
ape_metric = metrics.APE(pose_relation)
ape_metric.process_data(data)
if change_unit:
ape_metric.change_unit(change_unit)
title = str(ape_metric)
if align and not correct_scale:
title += "\n(with SE(3) Umeyama alignment)"
elif align and correct_scale:
title += "\n(with Sim(3) Umeyama alignment)"
elif only_scale:
title += "\n(scale corrected)"
elif align_origin:
title += "\n(with origin alignment)"
else:
title += "\n(not aligned)"
if (align or correct_scale) and n_to_align != -1:
title += " (aligned poses: {})".format(n_to_align)
ape_result = ape_metric.get_result(ref_name, est_name)
ape_result.info["title"] = title
logger.debug(SEP)
logger.info(ape_result.pretty_str())
ape_result.add_trajectory(ref_name, traj_ref)
ape_result.add_trajectory(est_name, traj_est)
if isinstance(traj_est, PoseTrajectory3D):
seconds_from_start = np.array(
[t - traj_est.timestamps[0] for t in traj_est.timestamps])
ape_result.add_np_array("seconds_from_start", seconds_from_start)
ape_result.add_np_array("timestamps", traj_est.timestamps)
ape_result.add_np_array("distances_from_start", traj_ref.distances)
ape_result.add_np_array("distances", traj_est.distances)
if alignment_transformation is not None:
ape_result.add_np_array("alignment_transformation_sim3",
alignment_transformation)
return ape_result
if __name__ == '__main__':
color_list = [np.array([133,164,195])/255.0,[1,0.6,1],[1,0,0]]
plt.figure('1',figsize=[6,6])
parser = argparse.ArgumentParser()
parser.add_argument('--seq', type=str, help='seq', default='outdoors1')
parser.add_argument('--batch', action="store_true")
args = parser.parse_args()
args.subcommand = 'tum'
seq = args.seq
args.ref_file = '/mnt/z/tum-vi/dataset-%s_512_16/dso/gt_imu.csv' % seq
# Convert GT format
dd = np.loadtxt(args.ref_file,delimiter=',',comments='#')
dd_new = np.copy(dd)
dd_new[:,0] /= 1e9
dd_new[:,4] = dd[:,7]
dd_new[:,5] = dd[:,4]
dd_new[:,6] = dd[:,5]
dd_new[:,7] = dd[:,6]
args.ref_file = os.path.join(os.path.dirname(args.ref_file),'gt_imu_temp.txt')
np.savetxt(args.ref_file, dd_new,delimiter=' ')
args.pose_relation = 'trans_part'
args.align = True
args.correct_scale = False
args.n_to_align = 1
args.align_origin = False
args.plot_mode = 'xyz'
args.plot_x_dimension = 'seconds'
args.plot_colormap_min = None
args.plot_colormap_max = None
args.plot_colormap_max_percentile = None
args.ros_map_yaml = None
args.plot = True
args.est_files = ['results/result_%s.txt' % seq]
label_list = ['DBA-Fusion (M)']
args.save_plot = False
args.serialize_plot = False
for iii in range(len(args.est_files)):
args.est_file = args.est_files[iii]
if args.est_file.find('visual') != -1:
args.correct_scale = True
else:
args.correct_scale = False
traj_ref, traj_est, ref_name, est_name = common.load_trajectories(args)
traj_ref_sel, traj_est_sel = sync.associate_trajectories(
traj_ref, traj_est, 0.01,0.0,
first_name=ref_name, snd_name=est_name)
# use the starting part for scale estimation
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
pose_relation = common.get_pose_relation(args)
result = ape(traj_ref=traj_ref_sel, traj_est=traj_est_sel,
pose_relation=pose_relation, align=args.align,
correct_scale=args.correct_scale, n_to_align=args.n_to_align,
align_origin=args.align_origin, ref_name=ref_name,
est_name=est_name)
traj_est_sel = copy.deepcopy(result.trajectories[est_name])
T01 = result.np_arrays['alignment_transformation_sim3']
# metric-scale APE calculation
result = ape(traj_ref=traj_ref_sel, traj_est=traj_est_sel,
pose_relation=pose_relation, align=args.align,
correct_scale=False, n_to_align=-1,
align_origin=args.align_origin, ref_name=ref_name,
est_name=est_name)
print(result)
if args.batch:
quit()
# visualization
traj_est.transform(T01)
if iii == 0:
x_series=[]
y_series=[]
z_series=[]
for i in range(len(traj_ref.poses_se3)):
TTT = traj_ref.poses_se3[i]
x_series.append(TTT[0,3])
y_series.append(TTT[1,3])
z_series.append(TTT[2,3])
plt.plot(x_series,y_series,c=[0,0,0],linestyle = '--',linewidth=0.5)
x_series=[]
y_series=[]
z_series=[]
for i in range(len(traj_est.poses_se3)):
TTT = traj_est.poses_se3[i]
x_series.append(TTT[0,3])
y_series.append(TTT[1,3])
z_series.append(TTT[2,3])
plt.plot(x_series,y_series,c=color_list[iii],label = label_list[iii],linewidth=0.5)
ll = max(max(x_series)-min(x_series),max(y_series)-min(y_series))
plt.xlim([(max(x_series)+min(x_series))/2 - 0.65*ll,(max(x_series)+min(x_series))/2+0.65*ll])
plt.ylim([(max(y_series)+min(y_series))/2 - 0.65*ll,(max(y_series)+min(y_series))/2+0.65*ll])
# plt.xlabel('X [m]')
# plt.ylabel('Y [m]')
plt.tick_params(labelsize=6,direction='in')
lg = plt.legend(loc='upper right',markerscale=3,fontsize=5,framealpha=1,ncol=1,columnspacing=0.3,handletextpad=0.3,edgecolor='black',fancybox=False)
lg.set_zorder(200)
lg.get_frame().set_linewidth(0.8)
plt.gca().yaxis.set_label_coords(-.1, .5)
plt.show()
================================================
FILE: results/PLACEHOLDER
================================================
================================================
FILE: setup.py
================================================
from setuptools import setup
from torch.utils.cpp_extension import BuildExtension, CUDAExtension
import os.path as osp
ROOT = osp.dirname(osp.abspath(__file__))
setup(
name='droid_backends',
ext_modules=[
CUDAExtension('droid_backends',
include_dirs=[osp.join(ROOT, 'thirdparty/eigen')],
sources=[
'src/droid.cpp',
'src/droid_kernels.cu',
'src/correlation_kernels.cu',
'src/altcorr_kernel.cu',
],
extra_compile_args={
'cxx': ['-O3'],
'nvcc': ['-O3',
'-gencode=arch=compute_60,code=sm_60',
'-gencode=arch=compute_61,code=sm_61',
'-gencode=arch=compute_70,code=sm_70',
'-gencode=arch=compute_75,code=sm_75',
'-gencode=arch=compute_80,code=sm_80',
'-gencode=arch=compute_86,code=sm_86',
]
}),
],
cmdclass={ 'build_ext' : BuildExtension }
)
setup(
name='lietorch',
version='0.2',
description='Lie Groups for PyTorch',
packages=['lietorch'],
package_dir={'': 'thirdparty/lietorch'},
ext_modules=[
CUDAExtension('lietorch_backends',
include_dirs=[
osp.join(ROOT, 'thirdparty/lietorch/lietorch/include'),
osp.join(ROOT, 'thirdparty/eigen')],
sources=[
'thirdparty/lietorch/lietorch/src/lietorch.cpp',
'thirdparty/lietorch/lietorch/src/lietorch_gpu.cu',
'thirdparty/lietorch/lietorch/src/lietorch_cpu.cpp'],
extra_compile_args={
'cxx': ['-O2'],
'nvcc': ['-O2',
'-gencode=arch=compute_60,code=sm_60',
'-gencode=arch=compute_61,code=sm_61',
'-gencode=arch=compute_70,code=sm_70',
'-gencode=arch=compute_75,code=sm_75',
'-gencode=arch=compute_80,code=sm_80',
'-gencode=arch=compute_86,code=sm_86',
]
}),
],
cmdclass={ 'build_ext' : BuildExtension }
)
================================================
FILE: src/altcorr_kernel.cu
================================================
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#define BLOCK_H 4
#define BLOCK_W 8
#define BLOCK_HW BLOCK_H * BLOCK_W
#define CHANNEL_STRIDE 32
__forceinline__ __device__
bool within_bounds(int h, int w, int H, int W) {
return h >= 0 && h < H && w >= 0 && w < W;
}
template
__global__ void altcorr_forward_kernel(
const torch::PackedTensorAccessor32 fmap1,
const torch::PackedTensorAccessor32 fmap2,
const torch::PackedTensorAccessor32 coords,
torch::PackedTensorAccessor32 corr,
int r)
{
const int b = blockIdx.x;
const int h0 = blockIdx.y * blockDim.x;
const int w0 = blockIdx.z * blockDim.y;
const int tid = threadIdx.x * blockDim.y + threadIdx.y;
const int H1 = fmap1.size(1);
const int W1 = fmap1.size(2);
const int H2 = fmap2.size(1);
const int W2 = fmap2.size(2);
const int N = coords.size(1);
const int C = fmap1.size(3);
__shared__ scalar_t f1[CHANNEL_STRIDE][BLOCK_HW];
__shared__ scalar_t f2[CHANNEL_STRIDE][BLOCK_HW];
__shared__ float x2s[BLOCK_HW];
__shared__ float y2s[BLOCK_HW];
for (int c=0; c(floor(y2s[k1])) - r + iy;
int w2 = static_cast(floor(x2s[k1])) - r + ix;
int c2 = tid % CHANNEL_STRIDE;
if (within_bounds(h2, w2, H2, W2))
f2[c2][k1] = fmap2[b][h2][w2][c+c2];
else
f2[c2][k1] = static_cast(0.0);
}
__syncthreads();
scalar_t s = 0.0;
for (int k=0; k((dy) * (dx));
scalar_t ne = s * static_cast((dy) * (1-dx));
scalar_t sw = s * static_cast((1-dy) * (dx));
scalar_t se = s * static_cast((1-dy) * (1-dx));
// if (iy > 0 && ix > 0 && within_bounds(h1, w1, H1, W1))
// corr[b][n][ix_nw][h1][w1] += nw;
// if (iy > 0 && ix < rd && within_bounds(h1, w1, H1, W1))
// corr[b][n][ix_ne][h1][w1] += ne;
// if (iy < rd && ix > 0 && within_bounds(h1, w1, H1, W1))
// corr[b][n][ix_sw][h1][w1] += sw;
// if (iy < rd && ix < rd && within_bounds(h1, w1, H1, W1))
// corr[b][n][ix_se][h1][w1] += se;
scalar_t* corr_ptr = &corr[b][n][0][h1][w1];
if (iy > 0 && ix > 0 && within_bounds(h1, w1, H1, W1))
*(corr_ptr + ix_nw) += nw;
if (iy > 0 && ix < rd && within_bounds(h1, w1, H1, W1))
*(corr_ptr + ix_ne) += ne;
if (iy < rd && ix > 0 && within_bounds(h1, w1, H1, W1))
*(corr_ptr + ix_sw) += sw;
if (iy < rd && ix < rd && within_bounds(h1, w1, H1, W1))
*(corr_ptr + ix_se) += se;
}
}
}
}
}
template
__global__ void altcorr_backward_kernel(
const torch::PackedTensorAccessor32 fmap1,
const torch::PackedTensorAccessor32 fmap2,
const torch::PackedTensorAccessor32 coords,
const torch::PackedTensorAccessor32 corr_grad,
torch::PackedTensorAccessor32 fmap1_grad,
torch::PackedTensorAccessor32 fmap2_grad,
torch::PackedTensorAccessor32 coords_grad,
int r)
{
const int b = blockIdx.x;
const int h0 = blockIdx.y * blockDim.x;
const int w0 = blockIdx.z * blockDim.y;
const int tid = threadIdx.x * blockDim.y + threadIdx.y;
const int H1 = fmap1.size(1);
const int W1 = fmap1.size(2);
const int H2 = fmap2.size(1);
const int W2 = fmap2.size(2);
const int N = coords.size(1);
const int C = fmap1.size(3);
__shared__ scalar_t f1[CHANNEL_STRIDE][BLOCK_HW+1];
__shared__ scalar_t f2[CHANNEL_STRIDE][BLOCK_HW+1];
__shared__ scalar_t f1_grad[CHANNEL_STRIDE][BLOCK_HW+1];
__shared__ scalar_t f2_grad[CHANNEL_STRIDE][BLOCK_HW+1];
__shared__ scalar_t x2s[BLOCK_HW];
__shared__ scalar_t y2s[BLOCK_HW];
for (int c=0; c(floor(y2s[k1]))-r+iy;
int w2 = static_cast(floor(x2s[k1]))-r+ix;
int c2 = tid % CHANNEL_STRIDE;
auto fptr = fmap2[b][h2][w2];
if (within_bounds(h2, w2, H2, W2))
f2[c2][k1] = fptr[c+c2];
else
f2[c2][k1] = 0.0;
f2_grad[c2][k1] = 0.0;
}
__syncthreads();
const scalar_t* grad_ptr = &corr_grad[b][n][0][h1][w1];
scalar_t g = 0.0;
int ix_nw = H1*W1*((iy-1) + rd*(ix-1));
int ix_ne = H1*W1*((iy-1) + rd*ix);
int ix_sw = H1*W1*(iy + rd*(ix-1));
int ix_se = H1*W1*(iy + rd*ix);
if (iy > 0 && ix > 0 && within_bounds(h1, w1, H1, W1))
g += *(grad_ptr + ix_nw) * dy * dx;
if (iy > 0 && ix < rd && within_bounds(h1, w1, H1, W1))
g += *(grad_ptr + ix_ne) * dy * (1-dx);
if (iy < rd && ix > 0 && within_bounds(h1, w1, H1, W1))
g += *(grad_ptr + ix_sw) * (1-dy) * dx;
if (iy < rd && ix < rd && within_bounds(h1, w1, H1, W1))
g += *(grad_ptr + ix_se) * (1-dy) * (1-dx);
for (int k=0; k(floor(y2s[k1]))-r+iy;
int w2 = static_cast(floor(x2s[k1]))-r+ix;
int c2 = tid % CHANNEL_STRIDE;
scalar_t* fptr = &fmap2_grad[b][h2][w2][0];
if (within_bounds(h2, w2, H2, W2))
atomicAdd(fptr+c+c2, f2_grad[c2][k1]);
}
}
}
}
__syncthreads();
for (int k=0; k altcorr_cuda_forward(
torch::Tensor fmap1,
torch::Tensor fmap2,
torch::Tensor coords,
int radius)
{
const auto B = coords.size(0);
const auto N = coords.size(1);
const auto H = coords.size(2);
const auto W = coords.size(3);
const auto rd = 2 * radius + 1;
auto opts = fmap1.options();
auto corr = torch::zeros({B, N, rd*rd, H, W}, opts);
const dim3 blocks(B, (H+BLOCK_H-1)/BLOCK_H, (W+BLOCK_W-1)/BLOCK_W);
const dim3 threads(BLOCK_H, BLOCK_W);
AT_DISPATCH_FLOATING_TYPES_AND_HALF(fmap1.type(), "altcorr_forward_kernel", ([&] {
altcorr_forward_kernel<<>>(
fmap1.packed_accessor32(),
fmap2.packed_accessor32(),
coords.packed_accessor32(),
corr.packed_accessor32(),
radius);
}));
return {corr};
}
std::vector altcorr_cuda_backward(
torch::Tensor fmap1,
torch::Tensor fmap2,
torch::Tensor coords,
torch::Tensor corr_grad,
int radius)
{
const auto B = coords.size(0);
const auto N = coords.size(1);
const auto H1 = fmap1.size(1);
const auto W1 = fmap1.size(2);
const auto H2 = fmap2.size(1);
const auto W2 = fmap2.size(2);
const auto C = fmap1.size(3);
auto opts = fmap1.options();
auto fmap1_grad = torch::zeros({B, H1, W1, C}, opts);
auto fmap2_grad = torch::zeros({B, H2, W2, C}, opts);
auto coords_grad = torch::zeros({B, N, H1, W1, 2}, opts);
const dim3 blocks(B, (H1+BLOCK_H-1)/BLOCK_H, (W1+BLOCK_W-1)/BLOCK_W);
const dim3 threads(BLOCK_H, BLOCK_W);
altcorr_backward_kernel<<>>(
fmap1.packed_accessor32(),
fmap2.packed_accessor32(),
coords.packed_accessor32(),
corr_grad.packed_accessor32(),
fmap1_grad.packed_accessor32(),
fmap2_grad.packed_accessor32(),
coords_grad.packed_accessor32(),
radius);
return {fmap1_grad, fmap2_grad, coords_grad};
}
================================================
FILE: src/bacore.h
================================================
#include
#include
class BACore
{
public:
BACore(){}
~BACore(){}
public:
void init(torch::Tensor _poses,
torch::Tensor _disps,
torch::Tensor _intrinsics,
torch::Tensor _disps_sens,
torch::Tensor _targets,
torch::Tensor _weights,
torch::Tensor _eta,
torch::Tensor _ii,
torch::Tensor _jj,
const int t0,
const int t1,
const int iterations,
const float lm,
const float ep,
const bool motion_only);
void hessian(torch::Tensor H, torch::Tensor v);
void optimize(torch::Tensor H, torch::Tensor v);
std::vector retract(torch::Tensor _dx);
public:
torch::Tensor poses;
torch::Tensor disps;
torch::Tensor intrinsics;
torch::Tensor disps_sens;
torch::Tensor targets;
torch::Tensor weights;
torch::Tensor eta;
torch::Tensor ii;
torch::Tensor jj;
int t0,t1;
float lm, ep;
torch::Tensor ts;
torch::Tensor ii_exp;
torch::Tensor jj_exp;
std::tuple kuniq;
torch::Tensor kx;
torch::Tensor kk_exp; // 不重复元素的索引
torch::Tensor dx;
torch::Tensor dz;
// initialize buffers
torch::Tensor Hs;
torch::Tensor vs;
torch::Tensor Eii;
torch::Tensor Eij;
torch::Tensor Cii;
torch::Tensor wi;
torch::Tensor m ;
torch::Tensor C ;
torch::Tensor w ;
torch::Tensor Q ;
torch::Tensor Ei;
torch::Tensor E ;
};
================================================
FILE: src/correlation_kernels.cu
================================================
#include
#include
#include
#include
#include
#include
#include
#include
#include
#define BLOCK 16
__forceinline__ __device__ bool within_bounds(int h, int w, int H, int W) {
return h >= 0 && h < H && w >= 0 && w < W;
}
template
__global__ void corr_index_forward_kernel(
const torch::PackedTensorAccessor32 volume,
const torch::PackedTensorAccessor32 coords,
torch::PackedTensorAccessor32 corr,
int r)
{
// batch index
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int n = blockIdx.z;
const int h1 = volume.size(1);
const int w1 = volume.size(2);
const int h2 = volume.size(3);
const int w2 = volume.size(4);
if (!within_bounds(y, x, h1, w1)) {
return;
}
float x0 = coords[n][0][y][x];
float y0 = coords[n][1][y][x];
float dx = x0 - floor(x0);
float dy = y0 - floor(y0);
int rd = 2*r + 1;
for (int i=0; i(floor(x0)) - r + i;
int y1 = static_cast(floor(y0)) - r + j;
if (within_bounds(y1, x1, h2, w2)) {
scalar_t s = volume[n][y][x][y1][x1];
if (i > 0 && j > 0)
corr[n][i-1][j-1][y][x] += s * scalar_t(dx * dy);
if (i > 0 && j < rd)
corr[n][i-1][j][y][x] += s * scalar_t(dx * (1.0f-dy));
if (i < rd && j > 0)
corr[n][i][j-1][y][x] += s * scalar_t((1.0f-dx) * dy);
if (i < rd && j < rd)
corr[n][i][j][y][x] += s * scalar_t((1.0f-dx) * (1.0f-dy));
}
}
}
}
template
__global__ void corr_index_backward_kernel(
const torch::PackedTensorAccessor32 coords,
const torch::PackedTensorAccessor32 corr_grad,
torch::PackedTensorAccessor32 volume_grad,
int r)
{
// batch index
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
const int n = blockIdx.z;
const int h1 = volume_grad.size(1);
const int w1 = volume_grad.size(2);
const int h2 = volume_grad.size(3);
const int w2 = volume_grad.size(4);
if (!within_bounds(y, x, h1, w1)) {
return;
}
float x0 = coords[n][0][y][x];
float y0 = coords[n][1][y][x];
float dx = x0 - floor(x0);
float dy = y0 - floor(y0);
int rd = 2*r + 1;
for (int i=0; i(floor(x0)) - r + i;
int y1 = static_cast(floor(y0)) - r + j;
if (within_bounds(y1, x1, h2, w2)) {
scalar_t g = 0.0;
if (i > 0 && j > 0)
g += corr_grad[n][i-1][j-1][y][x] * scalar_t(dx * dy);
if (i > 0 && j < rd)
g += corr_grad[n][i-1][j][y][x] * scalar_t(dx * (1.0f-dy));
if (i < rd && j > 0)
g += corr_grad[n][i][j-1][y][x] * scalar_t((1.0f-dx) * dy);
if (i < rd && j < rd)
g += corr_grad[n][i][j][y][x] * scalar_t((1.0f-dx) * (1.0f-dy));
volume_grad[n][y][x][y1][x1] += g;
}
}
}
}
std::vector corr_index_cuda_forward(
torch::Tensor volume,
torch::Tensor coords,
int radius)
{
const auto batch_size = volume.size(0);
const auto ht = volume.size(1);
const auto wd = volume.size(2);
const dim3 blocks((wd + BLOCK - 1) / BLOCK,
(ht + BLOCK - 1) / BLOCK,
batch_size);
const dim3 threads(BLOCK, BLOCK);
auto opts = volume.options();
torch::Tensor corr = torch::zeros(
{batch_size, 2*radius+1, 2*radius+1, ht, wd}, opts);
AT_DISPATCH_FLOATING_TYPES_AND_HALF(volume.type(), "sampler_forward_kernel", ([&] {
corr_index_forward_kernel<<>>(
volume.packed_accessor32(),
coords.packed_accessor32(),
corr.packed_accessor32(),
radius);
}));
return {corr};
}
std::vector corr_index_cuda_backward(
torch::Tensor volume,
torch::Tensor coords,
torch::Tensor corr_grad,
int radius)
{
const auto batch_size = volume.size(0);
const auto ht = volume.size(1);
const auto wd = volume.size(2);
auto volume_grad = torch::zeros_like(volume);
const dim3 blocks((wd + BLOCK - 1) / BLOCK,
(ht + BLOCK - 1) / BLOCK,
batch_size);
const dim3 threads(BLOCK, BLOCK);
AT_DISPATCH_FLOATING_TYPES_AND_HALF(volume.type(), "sampler_backward_kernel", ([&] {
corr_index_backward_kernel<<>>(
coords.packed_accessor32(),
corr_grad.packed_accessor32(),
volume_grad.packed_accessor32(),
radius);
}));
return {volume_grad};
}
================================================
FILE: src/droid.cpp
================================================
#include
#include
#include "bacore.h"
// CUDA forward declarations
std::vector projective_transform_cuda(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor ii,
torch::Tensor jj);
torch::Tensor depth_filter_cuda(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor ix,
torch::Tensor thresh);
torch::Tensor frame_distance_cuda(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor ii,
torch::Tensor jj,
const float beta);
std::vector projmap_cuda(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor ii,
torch::Tensor jj);
torch::Tensor iproj_cuda(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics);
std::vector ba_cuda(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor disps_sens,
torch::Tensor targets,
torch::Tensor weights,
torch::Tensor eta,
torch::Tensor ii,
torch::Tensor jj,
const int t0,
const int t1,
const int iterations,
const float lm,
const float ep,
const bool motion_only);
std::vector ba_cuda_extend(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor disps_sens,
torch::Tensor targets,
torch::Tensor weights,
torch::Tensor eta,
torch::Tensor ii,
torch::Tensor jj,
torch::Tensor H,
torch::Tensor v,
torch::Tensor A_prior,
const int t0,
const int t1,
const int iterations,
const float lm,
const float ep,
const bool motion_only,
const bool skip_solve);
std::vector corr_index_cuda_forward(
torch::Tensor volume,
torch::Tensor coords,
int radius);
std::vector corr_index_cuda_backward(
torch::Tensor volume,
torch::Tensor coords,
torch::Tensor corr_grad,
int radius);
std::vector altcorr_cuda_forward(
torch::Tensor fmap1,
torch::Tensor fmap2,
torch::Tensor coords,
int radius);
std::vector altcorr_cuda_backward(
torch::Tensor fmap1,
torch::Tensor fmap2,
torch::Tensor coords,
torch::Tensor corr_grad,
int radius);
#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")
#define CHECK_INPUT(x) CHECK_CONTIGUOUS(x)
std::vector ba(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor disps_sens,
torch::Tensor targets,
torch::Tensor weights,
torch::Tensor eta,
torch::Tensor ii,
torch::Tensor jj,
const int t0,
const int t1,
const int iterations,
const float lm,
const float ep,
const bool motion_only) {
CHECK_INPUT(targets);
CHECK_INPUT(weights);
CHECK_INPUT(poses);
CHECK_INPUT(disps);
CHECK_INPUT(intrinsics);
CHECK_INPUT(disps_sens);
CHECK_INPUT(ii);
CHECK_INPUT(jj);
return ba_cuda(poses, disps, intrinsics, disps_sens, targets, weights,
eta, ii, jj, t0, t1, iterations, lm, ep, motion_only);
}
std::vector ba_extend(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor disps_sens,
torch::Tensor targets,
torch::Tensor weights,
torch::Tensor eta,
torch::Tensor ii,
torch::Tensor jj,
torch::Tensor H,
torch::Tensor v,
torch::Tensor A_prior,
const int t0,
const int t1,
const int iterations,
const float lm,
const float ep,
const bool motion_only,
const bool skip_solve) {
CHECK_INPUT(targets);
CHECK_INPUT(weights);
CHECK_INPUT(poses);
CHECK_INPUT(disps);
CHECK_INPUT(intrinsics);
CHECK_INPUT(disps_sens);
CHECK_INPUT(ii);
CHECK_INPUT(jj);
CHECK_INPUT(H);
CHECK_INPUT(v);
CHECK_INPUT(A_prior);
std::vector dx_dz = ba_cuda_extend(poses, disps, intrinsics, disps_sens, targets, weights,
eta, ii, jj, H, v, A_prior, t0, t1, iterations, lm, ep, motion_only, skip_solve);
return dx_dz;
}
torch::Tensor frame_distance(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor ii,
torch::Tensor jj,
const float beta) {
CHECK_INPUT(poses);
CHECK_INPUT(disps);
CHECK_INPUT(intrinsics);
CHECK_INPUT(ii);
CHECK_INPUT(jj);
return frame_distance_cuda(poses, disps, intrinsics, ii, jj, beta);
}
std::vector projmap(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor ii,
torch::Tensor jj) {
CHECK_INPUT(poses);
CHECK_INPUT(disps);
CHECK_INPUT(intrinsics);
CHECK_INPUT(ii);
CHECK_INPUT(jj);
return projmap_cuda(poses, disps, intrinsics, ii, jj);
}
torch::Tensor iproj(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics) {
CHECK_INPUT(poses);
CHECK_INPUT(disps);
CHECK_INPUT(intrinsics);
return iproj_cuda(poses, disps, intrinsics);
}
// c++ python binding
std::vector corr_index_forward(
torch::Tensor volume,
torch::Tensor coords,
int radius) {
CHECK_INPUT(volume);
CHECK_INPUT(coords);
return corr_index_cuda_forward(volume, coords, radius);
}
std::vector corr_index_backward(
torch::Tensor volume,
torch::Tensor coords,
torch::Tensor corr_grad,
int radius) {
CHECK_INPUT(volume);
CHECK_INPUT(coords);
CHECK_INPUT(corr_grad);
auto volume_grad = corr_index_cuda_backward(volume, coords, corr_grad, radius);
return {volume_grad};
}
std::vector altcorr_forward(
torch::Tensor fmap1,
torch::Tensor fmap2,
torch::Tensor coords,
int radius) {
CHECK_INPUT(fmap1);
CHECK_INPUT(fmap2);
CHECK_INPUT(coords);
return altcorr_cuda_forward(fmap1, fmap2, coords, radius);
}
std::vector altcorr_backward(
torch::Tensor fmap1,
torch::Tensor fmap2,
torch::Tensor coords,
torch::Tensor corr_grad,
int radius) {
CHECK_INPUT(fmap1);
CHECK_INPUT(fmap2);
CHECK_INPUT(coords);
CHECK_INPUT(corr_grad);
return altcorr_cuda_backward(fmap1, fmap2, coords, corr_grad, radius);
}
torch::Tensor depth_filter(
torch::Tensor poses,
torch::Tensor disps,
torch::Tensor intrinsics,
torch::Tensor ix,
torch::Tensor thresh) {
CHECK_INPUT(poses);
CHECK_INPUT(disps);
CHECK_INPUT(intrinsics);
CHECK_INPUT(ix);
CHECK_INPUT(thresh);
return depth_filter_cuda(poses, disps, intrinsics, ix, thresh);
}
PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
// bundle adjustment kernels
m.def("ba", &ba, "bundle adjustment");
m.def("ba_extend", &ba_extend, "bundle adjustment (extended)");
m.def("frame_distance", &frame_distance, "frame_distance");
m.def("projmap", &projmap, "projmap");
m.def("depth_filter", &depth_filter, "depth_filter");
m.def("iproj", &iproj, "back projection");
// correlation volume kernels
m.def("altcorr_forward", &altcorr_forward, "ALTCORR forward");
m.def("altcorr_backward", &altcorr_backward, "ALTCORR backward");
m.def("corr_index_forward", &corr_index_forward, "INDEX forward");
m.def("corr_index_backward", &corr_index_backward, "INDEX backward");
py::class_(m, "BACore").def(py::init<>())
.def("init", &BACore::init)
.def("hessian", &BACore::hessian)
.def("optimize", &BACore::optimize)
.def("retract", &BACore::retract);
}
================================================
FILE: src/droid_kernels.cu
================================================
#include
#include
#include
#include
#include
#include
#include
#include
#include
// #include "utils.cuh"
#include
#include
#include
#include
#include "bacore.h"
typedef Eigen::SparseMatrix SpMat;
typedef Eigen::Triplet T;
typedef std::vector> graph_t;
typedef std::vector tensor_list_t;
#define MIN_DEPTH 0.25
#define THREADS 256
#define NUM_BLOCKS(batch_size) ((batch_size + THREADS - 1) / THREADS)
#define GPU_1D_KERNEL_LOOP(k, n) \
for (size_t k = threadIdx.x; k 1e-4) {
float a = (1 - cosf(theta)) / theta_sq;
crossInplace(phi, tau);
t[0] += a * tau[0];
t[1] += a * tau[1];
t[2] += a * tau[2];
float b = (theta - sinf(theta)) / (theta * theta_sq);
crossInplace(phi, tau);
t[0] += b * tau[0];
t[1] += b * tau[1];
t[2] += b * tau[2];
}
}
torch::Tensor solveDense(const Eigen::MatrixXd& A, const Eigen::MatrixXd& Aprior, const Eigen::VectorXd& b,
const int N, const int M,
const float lm=0.0001,const float ep = 0.1){
torch::Tensor dx;
Eigen::MatrixXd L = A + Aprior;
L.diagonal().array() += ep + lm * L.diagonal().array();
Eigen::LLT solver;
solver.compute(L);
if (solver.info() == Eigen::Success) {
Eigen::VectorXd x = solver.solve(b);
dx = torch::from_blob(x.data(), {N, M}, torch::TensorOptions()
.dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);
}
else {
dx = torch::zeros({N, M}, torch::TensorOptions()
.device(torch::kCUDA).dtype(torch::kFloat32));
}
return dx;
}
torch::Tensor solveDenseD(const Eigen::MatrixXd& A, const Eigen::VectorXd& b,
const int N, const int M,
const float lm=0.0001,const float ep = 0.1){
torch::Tensor dx;
Eigen::MatrixXd L(A);
L.diagonal().array() += ep + lm * L.diagonal().array();
Eigen::LLT solver;
solver.compute(L);
if (solver.info() == Eigen::Success) {
Eigen::VectorXd x = solver.solve(b);
dx = torch::from_blob(x.data(), {N, M}, torch::TensorOptions()
.dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);
}
else {
dx = torch::zeros({N, M}, torch::TensorOptions()
.device(torch::kCUDA).dtype(torch::kFloat32));
}
return dx;
}
__global__ void projective_transform_kernel(
const torch::PackedTensorAccessor32 target,
const torch::PackedTensorAccessor32 weight,
const torch::PackedTensorAccessor32 poses,
const torch::PackedTensorAccessor32 disps,
const torch::PackedTensorAccessor32 intrinsics,
const torch::PackedTensorAccessor32 ii,
const torch::PackedTensorAccessor32 jj,
torch::PackedTensorAccessor32 Hs, // Hessian of Poses
torch::PackedTensorAccessor32 vs, // Residuals (Pose)
torch::PackedTensorAccessor32 Eii,// Hessian, Disps_i - Pose_i
torch::PackedTensorAccessor32 Eij,// Hessian, Disps_i - Pose_j
torch::PackedTensorAccessor32 Cii,// Hessian, Disps_i - Disps_i
torch::PackedTensorAccessor32 bz) // Residuals (Disp)
{
const int block_id = blockIdx.x;
const int thread_id = threadIdx.x;
const int ht = disps.size(1);
const int wd = disps.size(2);
int ix = static_cast(ii[block_id]);
int jx = static_cast(jj[block_id]);
__shared__ float fx;
__shared__ float fy;
__shared__ float cx;
__shared__ float cy;
__shared__ float ti[3], tj[3], tij[3];
__shared__ float qi[4], qj[4], qij[4];
// load intrinsics from global memory
if (thread_id == 0) {
fx = intrinsics[0];
fy = intrinsics[1];
cx = intrinsics[2];
cy = intrinsics[3];
}
__syncthreads();
// stereo frames
if (ix == jx) {
if (thread_id == 0) {
tij[0] = -0.1;
tij[1] = 0;
tij[2] = 0;
qij[0] = 0;
qij[1] = 0;
qij[2] = 0;
qij[3] = 1;
}
}
else {
// load poses from global memory
if (thread_id < 3) {
ti[thread_id] = poses[ix][thread_id];
tj[thread_id] = poses[jx][thread_id];
}
if (thread_id < 4) {
qi[thread_id] = poses[ix][thread_id+3];
qj[thread_id] = poses[jx][thread_id+3];
}
__syncthreads();
if (thread_id == 0) {
relSE3(ti, qi, tj, qj, tij, qij);
}
}
__syncthreads();
//points
float Xi[4];
float Xj[4];
// jacobians
float Jx[12];
float Jz;
float* Ji = &Jx[0];
float* Jj = &Jx[6];
// hessians
float hij[12*(12+1)/2];
float vi[6], vj[6];
int l;
for (l=0; l<12*(12+1)/2; l++) {
hij[l] = 0;
}
for (int n=0; n<6; n++) {
vi[n] = 0;
vj[n] = 0;
}
__syncthreads();
GPU_1D_KERNEL_LOOP(k, ht*wd) {
const int i = k / wd;
const int j = k % wd;
const float u = static_cast(j);
const float v = static_cast(i);
// homogenous coordinates
Xi[0] = (u - cx) / fx;
Xi[1] = (v - cy) / fy;
Xi[2] = 1;
Xi[3] = disps[ix][i][j];
// transform homogenous point
actSE3(tij, qij, Xi, Xj);
const float x = Xj[0];
const float y = Xj[1];
const float h = Xj[3];
const float d = (Xj[2] < MIN_DEPTH) ? 0.0 : 1.0 / Xj[2];
const float d2 = d * d;
float wu = (Xj[2] < MIN_DEPTH) ? 0.0 : .001 * weight[block_id][0][i][j];
float wv = (Xj[2] < MIN_DEPTH) ? 0.0 : .001 * weight[block_id][1][i][j];
const float ru = target[block_id][0][i][j] - (fx * d * x + cx);
const float rv = target[block_id][1][i][j] - (fy * d * y + cy);
// x - coordinate
Jj[0] = fx * (h*d);
Jj[1] = fx * 0;
Jj[2] = fx * (-x*h*d2);
Jj[3] = fx * (-x*y*d2);
Jj[4] = fx * (1 + x*x*d2);
Jj[5] = fx * (-y*d);
Jz = fx * (tij[0] * d - tij[2] * (x * d2));
Cii[block_id][k] = wu * Jz * Jz;
bz[block_id][k] = wu * ru * Jz;
if (ix == jx) wu = 0;
adjSE3(tij, qij, Jj, Ji);
for (int n=0; n<6; n++) Ji[n] *= -1;
l=0;
for (int n=0; n<12; n++) {
for (int m=0; m<=n; m++) {
hij[l] += wu * Jx[n] * Jx[m];
l++;
}
}
for (int n=0; n<6; n++) {
vi[n] += wu * ru * Ji[n];
vj[n] += wu * ru * Jj[n];
Eii[block_id][n][k] = wu * Jz * Ji[n];
Eij[block_id][n][k] = wu * Jz * Jj[n];
}
Jj[0] = fy * 0;
Jj[1] = fy * (h*d);
Jj[2] = fy * (-y*h*d2);
Jj[3] = fy * (-1 - y*y*d2);
Jj[4] = fy * (x*y*d2);
Jj[5] = fy * (x*d);
Jz = fy * (tij[1] * d - tij[2] * (y * d2));
Cii[block_id][k] += wv * Jz * Jz;
bz[block_id][k] += wv * rv * Jz;
if (ix == jx) wv = 0;
adjSE3(tij, qij, Jj, Ji);
for (int n=0; n<6; n++) Ji[n] *= -1;
l=0;
for (int n=0; n<12; n++) {
for (int m=0; m<=n; m++) {
hij[l] += wv * Jx[n] * Jx[m];
l++;
}
}
for (int n=0; n<6; n++) {
vi[n] += wv * rv * Ji[n];
vj[n] += wv * rv * Jj[n];
Eii[block_id][n][k] += wv * Jz * Ji[n];
Eij[block_id][n][k] += wv * Jz * Jj[n];
}
}
__syncthreads();
__shared__ float sdata[THREADS];
for (int n=0; n<6; n++) {
sdata[threadIdx.x] = vi[n];
blockReduce(sdata);
if (threadIdx.x == 0) {
vs[0][block_id][n] = sdata[0];
}
__syncthreads();
sdata[threadIdx.x] = vj[n];
blockReduce(sdata);
if (threadIdx.x == 0) {
vs[1][block_id][n] = sdata[0];
}
}
l=0;
for (int n=0; n<12; n++) {
for (int m=0; m<=n; m++) {
sdata[threadIdx.x] = hij[l];
blockReduce(sdata);
if (threadIdx.x == 0) {
if (n<6 && m<6) {
Hs[0][block_id][n][m] = sdata[0];
Hs[0][block_id][m][n] = sdata[0];
}
else if (n >=6 && m<6) {
Hs[1][block_id][m][n-6] = sdata[0];
Hs[2][block_id][n-6][m] = sdata[0];
}
else {
Hs[3][block_id][n-6][m-6] = sdata[0];
Hs[3][block_id][m-6][n-6] = sdata[0];
}
}
l++;
}
}
}
__global__ void projmap_kernel(
const torch::PackedTensorAccessor32 poses,
const torch::PackedTensorAccessor32 disps,
const torch::PackedTensorAccessor32 intrinsics,
const torch::PackedTensorAccessor32 ii,
const torch::PackedTensorAccessor32 jj,
torch::PackedTensorAccessor32 coords,
torch::PackedTensorAccessor32 valid)
{
const int block_id = blockIdx.x;
const int thread_id = threadIdx.x;
const int ht = disps.size(1);
const int wd = disps.size(2);
__shared__ int ix;
__shared__ int jx;
__shared__ float fx;
__shared__ float fy;
__shared__ float cx;
__shared__ float cy;
__shared__ float ti[3], tj[3], tij[3];
__shared__ float qi[4], qj[4], qij[4];
// load intrinsics from global memory
if (thread_id == 0) {
ix = static_cast(ii[block_id]);
jx = static_cast(jj[block_id]);
fx = intrinsics[0];
fy = intrinsics[1];
cx = intrinsics[2];
cy = intrinsics[3];
}
__syncthreads();
// load poses from global memory
if (thread_id < 3) {
ti[thread_id] = poses[ix][thread_id];
tj[thread_id] = poses[jx][thread_id];
}
if (thread_id < 4) {
qi[thread_id] = poses[ix][thread_id+3];
qj[thread_id] = poses[jx][thread_id+3];
}
__syncthreads();
if (thread_id == 0) {
relSE3(ti, qi, tj, qj, tij, qij);
}
//points
float Xi[4];
float Xj[4];
__syncthreads();
GPU_1D_KERNEL_LOOP(k, ht*wd) {
const int i = k / wd;
const int j = k % wd;
const float u = static_cast(j);
const float v = static_cast(i);
// homogenous coordinates
Xi[0] = (u - cx) / fx;
Xi[1] = (v - cy) / fy;
Xi[2] = 1;
Xi[3] = disps[ix][i][j];
// transform homogenous point
actSE3(tij, qij, Xi, Xj);
coords[block_id][i][j][0] = u;
coords[block_id][i][j][1] = v;
if (Xj[2] > 0.01) {
coords[block_id][i][j][0] = fx * (Xj[0] / Xj[2]) + cx;
coords[block_id][i][j][1] = fy * (Xj[1] / Xj[2]) + cy;
}
valid[block_id][i][j][0] = (Xj[2] > MIN_DEPTH) ? 1.0 : 0.0;
}
}
__global__ void frame_distance_kernel(
const torch::PackedTensorAccessor32 poses,
const torch::PackedTensorAccessor32 disps,
const torch::PackedTensorAccessor32 intrinsics,
const torch::PackedTensorAccessor32 ii,
const torch::PackedTensorAccessor32 jj,
torch::PackedTensorAccessor32 dist,
const float beta) {
const int block_id = blockIdx.x;
const int thread_id = threadIdx.x;
const int ht = disps.size(1);
const int wd = disps.size(2);
__shared__ int ix;
__shared__ int jx;
__shared__ float fx;
__shared__ float fy;
__shared__ float cx;
__shared__ float cy;
__shared__ float ti[3], tj[3], tij[3];
__shared__ float qi[4], qj[4], qij[4];
// load intrinsics from global memory
if (thread_id == 0) {
ix = static_cast(ii[block_id]);
jx = static_cast(jj[block_id]);
fx = intrinsics[0];
fy = intrinsics[1];
cx = intrinsics[2];
cy = intrinsics[3];
}
__syncthreads();
//points
float Xi[4];
float Xj[4];
__shared__ float accum[THREADS]; accum[thread_id] = 0;
__shared__ float valid[THREADS]; valid[thread_id] = 0;
__shared__ float total[THREADS]; total[thread_id] = 0;
__syncthreads();
for (int n=0; n<1; n++) {
if (thread_id < 3) {
ti[thread_id] = poses[ix][thread_id];
tj[thread_id] = poses[jx][thread_id];
}
if (thread_id < 4) {
qi[thread_id] = poses[ix][thread_id+3];
qj[thread_id] = poses[jx][thread_id+3];
}
__syncthreads();
relSE3(ti, qi, tj, qj, tij, qij);
float d, du, dv;
GPU_1D_KERNEL_LOOP(k, ht*wd) {
const int i = k / wd;
const int j = k % wd;
const float u = static_cast(j);
const float v = static_cast(i);
// if (disps[ix][i][j] < 0.01) {
// continue;
// }
// homogenous coordinates
Xi[0] = (u - cx) / fx;
Xi[1] = (v - cy) / fy;
Xi[2] = 1;
Xi[3] = disps[ix][i][j];
// transform homogenous point
actSE3(tij, qij, Xi, Xj);
du = fx * (Xj[0] / Xj[2]) + cx - u;
dv = fy * (Xj[1] / Xj[2]) + cy - v;
d = sqrtf(du*du + dv*dv);
total[threadIdx.x] += beta;
if (Xj[2] > MIN_DEPTH) {
accum[threadIdx.x] += beta * d;
valid[threadIdx.x] += beta;
}
Xi[0] = (u - cx) / fx;
Xi[1] = (v - cy) / fy;
Xi[2] = 1;
Xi[3] = disps[ix][i][j];
Xj[0] = Xi[0] + Xi[3] * tij[0];
Xj[1] = Xi[1] + Xi[3] * tij[1];
Xj[2] = Xi[2] + Xi[3] * tij[2];
du = fx * (Xj[0] / Xj[2]) + cx - u;
dv = fy * (Xj[1] / Xj[2]) + cy - v;
d = sqrtf(du*du + dv*dv);
total[threadIdx.x] += (1 - beta);
if (Xj[2] > MIN_DEPTH) {
accum[threadIdx.x] += (1 - beta) * d;
valid[threadIdx.x] += (1 - beta);
}
}
if (threadIdx.x == 0) {
int tmp = ix;
ix = jx;
jx = tmp;
}
__syncthreads();
}
__syncthreads(); blockReduce(accum);
__syncthreads(); blockReduce(total);
__syncthreads(); blockReduce(valid);
__syncthreads();
if (thread_id == 0) {
// dist[block_id] = (valid[0] / (total[0] + 1e-8) < 0.75) ? 1000.0 : accum[0] / valid[0];
dist[block_id] = (valid[0] / (total[0] + 1e-8) < 0.75) ? 1000.0 : accum[0] / valid[0];
}
}
__global__ void depth_filter_kernel(
const torch::PackedTensorAccessor32 poses,
const torch::PackedTensorAccessor32 disps,
const torch::PackedTensorAccessor32 intrinsics,
const torch::PackedTensorAccessor32 inds,
const torch::PackedTensorAccessor32 thresh,
torch::PackedTensorAccessor32 counter)
{
const int block_id = blockIdx.x;
const int neigh_id = blockIdx.y;
const int index = blockIdx.z * blockDim.x + threadIdx.x;
// if (threadIdx.x == 0) {
// printf("%d %d %d %d\n", blockIdx.x, blockIdx.y, blockDim.x, threadIdx.x);
// }
const int num = disps.size(0);
const int ht = disps.size(1);
const int wd = disps.size(2);
__shared__ int ix;
__shared__ int jx;
__shared__ float fx;
__shared__ float fy;
__shared__ float cx;
__shared__ float cy;
__shared__ float ti[3], tj[3], tij[3];
__shared__ float qi[4], qj[4], qij[4];
if (threadIdx.x == 0) {
ix = static_cast(inds[block_id]);
jx = (neigh_id < 3) ? ix - neigh_id - 1 : ix + neigh_id;
fx = intrinsics[0];
fy = intrinsics[1];
cx = intrinsics[2];
cy = intrinsics[3];
}
__syncthreads();
if (jx < 0 || jx >= num) {
return;
}
const float t = thresh[block_id];
// load poses from global memory
if (threadIdx.x < 3) {
ti[threadIdx.x] = poses[ix][threadIdx.x];
tj[threadIdx.x] = poses[jx][threadIdx.x];
}
if (threadIdx.x < 4) {
qi[threadIdx.x] = poses[ix][threadIdx.x+3];
qj[threadIdx.x] = poses[jx][threadIdx.x+3];
}
__syncthreads();
if (threadIdx.x == 0) {
relSE3(ti, qi, tj, qj, tij, qij);
}
//points
float Xi[4];
float Xj[4];
__syncthreads();
if (index < ht*wd) {
const int i = index / wd;
const int j = index % wd;
const float ui = static_cast(j);
const float vi = static_cast(i);
const float di = disps[ix][i][j];
// homogenous coordinates
Xi[0] = (ui - cx) / fx;
Xi[1] = (vi - cy) / fy;
Xi[2] = 1;
Xi[3] = di;
// transform homogenous point
actSE3(tij, qij, Xi, Xj);
const float uj = fx * (Xj[0] / Xj[2]) + cx;
const float vj = fy * (Xj[1] / Xj[2]) + cy;
const float dj = Xj[3] / Xj[2];
const int u0 = static_cast(floor(uj));
const int v0 = static_cast(floor(vj));
if (u0 >= 0 && v0 >= 0 && u0 < wd-1 && v0 < ht-1) {
const float wx = ceil(uj) - uj;
const float wy = ceil(vj) - vj;
const float d00 = disps[jx][v0+0][u0+0];
const float d01 = disps[jx][v0+0][u0+1];
const float d10 = disps[jx][v0+1][u0+0];
const float d11 = disps[jx][v0+1][u0+1];
const float dj_hat = wy*wx*d00 + wy*(1-wx)*d01 + (1-wy)*wx*d10 + (1-wy)*(1-wx)*d11;
const float err = abs(1.0/dj - 1.0/dj_hat);
if (abs(1.0/dj - 1.0/d00) < t) atomicAdd(&counter[block_id][i][j], 1.0f);
else if (abs(1.0/dj - 1.0/d01) < t) atomicAdd(&counter[block_id][i][j], 1.0f);
else if (abs(1.0/dj - 1.0/d10) < t) atomicAdd(&counter[block_id][i][j], 1.0f);
else if (abs(1.0/dj - 1.0/d11) < t) atomicAdd(&counter[block_id][i][j], 1.0f);
}
}
}
__global__ void iproj_kernel(
const torch::PackedTensorAccessor32 poses,
const torch::PackedTensorAccessor32 disps,
const torch::PackedTensorAccessor32 intrinsics,
torch::PackedTensorAccessor32 points)
{
const int block_id = blockIdx.x;
const int index = blockIdx.y * blockDim.x + threadIdx.x;
const int num = disps.size(0);
const int ht = disps.size(1);
const int wd = disps.size(2);
__shared__ float fx;
__shared__ float fy;
__shared__ float cx;
__shared__ float cy;
__shared__ float t[3];
__shared__ float q[4];
if (threadIdx.x == 0) {
fx = intrinsics[0];
fy = intrinsics[1];
cx = intrinsics[2];
cy = intrinsics[3];
}
__syncthreads();
// load poses from global memory
if (threadIdx.x < 3) {
t[threadIdx.x] = poses[block_id][threadIdx.x];
}
if (threadIdx.x < 4) {
q[threadIdx.x] = poses[block_id][threadIdx.x+3];
}
__syncthreads();
//points
float Xi[4];
float Xj[4];
if (index < ht*wd) {
const int i = index / wd;
const int j = index % wd;
const float ui = static_cast(j);
const float vi = static_cast(i);
const float di = disps[block_id][i][j];
// homogenous coordinates
Xi[0] = (ui - cx) / fx;
Xi[1] = (vi - cy) / fy;
Xi[2] = 1;
Xi[3] = di;
// transform homogenous point
actSE3(t, q, Xi, Xj);
points[block_id][i][j][0] = Xj[0] / Xj[3];
points[block_id][i][j][1] = Xj[1] / Xj[3];
points[block_id][i][j][2] = Xj[2] / Xj[3];
}
}
__global__ void accum_kernel(
const torch::PackedTensorAccessor32 inps,
const torch::PackedTensorAccessor32 ptrs,
const torch::PackedTensorAccessor32 idxs,
torch::PackedTensorAccessor32 outs)
{
const int block_id = blockIdx.x;
const int D = inps.size(2);
const int start = ptrs[block_id];
const int end = ptrs[block_id+1];
for (int k=threadIdx.x; k poses,
const torch::PackedTensorAccessor32 dx,
const int t0, const int t1)
{
for (int k=t0+threadIdx.x; k disps,
const torch::PackedTensorAccessor32 dz,
const torch::PackedTensorAccessor32 inds)
{
const int i = inds[blockIdx.x];
const int ht = disps.size(1);
const int wd = disps.size(2);
for (int k=threadIdx.x; k();
long* jx_data = jx_cpu.data_ptr();
long* kx_data = inds.data_ptr();
int count = jx.size(0);
std::vector cols;
torch::Tensor ptrs_cpu = torch::zeros({count+1},
torch::TensorOptions().dtype(torch::kInt64));
long* ptrs_data = ptrs_cpu.data_ptr();
ptrs_data[0] = 0;
int i = 0;
for (int j=0; j();
for (int i=0; i>>(
data.packed_accessor32(),
ptrs.packed_accessor32(),
idxs.packed_accessor32(),
out.packed_accessor32());
return out;
}
__global__ void EEt6x6_kernel(
const torch::PackedTensorAccessor32 E,
const torch::PackedTensorAccessor32 Q,
const torch::PackedTensorAccessor32 idx,
torch::PackedTensorAccessor32 S)
{
// indicices
const int ix = idx[blockIdx.x][0];
const int jx = idx[blockIdx.x][1];
const int kx = idx[blockIdx.x][2];
const int D = E.size(2);
float dS[6][6];
float ei[6];
float ej[6];
for (int i=0; i<6; i++) {
for (int j=0; j<6; j++) {
dS[i][j] = 0;
}
}
for (int k=threadIdx.x; k E,
const torch::PackedTensorAccessor32 Q,
const torch::PackedTensorAccessor32 w,
const torch::PackedTensorAccessor32 idx,
torch::PackedTensorAccessor32 v)
{
const int D = E.size(2);
const int kx = idx[blockIdx.x][0];
float b[6];
for (int n=0; n<6; n++) {
b[n] = 0.0;
}
for (int k=threadIdx.x; k E,
const torch::PackedTensorAccessor32 x,
const torch::PackedTensorAccessor32 idx,
torch::PackedTensorAccessor32 w)
{
const int D = E.size(2);
const int ix = idx[blockIdx.x];
if (idx[blockIdx.x] <= 0 || idx[blockIdx.x] >= x.size(0))
return;
for (int k=threadIdx.x; k A;
Eigen::VectorX b;
SparseBlock(int N, int M) : N(N), M(M) {
A = Eigen::SparseMatrix(N*M, N*M);
b = Eigen::VectorXd::Zero(N*M);
}
SparseBlock(Eigen::SparseMatrix const& A, Eigen::VectorX const& b,
int N, int M) : A(A), b(b), N(N), M(M) {}
void update_lhs(torch::Tensor As, torch::Tensor ii, torch::Tensor jj) {
auto As_cpu = As.to(torch::kCPU).to(torch::kFloat64);
auto ii_cpu = ii.to(torch::kCPU).to(torch::kInt64);
auto jj_cpu = jj.to(torch::kCPU).to(torch::kInt64);
auto As_acc = As_cpu.accessor();
auto ii_acc = ii_cpu.accessor();
auto jj_acc = jj_cpu.accessor();
std::vector tripletList;
for (int n=0; n= 0 && j >= 0) {
for (int k=0; k();
auto ii_acc = ii_cpu.accessor();
for (int n=0; n= 0) {
for (int j=0; j get_dense() {
Eigen::MatrixXd Ad = Eigen::MatrixXd(A);
torch::Tensor H = torch::from_blob(Ad.data(), {N*M, N*M}, torch::TensorOptions()
.dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);
torch::Tensor v = torch::from_blob(b.data(), {N*M, 1}, torch::TensorOptions()
.dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);
return std::make_tuple(H, v);
}
void get_dense_extend(torch::PackedTensorAccessor32 H,
torch::PackedTensorAccessor32 v)
{
// Eigen::MatrixXd Ad = Eigen::MatrixXd(A);
// std::cerr< L(A);
L.diagonal().array() += ep + lm * L.diagonal().array();
Eigen::SimplicialLLT> solver;
solver.compute(L);
if (solver.info() == Eigen::Success) {
Eigen::VectorXd x = solver.solve(b);
dx = torch::from_blob(x.data(), {N, M}, torch::TensorOptions()
.dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);
}
else {
dx = torch::zeros({N, M}, torch::TensorOptions()
.device(torch::kCUDA).dtype(torch::kFloat32));
}
return dx;
}
torch::Tensor solve_dense(const float lm=0.0001,const float ep = 0.1){
torch::Tensor dx;
Eigen::MatrixXd L(A);
L.diagonal().array() += ep + lm * L.diagonal().array();
Eigen::LLT solver;
solver.compute(L);
if (solver.info() == Eigen::Success) {
Eigen::VectorXd x = solver.solve(b);
dx = torch::from_blob(x.data(), {N, M}, torch::TensorOptions()
.dtype(torch::kFloat64)).to(torch::kCUDA).to(torch::kFloat32);
}
else {
dx = torch::zeros({N, M}, torch::TensorOptions()
.device(torch::kCUDA).dtype(torch::kFloat32));
}
return dx;
}
private:
const int N;
const int M;
};
SparseBlock schur_block(torch::Tensor E,
torch::Tensor Q,
torch::Tensor w,
torch::Tensor ii,
torch::Tensor jj,
torch::Tensor kk,
const int t0,
const int t1)
{
torch::Tensor ii_cpu = ii.to(torch::kCPU);
torch::Tensor jj_cpu = jj.to(torch::kCPU);
torch::Tensor kk_cpu = kk.to(torch::kCPU);
const int P = t1 - t0;
const long* ii_data = ii_cpu.data_ptr();
const long* jj_data = jj_cpu.data_ptr();
const long* kk_data = kk_cpu.data_ptr();
std::vector> graph(P);
std::vector> index(P);
for (int n=0; n= t0 && j <= t1) {
const int t = j - t0;
graph[t].push_back(k);
index[t].push_back(n);
}
}
std::vector ii_list, jj_list, idx, jdx;
for (int i=0; i>>(
E.packed_accessor32(),
Q.packed_accessor32(),
ix_cuda.packed_accessor32(),
S.packed_accessor32