Showing preview only (203K chars total). Download the full file or copy to clipboard to get everything.
Repository: graphdeco-inria/gaussian-splatting
Branch: main
Commit: 54c035f7834b
Files: 32
Total size: 193.2 KB
Directory structure:
gitextract_uhkue7yy/
├── .gitignore
├── .gitmodules
├── LICENSE.md
├── README.md
├── arguments/
│ └── __init__.py
├── convert.py
├── environment.yml
├── full_eval.py
├── gaussian_renderer/
│ ├── __init__.py
│ └── network_gui.py
├── lpipsPyTorch/
│ ├── __init__.py
│ └── modules/
│ ├── lpips.py
│ ├── networks.py
│ └── utils.py
├── metrics.py
├── render.py
├── results.md
├── scene/
│ ├── __init__.py
│ ├── cameras.py
│ ├── colmap_loader.py
│ ├── dataset_readers.py
│ └── gaussian_model.py
├── train.py
└── utils/
├── camera_utils.py
├── general_utils.py
├── graphics_utils.py
├── image_utils.py
├── loss_utils.py
├── make_depth_scale.py
├── read_write_model.py
├── sh_utils.py
└── system_utils.py
================================================
FILE CONTENTS
================================================
================================================
FILE: .gitignore
================================================
*.pyc
.vscode
output
build
diff_rasterization/diff_rast.egg-info
diff_rasterization/dist
tensorboard_3d
screenshots
================================================
FILE: .gitmodules
================================================
[submodule "submodules/simple-knn"]
path = submodules/simple-knn
url = https://gitlab.inria.fr/bkerbl/simple-knn.git
[submodule "submodules/diff-gaussian-rasterization"]
path = submodules/diff-gaussian-rasterization
url = https://github.com/graphdeco-inria/diff-gaussian-rasterization.git
branch = dr_aa
[submodule "SIBR_viewers"]
path = SIBR_viewers
url = https://gitlab.inria.fr/sibr/sibr_core.git
[submodule "submodules/fused-ssim"]
path = submodules/fused-ssim
url = https://github.com/rahul-goel/fused-ssim.git
================================================
FILE: LICENSE.md
================================================
Gaussian-Splatting License
===========================
**Inria** and **the Max Planck Institut for Informatik (MPII)** hold all the ownership rights on the *Software* named **gaussian-splatting**.
The *Software* is in the process of being registered with the Agence pour la Protection des
Programmes (APP).
The *Software* is still being developed by the *Licensor*.
*Licensor*'s goal is to allow the research community to use, test and evaluate
the *Software*.
## 1. Definitions
*Licensee* means any person or entity that uses the *Software* and distributes
its *Work*.
*Licensor* means the owners of the *Software*, i.e Inria and MPII
*Software* means the original work of authorship made available under this
License ie gaussian-splatting.
*Work* means the *Software* and any additions to or derivative works of the
*Software* that are made available under this License.
## 2. Purpose
This license is intended to define the rights granted to the *Licensee* by
Licensors under the *Software*.
## 3. Rights granted
For the above reasons Licensors have decided to distribute the *Software*.
Licensors grant non-exclusive rights to use the *Software* for research purposes
to research users (both academic and industrial), free of charge, without right
to sublicense.. The *Software* may be used "non-commercially", i.e., for research
and/or evaluation purposes only.
Subject to the terms and conditions of this License, you are granted a
non-exclusive, royalty-free, license to reproduce, prepare derivative works of,
publicly display, publicly perform and distribute its *Work* and any resulting
derivative works in any form.
## 4. Limitations
**4.1 Redistribution.** You may reproduce or distribute the *Work* only if (a) you do
so under this License, (b) you include a complete copy of this License with
your distribution, and (c) you retain without modification any copyright,
patent, trademark, or attribution notices that are present in the *Work*.
**4.2 Derivative Works.** You may specify that additional or different terms apply
to the use, reproduction, and distribution of your derivative works of the *Work*
("Your Terms") only if (a) Your Terms provide that the use limitation in
Section 2 applies to your derivative works, and (b) you identify the specific
derivative works that are subject to Your Terms. Notwithstanding Your Terms,
this License (including the redistribution requirements in Section 3.1) will
continue to apply to the *Work* itself.
**4.3** Any other use without of prior consent of Licensors is prohibited. Research
users explicitly acknowledge having received from Licensors all information
allowing to appreciate the adequacy between of the *Software* and their needs and
to undertake all necessary precautions for its execution and use.
**4.4** The *Software* is provided both as a compiled library file and as source
code. In case of using the *Software* for a publication or other results obtained
through the use of the *Software*, users are strongly encouraged to cite the
corresponding publications as explained in the documentation of the *Software*.
## 5. Disclaimer
THE USER CANNOT USE, EXPLOIT OR DISTRIBUTE THE *SOFTWARE* FOR COMMERCIAL PURPOSES
WITHOUT PRIOR AND EXPLICIT CONSENT OF LICENSORS. YOU MUST CONTACT INRIA FOR ANY
UNAUTHORIZED USE: stip-sophia.transfert@inria.fr . ANY SUCH ACTION WILL
CONSTITUTE A FORGERY. THIS *SOFTWARE* IS PROVIDED "AS IS" WITHOUT ANY WARRANTIES
OF ANY NATURE AND ANY EXPRESS OR IMPLIED WARRANTIES, WITH REGARDS TO COMMERCIAL
USE, PROFESSIONNAL USE, LEGAL OR NOT, OR OTHER, OR COMMERCIALISATION OR
ADAPTATION. UNLESS EXPLICITLY PROVIDED BY LAW, IN NO EVENT, SHALL INRIA OR THE
AUTHOR BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE
GOODS OR SERVICES, LOSS OF USE, DATA, OR PROFITS OR BUSINESS INTERRUPTION)
HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING FROM, OUT OF OR
IN CONNECTION WITH THE *SOFTWARE* OR THE USE OR OTHER DEALINGS IN THE *SOFTWARE*.
## 6. Files subject to permissive licenses
The contents of the file ```utils/loss_utils.py``` are based on publicly available code authored by Evan Su, which falls under the permissive MIT license.
Title: pytorch-ssim\
Project code: https://github.com/Po-Hsun-Su/pytorch-ssim\
Copyright Evan Su, 2017\
License: https://github.com/Po-Hsun-Su/pytorch-ssim/blob/master/LICENSE.txt (MIT)
================================================
FILE: README.md
================================================
# 3D Gaussian Splatting for Real-Time Radiance Field Rendering
Bernhard Kerbl*, Georgios Kopanas*, Thomas Leimkühler, George Drettakis (* indicates equal contribution)<br>
| [Webpage](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/) | [Full Paper](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/3d_gaussian_splatting_high.pdf) | [Video](https://youtu.be/T_kXY43VZnk) | [Other GRAPHDECO Publications](http://www-sop.inria.fr/reves/publis/gdindex.php) | [FUNGRAPH project page](https://fungraph.inria.fr) |<br>
| [T&T+DB COLMAP (650MB)](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/datasets/input/tandt_db.zip) | [Pre-trained Models (14 GB)](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/datasets/pretrained/models.zip) | [Viewers for Windows (60MB)](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/binaries/viewers.zip) | [Evaluation Images (7 GB)](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/evaluation/images.zip) |<br>
This repository contains the official authors implementation associated with the paper "3D Gaussian Splatting for Real-Time Radiance Field Rendering", which can be found [here](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/). We further provide the reference images used to create the error metrics reported in the paper, as well as recently created, pre-trained models.
<a href="https://www.inria.fr/"><img height="100" src="assets/logo_inria.png"> </a>
<a href="https://univ-cotedazur.eu/"><img height="100" src="assets/logo_uca.png"> </a>
<a href="https://www.mpi-inf.mpg.de"><img height="100" src="assets/logo_mpi.png"> </a>
<a href="https://team.inria.fr/graphdeco/"> <img style="width:100%;" src="assets/logo_graphdeco.png"></a>
Abstract: *Radiance Field methods have recently revolutionized novel-view synthesis of scenes captured with multiple photos or videos. However, achieving high visual quality still requires neural networks that are costly to train and render, while recent faster methods inevitably trade off speed for quality. For unbounded and complete scenes (rather than isolated objects) and 1080p resolution rendering, no current method can achieve real-time display rates. We introduce three key elements that allow us to achieve state-of-the-art visual quality while maintaining competitive training times and importantly allow high-quality real-time (≥ 30 fps) novel-view synthesis at 1080p resolution. First, starting from sparse points produced during camera calibration, we represent the scene with 3D Gaussians that preserve desirable properties of continuous volumetric radiance fields for scene optimization while avoiding unnecessary computation in empty space; Second, we perform interleaved optimization/density control of the 3D Gaussians, notably optimizing anisotropic covariance to achieve an accurate representation of the scene; Third, we develop a fast visibility-aware rendering algorithm that supports anisotropic splatting and both accelerates training and allows realtime rendering. We demonstrate state-of-the-art visual quality and real-time rendering on several established datasets.*
<section class="section" id="BibTeX">
<div class="container is-max-desktop content">
<h2 class="title">BibTeX</h2>
<pre><code>@Article{kerbl3Dgaussians,
author = {Kerbl, Bernhard and Kopanas, Georgios and Leimk{\"u}hler, Thomas and Drettakis, George},
title = {3D Gaussian Splatting for Real-Time Radiance Field Rendering},
journal = {ACM Transactions on Graphics},
number = {4},
volume = {42},
month = {July},
year = {2023},
url = {https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/}
}</code></pre>
</div>
</section>
## Funding and Acknowledgments
This research was funded by the ERC Advanced grant FUNGRAPH No 788065. The authors are grateful to Adobe for generous donations, the OPAL infrastructure from Université Côte d’Azur and for the HPC resources from GENCI–IDRIS (Grant 2022-AD011013409). The authors thank the anonymous reviewers for their valuable feedback, P. Hedman and A. Tewari for proofreading earlier drafts also T. Müller, A. Yu and S. Fridovich-Keil for helping with the comparisons.
## NEW FEATURES !
We have limited resources for maintaining and updating the code. However, we have added a few new features since the original release that are inspired by some of the excellent work many other researchers have been doing on 3DGS. We will be adding other features within the ability of our resources.
**Update of October 2024**: We integrated [training speed acceleration](#training-speed-acceleration) and made it compatible with [depth regularization](#depth-regularization), [anti-aliasing](#anti-aliasing) and [exposure compensation](#exposure-compensation). We have enhanced the SIBR real time viewer by correcting bugs and adding features in the [Top View](#sibr-top-view) that allows visualization of input and user cameras.
**Update of Spring 2024**:
Orange Labs has kindly added [OpenXR support](#openxr-support) for VR viewing.
## Step-by-step Tutorial
Jonathan Stephens made a fantastic step-by-step tutorial for setting up Gaussian Splatting on your machine, along with instructions for creating usable datasets from videos. If the instructions below are too dry for you, go ahead and check it out [here](https://www.youtube.com/watch?v=UXtuigy_wYc).
## Colab
User [camenduru](https://github.com/camenduru) was kind enough to provide a Colab template that uses this repo's source (status: August 2023!) for quick and easy access to the method. Please check it out [here](https://github.com/camenduru/gaussian-splatting-colab).
## Cloning the Repository
The repository contains submodules, thus please check it out with
```shell
# SSH
git clone git@github.com:graphdeco-inria/gaussian-splatting.git --recursive
```
or
```shell
# HTTPS
git clone https://github.com/graphdeco-inria/gaussian-splatting --recursive
```
## Overview
The codebase has 4 main components:
- A PyTorch-based optimizer to produce a 3D Gaussian model from SfM inputs
- A network viewer that allows to connect to and visualize the optimization process
- An OpenGL-based real-time viewer to render trained models in real-time.
- A script to help you turn your own images into optimization-ready SfM data sets
The components have different requirements w.r.t. both hardware and software. They have been tested on Windows 10 and Ubuntu Linux 22.04. Instructions for setting up and running each of them are found in the sections below.
## Optimizer
The optimizer uses PyTorch and CUDA extensions in a Python environment to produce trained models.
### Hardware Requirements
- CUDA-ready GPU with Compute Capability 7.0+
- 24 GB VRAM (to train to paper evaluation quality)
- Please see FAQ for smaller VRAM configurations
### Software Requirements
- Conda (recommended for easy setup)
- C++ Compiler for PyTorch extensions (we used Visual Studio 2019 for Windows)
- CUDA SDK 11 for PyTorch extensions, install *after* Visual Studio (we used 11.8, **known issues with 11.6**)
- C++ Compiler and CUDA SDK must be compatible
### Setup
#### Local Setup
Our default, provided install method is based on Conda package and environment management:
```shell
SET DISTUTILS_USE_SDK=1 # Windows only
conda env create --file environment.yml
conda activate gaussian_splatting
```
Please note that this process assumes that you have CUDA SDK **11** installed, not **12**. For modifications, see below.
Tip: Downloading packages and creating a new environment with Conda can require a significant amount of disk space. By default, Conda will use the main system hard drive. You can avoid this by specifying a different package download location and an environment on a different drive:
```shell
conda config --add pkgs_dirs <Drive>/<pkg_path>
conda env create --file environment.yml --prefix <Drive>/<env_path>/gaussian_splatting
conda activate <Drive>/<env_path>/gaussian_splatting
```
#### Modifications
If you can afford the disk space, we recommend using our environment files for setting up a training environment identical to ours. If you want to make modifications, please note that major version changes might affect the results of our method. However, our (limited) experiments suggest that the codebase works just fine inside a more up-to-date environment (Python 3.8, PyTorch 2.0.0, CUDA 12). Make sure to create an environment where PyTorch and its CUDA runtime version match and the installed CUDA SDK has no major version difference with PyTorch's CUDA version.
#### Known Issues
Some users experience problems building the submodules on Windows (```cl.exe: File not found``` or similar). Please consider the workaround for this problem from the FAQ.
### Running
To run the optimizer, simply use
```shell
python train.py -s <path to COLMAP or NeRF Synthetic dataset>
```
<details>
<summary><span style="font-weight: bold;">Command Line Arguments for train.py</span></summary>
#### --source_path / -s
Path to the source directory containing a COLMAP or Synthetic NeRF data set.
#### --model_path / -m
Path where the trained model should be stored (```output/<random>``` by default).
#### --images / -i
Alternative subdirectory for COLMAP images (```images``` by default).
#### --eval
Add this flag to use a MipNeRF360-style training/test split for evaluation.
#### --resolution / -r
Specifies resolution of the loaded images before training. If provided ```1, 2, 4``` or ```8```, uses original, 1/2, 1/4 or 1/8 resolution, respectively. For all other values, rescales the width to the given number while maintaining image aspect. **If not set and input image width exceeds 1.6K pixels, inputs are automatically rescaled to this target.**
#### --data_device
Specifies where to put the source image data, ```cuda``` by default, recommended to use ```cpu``` if training on large/high-resolution dataset, will reduce VRAM consumption, but slightly slow down training. Thanks to [HrsPythonix](https://github.com/HrsPythonix).
#### --white_background / -w
Add this flag to use white background instead of black (default), e.g., for evaluation of NeRF Synthetic dataset.
#### --sh_degree
Order of spherical harmonics to be used (no larger than 3). ```3``` by default.
#### --convert_SHs_python
Flag to make pipeline compute forward and backward of SHs with PyTorch instead of ours.
#### --convert_cov3D_python
Flag to make pipeline compute forward and backward of the 3D covariance with PyTorch instead of ours.
#### --debug
Enables debug mode if you experience erros. If the rasterizer fails, a ```dump``` file is created that you may forward to us in an issue so we can take a look.
#### --debug_from
Debugging is **slow**. You may specify an iteration (starting from 0) after which the above debugging becomes active.
#### --iterations
Number of total iterations to train for, ```30_000``` by default.
#### --ip
IP to start GUI server on, ```127.0.0.1``` by default.
#### --port
Port to use for GUI server, ```6009``` by default.
#### --test_iterations
Space-separated iterations at which the training script computes L1 and PSNR over test set, ```7000 30000``` by default.
#### --save_iterations
Space-separated iterations at which the training script saves the Gaussian model, ```7000 30000 <iterations>``` by default.
#### --checkpoint_iterations
Space-separated iterations at which to store a checkpoint for continuing later, saved in the model directory.
#### --start_checkpoint
Path to a saved checkpoint to continue training from.
#### --quiet
Flag to omit any text written to standard out pipe.
#### --feature_lr
Spherical harmonics features learning rate, ```0.0025``` by default.
#### --opacity_lr
Opacity learning rate, ```0.05``` by default.
#### --scaling_lr
Scaling learning rate, ```0.005``` by default.
#### --rotation_lr
Rotation learning rate, ```0.001``` by default.
#### --position_lr_max_steps
Number of steps (from 0) where position learning rate goes from ```initial``` to ```final```. ```30_000``` by default.
#### --position_lr_init
Initial 3D position learning rate, ```0.00016``` by default.
#### --position_lr_final
Final 3D position learning rate, ```0.0000016``` by default.
#### --position_lr_delay_mult
Position learning rate multiplier (cf. Plenoxels), ```0.01``` by default.
#### --densify_from_iter
Iteration where densification starts, ```500``` by default.
#### --densify_until_iter
Iteration where densification stops, ```15_000``` by default.
#### --densify_grad_threshold
Limit that decides if points should be densified based on 2D position gradient, ```0.0002``` by default.
#### --densification_interval
How frequently to densify, ```100``` (every 100 iterations) by default.
#### --opacity_reset_interval
How frequently to reset opacity, ```3_000``` by default.
#### --lambda_dssim
Influence of SSIM on total loss from 0 to 1, ```0.2``` by default.
#### --percent_dense
Percentage of scene extent (0--1) a point must exceed to be forcibly densified, ```0.01``` by default.
</details>
<br>
Note that similar to MipNeRF360, we target images at resolutions in the 1-1.6K pixel range. For convenience, arbitrary-size inputs can be passed and will be automatically resized if their width exceeds 1600 pixels. We recommend to keep this behavior, but you may force training to use your higher-resolution images by setting ```-r 1```.
The MipNeRF360 scenes are hosted by the paper authors [here](https://jonbarron.info/mipnerf360/). You can find our SfM data sets for Tanks&Temples and Deep Blending [here](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/datasets/input/tandt_db.zip). If you do not provide an output model directory (```-m```), trained models are written to folders with randomized unique names inside the ```output``` directory. At this point, the trained models may be viewed with the real-time viewer (see further below).
### Evaluation
By default, the trained models use all available images in the dataset. To train them while withholding a test set for evaluation, use the ```--eval``` flag. This way, you can render training/test sets and produce error metrics as follows:
```shell
python train.py -s <path to COLMAP or NeRF Synthetic dataset> --eval # Train with train/test split
python render.py -m <path to trained model> # Generate renderings
python metrics.py -m <path to trained model> # Compute error metrics on renderings
```
If you want to evaluate our [pre-trained models](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/datasets/pretrained/models.zip), you will have to download the corresponding source data sets and indicate their location to ```render.py``` with an additional ```--source_path/-s``` flag. Note: The pre-trained models were created with the release codebase. This code base has been cleaned up and includes bugfixes, hence the metrics you get from evaluating them will differ from those in the paper.
```shell
python render.py -m <path to pre-trained model> -s <path to COLMAP dataset>
python metrics.py -m <path to pre-trained model>
```
<details>
<summary><span style="font-weight: bold;">Command Line Arguments for render.py</span></summary>
#### --model_path / -m
Path to the trained model directory you want to create renderings for.
#### --skip_train
Flag to skip rendering the training set.
#### --skip_test
Flag to skip rendering the test set.
#### --quiet
Flag to omit any text written to standard out pipe.
**The below parameters will be read automatically from the model path, based on what was used for training. However, you may override them by providing them explicitly on the command line.**
#### --source_path / -s
Path to the source directory containing a COLMAP or Synthetic NeRF data set.
#### --images / -i
Alternative subdirectory for COLMAP images (```images``` by default).
#### --eval
Add this flag to use a MipNeRF360-style training/test split for evaluation.
#### --resolution / -r
Changes the resolution of the loaded images before training. If provided ```1, 2, 4``` or ```8```, uses original, 1/2, 1/4 or 1/8 resolution, respectively. For all other values, rescales the width to the given number while maintaining image aspect. ```1``` by default.
#### --white_background / -w
Add this flag to use white background instead of black (default), e.g., for evaluation of NeRF Synthetic dataset.
#### --convert_SHs_python
Flag to make pipeline render with computed SHs from PyTorch instead of ours.
#### --convert_cov3D_python
Flag to make pipeline render with computed 3D covariance from PyTorch instead of ours.
</details>
<details>
<summary><span style="font-weight: bold;">Command Line Arguments for metrics.py</span></summary>
#### --model_paths / -m
Space-separated list of model paths for which metrics should be computed.
</details>
<br>
We further provide the ```full_eval.py``` script. This script specifies the routine used in our evaluation and demonstrates the use of some additional parameters, e.g., ```--images (-i)``` to define alternative image directories within COLMAP data sets. If you have downloaded and extracted all the training data, you can run it like this:
```shell
python full_eval.py -m360 <mipnerf360 folder> -tat <tanks and temples folder> -db <deep blending folder>
```
In the current version, this process takes about 7h on our reference machine containing an A6000. If you want to do the full evaluation on our pre-trained models, you can specify their download location and skip training.
```shell
python full_eval.py -o <directory with pretrained models> --skip_training -m360 <mipnerf360 folder> -tat <tanks and temples folder> -db <deep blending folder>
```
If you want to compute the metrics on our paper's [evaluation images](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/evaluation/images.zip), you can also skip rendering. In this case it is not necessary to provide the source datasets. You can compute metrics for multiple image sets at a time.
```shell
python full_eval.py -m <directory with evaluation images>/garden ... --skip_training --skip_rendering
```
<details>
<summary><span style="font-weight: bold;">Command Line Arguments for full_eval.py</span></summary>
#### --skip_training
Flag to skip training stage.
#### --skip_rendering
Flag to skip rendering stage.
#### --skip_metrics
Flag to skip metrics calculation stage.
#### --output_path
Directory to put renderings and results in, ```./eval``` by default, set to pre-trained model location if evaluating them.
#### --mipnerf360 / -m360
Path to MipNeRF360 source datasets, required if training or rendering.
#### --tanksandtemples / -tat
Path to Tanks&Temples source datasets, required if training or rendering.
#### --deepblending / -db
Path to Deep Blending source datasets, required if training or rendering.
</details>
<br>
## Interactive Viewers
We provide two interactive viewers for our method: remote and real-time. Our viewing solutions are based on the [SIBR](https://sibr.gitlabpages.inria.fr/) framework, developed by the GRAPHDECO group for several novel-view synthesis projects.
### Hardware Requirements
- OpenGL 4.5-ready GPU and drivers (or latest MESA software)
- 4 GB VRAM recommended
- CUDA-ready GPU with Compute Capability 7.0+ (only for Real-Time Viewer)
### Software Requirements
- Visual Studio or g++, **not Clang** (we used Visual Studio 2019 for Windows)
- CUDA SDK 11, install *after* Visual Studio (we used 11.8)
- CMake (recent version, we used 3.24)
- 7zip (only on Windows)
### Pre-built Windows Binaries
We provide pre-built binaries for Windows [here](https://repo-sam.inria.fr/fungraph/3d-gaussian-splatting/binaries/viewers.zip). We recommend using them on Windows for an efficient setup, since the building of SIBR involves several external dependencies that must be downloaded and compiled on-the-fly.
### Installation from Source
If you cloned with submodules (e.g., using ```--recursive```), the source code for the viewers is found in ```SIBR_viewers```. The network viewer runs within the SIBR framework for Image-based Rendering applications.
#### Windows
CMake should take care of your dependencies.
```shell
cd SIBR_viewers
cmake -Bbuild .
cmake --build build --target install --config RelWithDebInfo
```
You may specify a different configuration, e.g. ```Debug``` if you need more control during development.
#### Ubuntu 22.04
You will need to install a few dependencies before running the project setup.
```shell
# Dependencies
sudo apt install -y libglew-dev libassimp-dev libboost-all-dev libgtk-3-dev libopencv-dev libglfw3-dev libavdevice-dev libavcodec-dev libeigen3-dev libxxf86vm-dev libembree-dev
# Project setup
cd SIBR_viewers
cmake -Bbuild . -DCMAKE_BUILD_TYPE=Release # add -G Ninja to build faster
cmake --build build -j24 --target install
```
#### Ubuntu 20.04
Backwards compatibility with Focal Fossa is not fully tested, but building SIBR with CMake should still work after invoking
```shell
git checkout fossa_compatibility
```
### Navigation in SIBR Viewers
The SIBR interface provides several methods of navigating the scene. By default, you will be started with an FPS navigator, which you can control with ```W, A, S, D, Q, E``` for camera translation and ```I, K, J, L, U, O``` for rotation. Alternatively, you may want to use a Trackball-style navigator (select from the floating menu). You can also snap to a camera from the data set with the ```Snap to``` button or find the closest camera with ```Snap to closest```. The floating menues also allow you to change the navigation speed. You can use the ```Scaling Modifier``` to control the size of the displayed Gaussians, or show the initial point cloud.
### Running the Network Viewer
https://github.com/graphdeco-inria/gaussian-splatting/assets/40643808/90a2e4d3-cf2e-4633-b35f-bfe284e28ff7
After extracting or installing the viewers, you may run the compiled ```SIBR_remoteGaussian_app[_config]``` app in ```<SIBR install dir>/bin```, e.g.:
```shell
./<SIBR install dir>/bin/SIBR_remoteGaussian_app
```
The network viewer allows you to connect to a running training process on the same or a different machine. If you are training on the same machine and OS, no command line parameters should be required: the optimizer communicates the location of the training data to the network viewer. By default, optimizer and network viewer will try to establish a connection on **localhost** on port **6009**. You can change this behavior by providing matching ```--ip``` and ```--port``` parameters to both the optimizer and the network viewer. If for some reason the path used by the optimizer to find the training data is not reachable by the network viewer (e.g., due to them running on different (virtual) machines), you may specify an override location to the viewer by using ```-s <source path>```.
<details>
<summary><span style="font-weight: bold;">Primary Command Line Arguments for Network Viewer</span></summary>
#### --path / -s
Argument to override model's path to source dataset.
#### --ip
IP to use for connection to a running training script.
#### --port
Port to use for connection to a running training script.
#### --rendering-size
Takes two space separated numbers to define the resolution at which network rendering occurs, ```1200``` width by default.
Note that to enforce an aspect that differs from the input images, you need ```--force-aspect-ratio``` too.
#### --load_images
Flag to load source dataset images to be displayed in the top view for each camera.
</details>
<br>
### Running the Real-Time Viewer
https://github.com/graphdeco-inria/gaussian-splatting/assets/40643808/0940547f-1d82-4c2f-a616-44eabbf0f816
After extracting or installing the viewers, you may run the compiled ```SIBR_gaussianViewer_app[_config]``` app in ```<SIBR install dir>/bin```, e.g.:
```shell
./<SIBR install dir>/bin/SIBR_gaussianViewer_app -m <path to trained model>
```
It should suffice to provide the ```-m``` parameter pointing to a trained model directory. Alternatively, you can specify an override location for training input data using ```-s```. To use a specific resolution other than the auto-chosen one, specify ```--rendering-size <width> <height>```. Combine it with ```--force-aspect-ratio``` if you want the exact resolution and don't mind image distortion.
**To unlock the full frame rate, please disable V-Sync on your machine and also in the application (Menu → Display). In a multi-GPU system (e.g., laptop) your OpenGL/Display GPU should be the same as your CUDA GPU (e.g., by setting the application's GPU preference on Windows, see below) for maximum performance.**
In addition to the initial point cloud and the splats, you also have the option to visualize the Gaussians by rendering them as ellipsoids from the floating menu.
SIBR has many other functionalities, please see the [documentation](https://sibr.gitlabpages.inria.fr/) for more details on the viewer, navigation options etc. There is also a Top View (available from the menu) that shows the placement of the input cameras and the original SfM point cloud; please note that Top View slows rendering when enabled. The real-time viewer also uses slightly more aggressive, fast culling, which can be toggled in the floating menu. If you ever encounter an issue that can be solved by turning fast culling off, please let us know.
<details>
<summary><span style="font-weight: bold;">Primary Command Line Arguments for Real-Time Viewer</span></summary>
#### --model-path / -m
Path to trained model.
#### --iteration
Specifies which of state to load if multiple are available. Defaults to latest available iteration.
#### --path / -s
Argument to override model's path to source dataset.
#### --rendering-size
Takes two space separated numbers to define the resolution at which real-time rendering occurs, ```1200``` width by default. Note that to enforce an aspect that differs from the input images, you need ```--force-aspect-ratio``` too.
#### --load_images
Flag to load source dataset images to be displayed in the top view for each camera.
#### --device
Index of CUDA device to use for rasterization if multiple are available, ```0``` by default.
#### --no_interop
Disables CUDA/GL interop forcibly. Use on systems that may not behave according to spec (e.g., WSL2 with MESA GL 4.5 software rendering).
</details>
<br>
## Processing your own Scenes
Our COLMAP loaders expect the following dataset structure in the source path location:
```
<location>
|---images
| |---<image 0>
| |---<image 1>
| |---...
|---sparse
|---0
|---cameras.bin
|---images.bin
|---points3D.bin
```
For rasterization, the camera models must be either a SIMPLE_PINHOLE or PINHOLE camera. We provide a converter script ```convert.py```, to extract undistorted images and SfM information from input images. Optionally, you can use ImageMagick to resize the undistorted images. This rescaling is similar to MipNeRF360, i.e., it creates images with 1/2, 1/4 and 1/8 the original resolution in corresponding folders. To use them, please first install a recent version of COLMAP (ideally CUDA-powered) and ImageMagick. Put the images you want to use in a directory ```<location>/input```.
```
<location>
|---input
|---<image 0>
|---<image 1>
|---...
```
If you have COLMAP and ImageMagick on your system path, you can simply run
```shell
python convert.py -s <location> [--resize] #If not resizing, ImageMagick is not needed
```
Alternatively, you can use the optional parameters ```--colmap_executable``` and ```--magick_executable``` to point to the respective paths. Please note that on Windows, the executable should point to the COLMAP ```.bat``` file that takes care of setting the execution environment. Once done, ```<location>``` will contain the expected COLMAP data set structure with undistorted, resized input images, in addition to your original images and some temporary (distorted) data in the directory ```distorted```.
If you have your own COLMAP dataset without undistortion (e.g., using ```OPENCV``` camera), you can try to just run the last part of the script: Put the images in ```input``` and the COLMAP info in a subdirectory ```distorted```:
```
<location>
|---input
| |---<image 0>
| |---<image 1>
| |---...
|---distorted
|---database.db
|---sparse
|---0
|---...
```
Then run
```shell
python convert.py -s <location> --skip_matching [--resize] #If not resizing, ImageMagick is not needed
```
<details>
<summary><span style="font-weight: bold;">Command Line Arguments for convert.py</span></summary>
#### --no_gpu
Flag to avoid using GPU in COLMAP.
#### --skip_matching
Flag to indicate that COLMAP info is available for images.
#### --source_path / -s
Location of the inputs.
#### --camera
Which camera model to use for the early matching steps, ```OPENCV``` by default.
#### --resize
Flag for creating resized versions of input images.
#### --colmap_executable
Path to the COLMAP executable (```.bat``` on Windows).
#### --magick_executable
Path to the ImageMagick executable.
</details>
<br>
### Training speed acceleration
We integrated the drop-in replacements from [Taming-3dgs](https://humansensinglab.github.io/taming-3dgs/)<sup>1</sup> with [fused ssim](https://github.com/rahul-goel/fused-ssim/tree/main) into the original codebase to speed up training times. Once installed, the accelerated rasterizer delivers a **$\times$ 1.6 training time speedup** using `--optimizer_type default` and a **$\times$ 2.7 training time speedup** using `--optimizer_type sparse_adam`.
To get faster training times you must first install the accelerated rasterizer to your environment:
```bash
pip uninstall diff-gaussian-rasterization -y
cd submodules/diff-gaussian-rasterization
rm -r build
git checkout 3dgs_accel
pip install .
```
Then you can add the following parameter to use the sparse adam optimizer when running `train.py`:
```bash
--optimizer_type sparse_adam
```
*Note that this custom rasterizer has a different behaviour than the original version, for more details on training times please see [stats for training times](results.md/#training-times-comparisons)*.
*1. Mallick and Goel, et al. ‘Taming 3DGS: High-Quality Radiance Fields with Limited Resources’. SIGGRAPH Asia 2024 Conference Papers, 2024, https://doi.org/10.1145/3680528.3687694, [github](https://github.com/humansensinglab/taming-3dgs)*
### Depth regularization
To have better reconstructed scenes we use depth maps as priors during optimization with each input images. It works best on untextured parts ex: roads and can remove floaters. Several papers have used similar ideas to improve various aspects of 3DGS; (e.g. [DepthRegularizedGS](https://robot0321.github.io/DepthRegGS/index.html), [SparseGS](https://formycat.github.io/SparseGS-Real-Time-360-Sparse-View-Synthesis-using-Gaussian-Splatting/), [DNGaussian](https://fictionarry.github.io/DNGaussian/)). The depth regularization we integrated is that used in our [Hierarchical 3DGS](https://repo-sam.inria.fr/fungraph/hierarchical-3d-gaussians/) paper, but applied to the original 3DGS; for some scenes (e.g., the DeepBlending scenes) it improves quality significantly; for others it either makes a small difference or can even be worse. For example results showing the potential benefit and statistics on quality please see here: [Stats for depth regularization](results.md).
When training on a synthetic dataset, depth maps can be produced and they do not require further processing to be used in our method.
For real world datasets depth maps should be generated for each input images, to generate them please do the following:
1. Clone [Depth Anything v2](https://github.com/DepthAnything/Depth-Anything-V2?tab=readme-ov-file#usage):
```
git clone https://github.com/DepthAnything/Depth-Anything-V2.git
```
2. Download weights from [Depth-Anything-V2-Large](https://huggingface.co/depth-anything/Depth-Anything-V2-Large/resolve/main/depth_anything_v2_vitl.pth?download=true) and place it under `Depth-Anything-V2/checkpoints/`
3. Generate depth maps:
```
python Depth-Anything-V2/run.py --encoder vitl --pred-only --grayscale --img-path <path to input images> --outdir <output path>
```
5. Generate a `depth_params.json` file using:
```
python utils/make_depth_scale.py --base_dir <path to colmap> --depths_dir <path to generated depths>
```
A new parameter should be set when training if you want to use depth regularization `-d <path to depth maps>`.
### Exposure compensation
To compensate for exposure changes in the different input images we optimize an affine transformation for each image just as in [Hierarchical 3dgs](https://repo-sam.inria.fr/fungraph/hierarchical-3d-gaussians/).
This can greatly improve reconstruction results for "in the wild" captures, e.g., with a smartphone when the exposure setting of the camera is not fixed. For example results showing the potential benefit and statistics on quality please see here: [Stats for exposure compensation](results.md).
Add the following parameters to enable it:
```
--exposure_lr_init 0.001 --exposure_lr_final 0.0001 --exposure_lr_delay_steps 5000 --exposure_lr_delay_mult 0.001 --train_test_exp
```
Again, other excellent papers have used similar ideas e.g. [NeRF-W](https://nerf-w.github.io/), [URF](https://urban-radiance-fields.github.io/).
### Anti-aliasing
We added the EWA Filter from [Mip Splatting](https://niujinshuchong.github.io/mip-splatting/) in our codebase to remove aliasing. It is disabled by default but you can enable it by adding `--antialiasing` when training on a scene using `train.py` or rendering using `render.py`. Antialiasing can be toggled in the SIBR viewer, it is disabled by default but you should enable it when viewing a scene trained using `--antialiasing`.
*this scene was trained using `--antialiasing`*.
### SIBR: Top view
> `Views > Top view`
The `Top view` renders the SfM point cloud in another view with the corresponding input cameras and the `Point view` user camera. This allows visualization of how far the viewer is from the input cameras for example.
It is a 3D view so the user can navigate through it just as in the `Point view` (modes available: FPS, trackball, orbit).
<!-- _gif showing the top view, showing it is realtime_ -->
<!--  -->
Options are available to customize this view, meshes can be disabled/enabled and their scales can be modified.
<!-- _gif showing different options_ -->
<!--  -->
A useful additional functionality is to move to the position of an input image, and progressively fade out to the SfM point view in that position (e.g., to verify camera alignment). Views from input cameras can be displayed in the `Top view` (*note that `--images-path` must be set in the command line*). One can snap the `Top view` camera to the closest input camera from the user camera in the `Point view` by clicking `Top view settings > Cameras > Snap to closest`.
<!-- _gif showing for a snapped camera the ground truth image with alpha_ -->
<!--  -->
### OpenXR support
OpenXR is supported in the branch `gaussian_code_release_openxr`
Within that branch, you can find documentation for VR support [here](https://gitlab.inria.fr/sibr/sibr_core/-/tree/gaussian_code_release_openxr?ref_type=heads).
## FAQ
- *Where do I get data sets, e.g., those referenced in ```full_eval.py```?* The MipNeRF360 data set is provided by the authors of the original paper on the project site. Note that two of the data sets cannot be openly shared and require you to consult the authors directly. For Tanks&Temples and Deep Blending, please use the download links provided at the top of the page. Alternatively, you may access the cloned data (status: August 2023!) from [HuggingFace](https://huggingface.co/camenduru/gaussian-splatting)
- *How can I use this for a much larger dataset, like a city district?* The current method was not designed for these, but given enough memory, it should work out. However, the approach can struggle in multi-scale detail scenes (extreme close-ups, mixed with far-away shots). This is usually the case in, e.g., driving data sets (cars close up, buildings far away). For such scenes, you can lower the ```--position_lr_init```, ```--position_lr_final``` and ```--scaling_lr``` (x0.3, x0.1, ...). The more extensive the scene, the lower these values should be. Below, we use default learning rates (left) and ```--position_lr_init 0.000016 --scaling_lr 0.001"``` (right).
|  <!-- --> | <!-- -->  |
| --- | --- |
- *I'm on Windows and I can't manage to build the submodules, what do I do?* Consider following the steps in the excellent video tutorial [here](https://www.youtube.com/watch?v=UXtuigy_wYc), hopefully they should help. The order in which the steps are done is important! Alternatively, consider using the linked Colab template.
- *It still doesn't work. It says something about ```cl.exe```. What do I do?* User Henry Pearce found a workaround. You can you try adding the visual studio path to your environment variables (your version number might differ);
```C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30133\bin\Hostx64\x64```
Then make sure you start a new conda prompt and cd to your repo location and try this;
```
conda activate gaussian_splatting
cd <dir_to_repo>/gaussian-splatting
pip install submodules\diff-gaussian-rasterization
pip install submodules\simple-knn
```
- *I'm on macOS/Puppy Linux/Greenhat and I can't manage to build, what do I do?* Sorry, we can't provide support for platforms outside of the ones we list in this README. Consider using the linked Colab template.
- *I don't have 24 GB of VRAM for training, what do I do?* The VRAM consumption is determined by the number of points that are being optimized, which increases over time. If you only want to train to 7k iterations, you will need significantly less. To do the full training routine and avoid running out of memory, you can increase the ```--densify_grad_threshold```, ```--densification_interval``` or reduce the value of ```--densify_until_iter```. Note however that this will affect the quality of the result. Also try setting ```--test_iterations``` to ```-1``` to avoid memory spikes during testing. If ```--densify_grad_threshold``` is very high, no densification should occur and training should complete if the scene itself loads successfully.
- *24 GB of VRAM for reference quality training is still a lot! Can't we do it with less?* Yes, most likely. By our calculations it should be possible with **way** less memory (~8GB). If we can find the time we will try to achieve this. If some PyTorch veteran out there wants to tackle this, we look forward to your pull request!
- *How can I use the differentiable Gaussian rasterizer for my own project?* Easy, it is included in this repo as a submodule ```diff-gaussian-rasterization```. Feel free to check out and install the package. It's not really documented, but using it from the Python side is very straightforward (cf. ```gaussian_renderer/__init__.py```).
- *Wait, but ```<insert feature>``` isn't optimized and could be much better?* There are several parts we didn't even have time to think about improving (yet). The performance you get with this prototype is probably a rather slow baseline for what is physically possible.
- *Something is broken, how did this happen?* We tried hard to provide a solid and comprehensible basis to make use of the paper's method. We have refactored the code quite a bit, but we have limited capacity to test all possible usage scenarios. Thus, if part of the website, the code or the performance is lacking, please create an issue. If we find the time, we will do our best to address it.
================================================
FILE: arguments/__init__.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
from argparse import ArgumentParser, Namespace
import sys
import os
class GroupParams:
pass
class ParamGroup:
def __init__(self, parser: ArgumentParser, name : str, fill_none = False):
group = parser.add_argument_group(name)
for key, value in vars(self).items():
shorthand = False
if key.startswith("_"):
shorthand = True
key = key[1:]
t = type(value)
value = value if not fill_none else None
if shorthand:
if t == bool:
group.add_argument("--" + key, ("-" + key[0:1]), default=value, action="store_true")
else:
group.add_argument("--" + key, ("-" + key[0:1]), default=value, type=t)
else:
if t == bool:
group.add_argument("--" + key, default=value, action="store_true")
else:
group.add_argument("--" + key, default=value, type=t)
def extract(self, args):
group = GroupParams()
for arg in vars(args).items():
if arg[0] in vars(self) or ("_" + arg[0]) in vars(self):
setattr(group, arg[0], arg[1])
return group
class ModelParams(ParamGroup):
def __init__(self, parser, sentinel=False):
self.sh_degree = 3
self._source_path = ""
self._model_path = ""
self._images = "images"
self._depths = ""
self._resolution = -1
self._white_background = False
self.train_test_exp = False
self.data_device = "cuda"
self.eval = False
super().__init__(parser, "Loading Parameters", sentinel)
def extract(self, args):
g = super().extract(args)
g.source_path = os.path.abspath(g.source_path)
return g
class PipelineParams(ParamGroup):
def __init__(self, parser):
self.convert_SHs_python = False
self.compute_cov3D_python = False
self.debug = False
self.antialiasing = False
super().__init__(parser, "Pipeline Parameters")
class OptimizationParams(ParamGroup):
def __init__(self, parser):
self.iterations = 30_000
self.position_lr_init = 0.00016
self.position_lr_final = 0.0000016
self.position_lr_delay_mult = 0.01
self.position_lr_max_steps = 30_000
self.feature_lr = 0.0025
self.opacity_lr = 0.025
self.scaling_lr = 0.005
self.rotation_lr = 0.001
self.exposure_lr_init = 0.01
self.exposure_lr_final = 0.001
self.exposure_lr_delay_steps = 0
self.exposure_lr_delay_mult = 0.0
self.percent_dense = 0.01
self.lambda_dssim = 0.2
self.densification_interval = 100
self.opacity_reset_interval = 3000
self.densify_from_iter = 500
self.densify_until_iter = 15_000
self.densify_grad_threshold = 0.0002
self.depth_l1_weight_init = 1.0
self.depth_l1_weight_final = 0.01
self.random_background = False
self.optimizer_type = "default"
super().__init__(parser, "Optimization Parameters")
def get_combined_args(parser : ArgumentParser):
cmdlne_string = sys.argv[1:]
cfgfile_string = "Namespace()"
args_cmdline = parser.parse_args(cmdlne_string)
try:
cfgfilepath = os.path.join(args_cmdline.model_path, "cfg_args")
print("Looking for config file in", cfgfilepath)
with open(cfgfilepath) as cfg_file:
print("Config file found: {}".format(cfgfilepath))
cfgfile_string = cfg_file.read()
except TypeError:
print("Config file not found at")
pass
args_cfgfile = eval(cfgfile_string)
merged_dict = vars(args_cfgfile).copy()
for k,v in vars(args_cmdline).items():
if v != None:
merged_dict[k] = v
return Namespace(**merged_dict)
================================================
FILE: convert.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import os
import logging
from argparse import ArgumentParser
import shutil
# This Python script is based on the shell converter script provided in the MipNerF 360 repository.
parser = ArgumentParser("Colmap converter")
parser.add_argument("--no_gpu", action='store_true')
parser.add_argument("--skip_matching", action='store_true')
parser.add_argument("--source_path", "-s", required=True, type=str)
parser.add_argument("--camera", default="OPENCV", type=str)
parser.add_argument("--colmap_executable", default="", type=str)
parser.add_argument("--resize", action="store_true")
parser.add_argument("--magick_executable", default="", type=str)
args = parser.parse_args()
colmap_command = '"{}"'.format(args.colmap_executable) if len(args.colmap_executable) > 0 else "colmap"
magick_command = '"{}"'.format(args.magick_executable) if len(args.magick_executable) > 0 else "magick"
use_gpu = 1 if not args.no_gpu else 0
if not args.skip_matching:
os.makedirs(args.source_path + "/distorted/sparse", exist_ok=True)
## Feature extraction
feat_extracton_cmd = colmap_command + " feature_extractor "\
"--database_path " + args.source_path + "/distorted/database.db \
--image_path " + args.source_path + "/input \
--ImageReader.single_camera 1 \
--ImageReader.camera_model " + args.camera + " \
--SiftExtraction.use_gpu " + str(use_gpu)
exit_code = os.system(feat_extracton_cmd)
if exit_code != 0:
logging.error(f"Feature extraction failed with code {exit_code}. Exiting.")
exit(exit_code)
## Feature matching
feat_matching_cmd = colmap_command + " exhaustive_matcher \
--database_path " + args.source_path + "/distorted/database.db \
--SiftMatching.use_gpu " + str(use_gpu)
exit_code = os.system(feat_matching_cmd)
if exit_code != 0:
logging.error(f"Feature matching failed with code {exit_code}. Exiting.")
exit(exit_code)
### Bundle adjustment
# The default Mapper tolerance is unnecessarily large,
# decreasing it speeds up bundle adjustment steps.
mapper_cmd = (colmap_command + " mapper \
--database_path " + args.source_path + "/distorted/database.db \
--image_path " + args.source_path + "/input \
--output_path " + args.source_path + "/distorted/sparse \
--Mapper.ba_global_function_tolerance=0.000001")
exit_code = os.system(mapper_cmd)
if exit_code != 0:
logging.error(f"Mapper failed with code {exit_code}. Exiting.")
exit(exit_code)
### Image undistortion
## We need to undistort our images into ideal pinhole intrinsics.
img_undist_cmd = (colmap_command + " image_undistorter \
--image_path " + args.source_path + "/input \
--input_path " + args.source_path + "/distorted/sparse/0 \
--output_path " + args.source_path + "\
--output_type COLMAP")
exit_code = os.system(img_undist_cmd)
if exit_code != 0:
logging.error(f"Mapper failed with code {exit_code}. Exiting.")
exit(exit_code)
files = os.listdir(args.source_path + "/sparse")
os.makedirs(args.source_path + "/sparse/0", exist_ok=True)
# Copy each file from the source directory to the destination directory
for file in files:
if file == '0':
continue
source_file = os.path.join(args.source_path, "sparse", file)
destination_file = os.path.join(args.source_path, "sparse", "0", file)
shutil.move(source_file, destination_file)
if(args.resize):
print("Copying and resizing...")
# Resize images.
os.makedirs(args.source_path + "/images_2", exist_ok=True)
os.makedirs(args.source_path + "/images_4", exist_ok=True)
os.makedirs(args.source_path + "/images_8", exist_ok=True)
# Get the list of files in the source directory
files = os.listdir(args.source_path + "/images")
# Copy each file from the source directory to the destination directory
for file in files:
source_file = os.path.join(args.source_path, "images", file)
destination_file = os.path.join(args.source_path, "images_2", file)
shutil.copy2(source_file, destination_file)
exit_code = os.system(magick_command + " mogrify -resize 50% " + destination_file)
if exit_code != 0:
logging.error(f"50% resize failed with code {exit_code}. Exiting.")
exit(exit_code)
destination_file = os.path.join(args.source_path, "images_4", file)
shutil.copy2(source_file, destination_file)
exit_code = os.system(magick_command + " mogrify -resize 25% " + destination_file)
if exit_code != 0:
logging.error(f"25% resize failed with code {exit_code}. Exiting.")
exit(exit_code)
destination_file = os.path.join(args.source_path, "images_8", file)
shutil.copy2(source_file, destination_file)
exit_code = os.system(magick_command + " mogrify -resize 12.5% " + destination_file)
if exit_code != 0:
logging.error(f"12.5% resize failed with code {exit_code}. Exiting.")
exit(exit_code)
print("Done.")
================================================
FILE: environment.yml
================================================
name: gaussian_splatting
channels:
- pytorch
- conda-forge
- defaults
dependencies:
- cudatoolkit=11.6
- plyfile
- python=3.7.13
- pip=22.3.1
- pytorch=1.12.1
- torchaudio=0.12.1
- torchvision=0.13.1
- tqdm
- pip:
- submodules/diff-gaussian-rasterization
- submodules/simple-knn
- submodules/fused-ssim
- opencv-python
- joblib
================================================
FILE: full_eval.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import os
from argparse import ArgumentParser
import time
mipnerf360_outdoor_scenes = ["bicycle", "flowers", "garden", "stump", "treehill"]
mipnerf360_indoor_scenes = ["room", "counter", "kitchen", "bonsai"]
tanks_and_temples_scenes = ["truck", "train"]
deep_blending_scenes = ["drjohnson", "playroom"]
parser = ArgumentParser(description="Full evaluation script parameters")
parser.add_argument("--skip_training", action="store_true")
parser.add_argument("--skip_rendering", action="store_true")
parser.add_argument("--skip_metrics", action="store_true")
parser.add_argument("--output_path", default="./eval")
parser.add_argument("--use_depth", action="store_true")
parser.add_argument("--use_expcomp", action="store_true")
parser.add_argument("--fast", action="store_true")
parser.add_argument("--aa", action="store_true")
args, _ = parser.parse_known_args()
all_scenes = []
all_scenes.extend(mipnerf360_outdoor_scenes)
all_scenes.extend(mipnerf360_indoor_scenes)
all_scenes.extend(tanks_and_temples_scenes)
all_scenes.extend(deep_blending_scenes)
if not args.skip_training or not args.skip_rendering:
parser.add_argument('--mipnerf360', "-m360", required=True, type=str)
parser.add_argument("--tanksandtemples", "-tat", required=True, type=str)
parser.add_argument("--deepblending", "-db", required=True, type=str)
args = parser.parse_args()
if not args.skip_training:
common_args = " --disable_viewer --quiet --eval --test_iterations -1 "
if args.aa:
common_args += " --antialiasing "
if args.use_depth:
common_args += " -d depths2/ "
if args.use_expcomp:
common_args += " --exposure_lr_init 0.001 --exposure_lr_final 0.0001 --exposure_lr_delay_steps 5000 --exposure_lr_delay_mult 0.001 --train_test_exp "
if args.fast:
common_args += " --optimizer_type sparse_adam "
start_time = time.time()
for scene in mipnerf360_outdoor_scenes:
source = args.mipnerf360 + "/" + scene
os.system("python train.py -s " + source + " -i images_4 -m " + args.output_path + "/" + scene + common_args)
for scene in mipnerf360_indoor_scenes:
source = args.mipnerf360 + "/" + scene
os.system("python train.py -s " + source + " -i images_2 -m " + args.output_path + "/" + scene + common_args)
m360_timing = (time.time() - start_time)/60.0
start_time = time.time()
for scene in tanks_and_temples_scenes:
source = args.tanksandtemples + "/" + scene
os.system("python train.py -s " + source + " -m " + args.output_path + "/" + scene + common_args)
tandt_timing = (time.time() - start_time)/60.0
start_time = time.time()
for scene in deep_blending_scenes:
source = args.deepblending + "/" + scene
os.system("python train.py -s " + source + " -m " + args.output_path + "/" + scene + common_args)
db_timing = (time.time() - start_time)/60.0
with open(os.path.join(args.output_path,"timing.txt"), 'w') as file:
file.write(f"m360: {m360_timing} minutes \n tandt: {tandt_timing} minutes \n db: {db_timing} minutes\n")
if not args.skip_rendering:
all_sources = []
for scene in mipnerf360_outdoor_scenes:
all_sources.append(args.mipnerf360 + "/" + scene)
for scene in mipnerf360_indoor_scenes:
all_sources.append(args.mipnerf360 + "/" + scene)
for scene in tanks_and_temples_scenes:
all_sources.append(args.tanksandtemples + "/" + scene)
for scene in deep_blending_scenes:
all_sources.append(args.deepblending + "/" + scene)
common_args = " --quiet --eval --skip_train"
if args.aa:
common_args += " --antialiasing "
if args.use_expcomp:
common_args += " --train_test_exp "
for scene, source in zip(all_scenes, all_sources):
os.system("python render.py --iteration 7000 -s " + source + " -m " + args.output_path + "/" + scene + common_args)
os.system("python render.py --iteration 30000 -s " + source + " -m " + args.output_path + "/" + scene + common_args)
if not args.skip_metrics:
scenes_string = ""
for scene in all_scenes:
scenes_string += "\"" + args.output_path + "/" + scene + "\" "
os.system("python metrics.py -m " + scenes_string)
================================================
FILE: gaussian_renderer/__init__.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import torch
import math
from diff_gaussian_rasterization import GaussianRasterizationSettings, GaussianRasterizer
from scene.gaussian_model import GaussianModel
from utils.sh_utils import eval_sh
def render(viewpoint_camera, pc : GaussianModel, pipe, bg_color : torch.Tensor, scaling_modifier = 1.0, separate_sh = False, override_color = None, use_trained_exp=False):
"""
Render the scene.
Background tensor (bg_color) must be on GPU!
"""
# Create zero tensor. We will use it to make pytorch return gradients of the 2D (screen-space) means
screenspace_points = torch.zeros_like(pc.get_xyz, dtype=pc.get_xyz.dtype, requires_grad=True, device="cuda") + 0
try:
screenspace_points.retain_grad()
except:
pass
# Set up rasterization configuration
tanfovx = math.tan(viewpoint_camera.FoVx * 0.5)
tanfovy = math.tan(viewpoint_camera.FoVy * 0.5)
raster_settings = GaussianRasterizationSettings(
image_height=int(viewpoint_camera.image_height),
image_width=int(viewpoint_camera.image_width),
tanfovx=tanfovx,
tanfovy=tanfovy,
bg=bg_color,
scale_modifier=scaling_modifier,
viewmatrix=viewpoint_camera.world_view_transform,
projmatrix=viewpoint_camera.full_proj_transform,
sh_degree=pc.active_sh_degree,
campos=viewpoint_camera.camera_center,
prefiltered=False,
debug=pipe.debug,
antialiasing=pipe.antialiasing
)
rasterizer = GaussianRasterizer(raster_settings=raster_settings)
means3D = pc.get_xyz
means2D = screenspace_points
opacity = pc.get_opacity
# If precomputed 3d covariance is provided, use it. If not, then it will be computed from
# scaling / rotation by the rasterizer.
scales = None
rotations = None
cov3D_precomp = None
if pipe.compute_cov3D_python:
cov3D_precomp = pc.get_covariance(scaling_modifier)
else:
scales = pc.get_scaling
rotations = pc.get_rotation
# If precomputed colors are provided, use them. Otherwise, if it is desired to precompute colors
# from SHs in Python, do it. If not, then SH -> RGB conversion will be done by rasterizer.
shs = None
colors_precomp = None
if override_color is None:
if pipe.convert_SHs_python:
shs_view = pc.get_features.transpose(1, 2).view(-1, 3, (pc.max_sh_degree+1)**2)
dir_pp = (pc.get_xyz - viewpoint_camera.camera_center.repeat(pc.get_features.shape[0], 1))
dir_pp_normalized = dir_pp/dir_pp.norm(dim=1, keepdim=True)
sh2rgb = eval_sh(pc.active_sh_degree, shs_view, dir_pp_normalized)
colors_precomp = torch.clamp_min(sh2rgb + 0.5, 0.0)
else:
if separate_sh:
dc, shs = pc.get_features_dc, pc.get_features_rest
else:
shs = pc.get_features
else:
colors_precomp = override_color
# Rasterize visible Gaussians to image, obtain their radii (on screen).
if separate_sh:
rendered_image, radii, depth_image = rasterizer(
means3D = means3D,
means2D = means2D,
dc = dc,
shs = shs,
colors_precomp = colors_precomp,
opacities = opacity,
scales = scales,
rotations = rotations,
cov3D_precomp = cov3D_precomp)
else:
rendered_image, radii, depth_image = rasterizer(
means3D = means3D,
means2D = means2D,
shs = shs,
colors_precomp = colors_precomp,
opacities = opacity,
scales = scales,
rotations = rotations,
cov3D_precomp = cov3D_precomp)
# Apply exposure to rendered image (training only)
if use_trained_exp:
exposure = pc.get_exposure_from_name(viewpoint_camera.image_name)
rendered_image = torch.matmul(rendered_image.permute(1, 2, 0), exposure[:3, :3]).permute(2, 0, 1) + exposure[:3, 3, None, None]
# Those Gaussians that were frustum culled or had a radius of 0 were not visible.
# They will be excluded from value updates used in the splitting criteria.
rendered_image = rendered_image.clamp(0, 1)
out = {
"render": rendered_image,
"viewspace_points": screenspace_points,
"visibility_filter" : (radii > 0).nonzero(),
"radii": radii,
"depth" : depth_image
}
return out
================================================
FILE: gaussian_renderer/network_gui.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import torch
import traceback
import socket
import json
from scene.cameras import MiniCam
host = "127.0.0.1"
port = 6009
conn = None
addr = None
listener = socket.socket(socket.AF_INET, socket.SOCK_STREAM)
def init(wish_host, wish_port):
global host, port, listener
host = wish_host
port = wish_port
listener.bind((host, port))
listener.listen()
listener.settimeout(0)
def try_connect():
global conn, addr, listener
try:
conn, addr = listener.accept()
print(f"\nConnected by {addr}")
conn.settimeout(None)
except Exception as inst:
pass
def read():
global conn
messageLength = conn.recv(4)
messageLength = int.from_bytes(messageLength, 'little')
message = conn.recv(messageLength)
return json.loads(message.decode("utf-8"))
def send(message_bytes, verify):
global conn
if message_bytes != None:
conn.sendall(message_bytes)
conn.sendall(len(verify).to_bytes(4, 'little'))
conn.sendall(bytes(verify, 'ascii'))
def receive():
message = read()
width = message["resolution_x"]
height = message["resolution_y"]
if width != 0 and height != 0:
try:
do_training = bool(message["train"])
fovy = message["fov_y"]
fovx = message["fov_x"]
znear = message["z_near"]
zfar = message["z_far"]
do_shs_python = bool(message["shs_python"])
do_rot_scale_python = bool(message["rot_scale_python"])
keep_alive = bool(message["keep_alive"])
scaling_modifier = message["scaling_modifier"]
world_view_transform = torch.reshape(torch.tensor(message["view_matrix"]), (4, 4)).cuda()
world_view_transform[:,1] = -world_view_transform[:,1]
world_view_transform[:,2] = -world_view_transform[:,2]
full_proj_transform = torch.reshape(torch.tensor(message["view_projection_matrix"]), (4, 4)).cuda()
full_proj_transform[:,1] = -full_proj_transform[:,1]
custom_cam = MiniCam(width, height, fovy, fovx, znear, zfar, world_view_transform, full_proj_transform)
except Exception as e:
print("")
traceback.print_exc()
raise e
return custom_cam, do_training, do_shs_python, do_rot_scale_python, keep_alive, scaling_modifier
else:
return None, None, None, None, None, None
================================================
FILE: lpipsPyTorch/__init__.py
================================================
import torch
from .modules.lpips import LPIPS
def lpips(x: torch.Tensor,
y: torch.Tensor,
net_type: str = 'alex',
version: str = '0.1'):
r"""Function that measures
Learned Perceptual Image Patch Similarity (LPIPS).
Arguments:
x, y (torch.Tensor): the input tensors to compare.
net_type (str): the network type to compare the features:
'alex' | 'squeeze' | 'vgg'. Default: 'alex'.
version (str): the version of LPIPS. Default: 0.1.
"""
device = x.device
criterion = LPIPS(net_type, version).to(device)
return criterion(x, y)
================================================
FILE: lpipsPyTorch/modules/lpips.py
================================================
import torch
import torch.nn as nn
from .networks import get_network, LinLayers
from .utils import get_state_dict
class LPIPS(nn.Module):
r"""Creates a criterion that measures
Learned Perceptual Image Patch Similarity (LPIPS).
Arguments:
net_type (str): the network type to compare the features:
'alex' | 'squeeze' | 'vgg'. Default: 'alex'.
version (str): the version of LPIPS. Default: 0.1.
"""
def __init__(self, net_type: str = 'alex', version: str = '0.1'):
assert version in ['0.1'], 'v0.1 is only supported now'
super(LPIPS, self).__init__()
# pretrained network
self.net = get_network(net_type)
# linear layers
self.lin = LinLayers(self.net.n_channels_list)
self.lin.load_state_dict(get_state_dict(net_type, version))
def forward(self, x: torch.Tensor, y: torch.Tensor):
feat_x, feat_y = self.net(x), self.net(y)
diff = [(fx - fy) ** 2 for fx, fy in zip(feat_x, feat_y)]
res = [l(d).mean((2, 3), True) for d, l in zip(diff, self.lin)]
return torch.sum(torch.cat(res, 0), 0, True)
================================================
FILE: lpipsPyTorch/modules/networks.py
================================================
from typing import Sequence
from itertools import chain
import torch
import torch.nn as nn
from torchvision import models
from .utils import normalize_activation
def get_network(net_type: str):
if net_type == 'alex':
return AlexNet()
elif net_type == 'squeeze':
return SqueezeNet()
elif net_type == 'vgg':
return VGG16()
else:
raise NotImplementedError('choose net_type from [alex, squeeze, vgg].')
class LinLayers(nn.ModuleList):
def __init__(self, n_channels_list: Sequence[int]):
super(LinLayers, self).__init__([
nn.Sequential(
nn.Identity(),
nn.Conv2d(nc, 1, 1, 1, 0, bias=False)
) for nc in n_channels_list
])
for param in self.parameters():
param.requires_grad = False
class BaseNet(nn.Module):
def __init__(self):
super(BaseNet, self).__init__()
# register buffer
self.register_buffer(
'mean', torch.Tensor([-.030, -.088, -.188])[None, :, None, None])
self.register_buffer(
'std', torch.Tensor([.458, .448, .450])[None, :, None, None])
def set_requires_grad(self, state: bool):
for param in chain(self.parameters(), self.buffers()):
param.requires_grad = state
def z_score(self, x: torch.Tensor):
return (x - self.mean) / self.std
def forward(self, x: torch.Tensor):
x = self.z_score(x)
output = []
for i, (_, layer) in enumerate(self.layers._modules.items(), 1):
x = layer(x)
if i in self.target_layers:
output.append(normalize_activation(x))
if len(output) == len(self.target_layers):
break
return output
class SqueezeNet(BaseNet):
def __init__(self):
super(SqueezeNet, self).__init__()
self.layers = models.squeezenet1_1(True).features
self.target_layers = [2, 5, 8, 10, 11, 12, 13]
self.n_channels_list = [64, 128, 256, 384, 384, 512, 512]
self.set_requires_grad(False)
class AlexNet(BaseNet):
def __init__(self):
super(AlexNet, self).__init__()
self.layers = models.alexnet(True).features
self.target_layers = [2, 5, 8, 10, 12]
self.n_channels_list = [64, 192, 384, 256, 256]
self.set_requires_grad(False)
class VGG16(BaseNet):
def __init__(self):
super(VGG16, self).__init__()
self.layers = models.vgg16(weights=models.VGG16_Weights.IMAGENET1K_V1).features
self.target_layers = [4, 9, 16, 23, 30]
self.n_channels_list = [64, 128, 256, 512, 512]
self.set_requires_grad(False)
================================================
FILE: lpipsPyTorch/modules/utils.py
================================================
from collections import OrderedDict
import torch
def normalize_activation(x, eps=1e-10):
norm_factor = torch.sqrt(torch.sum(x ** 2, dim=1, keepdim=True))
return x / (norm_factor + eps)
def get_state_dict(net_type: str = 'alex', version: str = '0.1'):
# build url
url = 'https://raw.githubusercontent.com/richzhang/PerceptualSimilarity/' \
+ f'master/lpips/weights/v{version}/{net_type}.pth'
# download
old_state_dict = torch.hub.load_state_dict_from_url(
url, progress=True,
map_location=None if torch.cuda.is_available() else torch.device('cpu')
)
# rename keys
new_state_dict = OrderedDict()
for key, val in old_state_dict.items():
new_key = key
new_key = new_key.replace('lin', '')
new_key = new_key.replace('model.', '')
new_state_dict[new_key] = val
return new_state_dict
================================================
FILE: metrics.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
from pathlib import Path
import os
from PIL import Image
import torch
import torchvision.transforms.functional as tf
from utils.loss_utils import ssim
from lpipsPyTorch import lpips
import json
from tqdm import tqdm
from utils.image_utils import psnr
from argparse import ArgumentParser
def readImages(renders_dir, gt_dir):
renders = []
gts = []
image_names = []
for fname in os.listdir(renders_dir):
render = Image.open(renders_dir / fname)
gt = Image.open(gt_dir / fname)
renders.append(tf.to_tensor(render).unsqueeze(0)[:, :3, :, :].cuda())
gts.append(tf.to_tensor(gt).unsqueeze(0)[:, :3, :, :].cuda())
image_names.append(fname)
return renders, gts, image_names
def evaluate(model_paths):
full_dict = {}
per_view_dict = {}
full_dict_polytopeonly = {}
per_view_dict_polytopeonly = {}
print("")
for scene_dir in model_paths:
try:
print("Scene:", scene_dir)
full_dict[scene_dir] = {}
per_view_dict[scene_dir] = {}
full_dict_polytopeonly[scene_dir] = {}
per_view_dict_polytopeonly[scene_dir] = {}
test_dir = Path(scene_dir) / "test"
for method in os.listdir(test_dir):
print("Method:", method)
full_dict[scene_dir][method] = {}
per_view_dict[scene_dir][method] = {}
full_dict_polytopeonly[scene_dir][method] = {}
per_view_dict_polytopeonly[scene_dir][method] = {}
method_dir = test_dir / method
gt_dir = method_dir/ "gt"
renders_dir = method_dir / "renders"
renders, gts, image_names = readImages(renders_dir, gt_dir)
ssims = []
psnrs = []
lpipss = []
for idx in tqdm(range(len(renders)), desc="Metric evaluation progress"):
ssims.append(ssim(renders[idx], gts[idx]))
psnrs.append(psnr(renders[idx], gts[idx]))
lpipss.append(lpips(renders[idx], gts[idx], net_type='vgg'))
print(" SSIM : {:>12.7f}".format(torch.tensor(ssims).mean(), ".5"))
print(" PSNR : {:>12.7f}".format(torch.tensor(psnrs).mean(), ".5"))
print(" LPIPS: {:>12.7f}".format(torch.tensor(lpipss).mean(), ".5"))
print("")
full_dict[scene_dir][method].update({"SSIM": torch.tensor(ssims).mean().item(),
"PSNR": torch.tensor(psnrs).mean().item(),
"LPIPS": torch.tensor(lpipss).mean().item()})
per_view_dict[scene_dir][method].update({"SSIM": {name: ssim for ssim, name in zip(torch.tensor(ssims).tolist(), image_names)},
"PSNR": {name: psnr for psnr, name in zip(torch.tensor(psnrs).tolist(), image_names)},
"LPIPS": {name: lp for lp, name in zip(torch.tensor(lpipss).tolist(), image_names)}})
with open(scene_dir + "/results.json", 'w') as fp:
json.dump(full_dict[scene_dir], fp, indent=True)
with open(scene_dir + "/per_view.json", 'w') as fp:
json.dump(per_view_dict[scene_dir], fp, indent=True)
except:
print("Unable to compute metrics for model", scene_dir)
if __name__ == "__main__":
device = torch.device("cuda:0")
torch.cuda.set_device(device)
# Set up command line argument parser
parser = ArgumentParser(description="Training script parameters")
parser.add_argument('--model_paths', '-m', required=True, nargs="+", type=str, default=[])
args = parser.parse_args()
evaluate(args.model_paths)
================================================
FILE: render.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import torch
from scene import Scene
import os
from tqdm import tqdm
from os import makedirs
from gaussian_renderer import render
import torchvision
from utils.general_utils import safe_state
from argparse import ArgumentParser
from arguments import ModelParams, PipelineParams, get_combined_args
from gaussian_renderer import GaussianModel
try:
from diff_gaussian_rasterization import SparseGaussianAdam
SPARSE_ADAM_AVAILABLE = True
except:
SPARSE_ADAM_AVAILABLE = False
def render_set(model_path, name, iteration, views, gaussians, pipeline, background, train_test_exp, separate_sh):
render_path = os.path.join(model_path, name, "ours_{}".format(iteration), "renders")
gts_path = os.path.join(model_path, name, "ours_{}".format(iteration), "gt")
makedirs(render_path, exist_ok=True)
makedirs(gts_path, exist_ok=True)
for idx, view in enumerate(tqdm(views, desc="Rendering progress")):
rendering = render(view, gaussians, pipeline, background, use_trained_exp=train_test_exp, separate_sh=separate_sh)["render"]
gt = view.original_image[0:3, :, :]
if args.train_test_exp:
rendering = rendering[..., rendering.shape[-1] // 2:]
gt = gt[..., gt.shape[-1] // 2:]
torchvision.utils.save_image(rendering, os.path.join(render_path, '{0:05d}'.format(idx) + ".png"))
torchvision.utils.save_image(gt, os.path.join(gts_path, '{0:05d}'.format(idx) + ".png"))
def render_sets(dataset : ModelParams, iteration : int, pipeline : PipelineParams, skip_train : bool, skip_test : bool, separate_sh: bool):
with torch.no_grad():
gaussians = GaussianModel(dataset.sh_degree)
scene = Scene(dataset, gaussians, load_iteration=iteration, shuffle=False)
bg_color = [1,1,1] if dataset.white_background else [0, 0, 0]
background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")
if not skip_train:
render_set(dataset.model_path, "train", scene.loaded_iter, scene.getTrainCameras(), gaussians, pipeline, background, dataset.train_test_exp, separate_sh)
if not skip_test:
render_set(dataset.model_path, "test", scene.loaded_iter, scene.getTestCameras(), gaussians, pipeline, background, dataset.train_test_exp, separate_sh)
if __name__ == "__main__":
# Set up command line argument parser
parser = ArgumentParser(description="Testing script parameters")
model = ModelParams(parser, sentinel=True)
pipeline = PipelineParams(parser)
parser.add_argument("--iteration", default=-1, type=int)
parser.add_argument("--skip_train", action="store_true")
parser.add_argument("--skip_test", action="store_true")
parser.add_argument("--quiet", action="store_true")
args = get_combined_args(parser)
print("Rendering " + args.model_path)
# Initialize system state (RNG)
safe_state(args.quiet)
render_sets(model.extract(args), args.iteration, pipeline.extract(args), args.skip_train, args.skip_test, SPARSE_ADAM_AVAILABLE)
================================================
FILE: results.md
================================================
# Evaluations
We evaluated the impact of the features we added on MipNeRF360, Tanks&Temples and Deep Blending datasets. [Exposure Compensation](#exposure-compensation) is evaluated separately. Note that [Default rasterizer](#default-rasterizer) refers to the original [3dgs rasterizer](https://github.com/graphdeco-inria/diff-gaussian-rasterization/tree/9c5c2028f6fbee2be239bc4c9421ff894fe4fbe0) and [Accelerated rasterizer](#accelerated-rasterizer) refers to the [taming-3dgs rasterizer](https://github.com/graphdeco-inria/diff-gaussian-rasterization/tree/3dgs_accel).
## Default rasterizer
### PSNR
***DR**:depth regularization, **AA**:antialiasing*
<br>
### SSIM
***DR**:depth regularization, **AA**:antialiasing*
### LPIPS
*lower is better, **DR**:depth regularization, **AA**:antialiasing*
## Accelerated rasterizer
### Default optimizer
These numbers were obtained using the accelerated rasterizer and `--optimizer_type default` when training.
#### PSNR
***DR**:depth regularization, **AA**:antialiasing*
#### SSIM
***DR**:depth regularization, **AA**:antialiasing*
#### LPIPS
*lower is better, **DR**:depth regularization, **AA**:antialiasing*
### Sparse Adam optimizer
These numbers were obtained using the accelerated rasterizer and `--optimizer_type sparse_adam` when training.
#### PSNR
***DR**:depth regularization, **AA**:antialiasing*
#### SSIM
***DR**:depth regularization, **AA**:antialiasing*
#### LPIPS
*lower is better, **DR**:depth regularization, **AA**:antialiasing*
## Exposure compensation
We account for exposure variations between images by optimizing a 3x4 affine transform for each image. During training, this transform is applied to the colour of the rendered images.
The exposure compensation is designed to improve the inputs' coherence during training and is not applied during real-time navigation.
Enabling the `--train_test_exp` option includes the left half of the test images in the training set, using only their right halves for testing, following the same testing methodology as NeRF-W and Mega-NeRF. This allows us to optimize the exposure affine transform for test views. However, since this setting alters the train/test splits, the resulting metrics are not comparable to those from models trained without it. Here we provide results with `--train_test_exp`, with and without exposure compensation.
### PSNR
### SSIM
### LPIPS
*Lower is better.*
## Training times comparisons
We report the training times with all features enabled using the original 3dgs rasterizer *(baseline)* and the accelerated rasterizer with default optimizer then sparse adam.
================================================
FILE: scene/__init__.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import os
import random
import json
from utils.system_utils import searchForMaxIteration
from scene.dataset_readers import sceneLoadTypeCallbacks
from scene.gaussian_model import GaussianModel
from arguments import ModelParams
from utils.camera_utils import cameraList_from_camInfos, camera_to_JSON
class Scene:
gaussians : GaussianModel
def __init__(self, args : ModelParams, gaussians : GaussianModel, load_iteration=None, shuffle=True, resolution_scales=[1.0]):
"""b
:param path: Path to colmap scene main folder.
"""
self.model_path = args.model_path
self.loaded_iter = None
self.gaussians = gaussians
if load_iteration:
if load_iteration == -1:
self.loaded_iter = searchForMaxIteration(os.path.join(self.model_path, "point_cloud"))
else:
self.loaded_iter = load_iteration
print("Loading trained model at iteration {}".format(self.loaded_iter))
self.train_cameras = {}
self.test_cameras = {}
if os.path.exists(os.path.join(args.source_path, "sparse")):
scene_info = sceneLoadTypeCallbacks["Colmap"](args.source_path, args.images, args.depths, args.eval, args.train_test_exp)
elif os.path.exists(os.path.join(args.source_path, "transforms_train.json")):
print("Found transforms_train.json file, assuming Blender data set!")
scene_info = sceneLoadTypeCallbacks["Blender"](args.source_path, args.white_background, args.depths, args.eval)
else:
assert False, "Could not recognize scene type!"
if not self.loaded_iter:
with open(scene_info.ply_path, 'rb') as src_file, open(os.path.join(self.model_path, "input.ply") , 'wb') as dest_file:
dest_file.write(src_file.read())
json_cams = []
camlist = []
if scene_info.test_cameras:
camlist.extend(scene_info.test_cameras)
if scene_info.train_cameras:
camlist.extend(scene_info.train_cameras)
for id, cam in enumerate(camlist):
json_cams.append(camera_to_JSON(id, cam))
with open(os.path.join(self.model_path, "cameras.json"), 'w') as file:
json.dump(json_cams, file)
if shuffle:
random.shuffle(scene_info.train_cameras) # Multi-res consistent random shuffling
random.shuffle(scene_info.test_cameras) # Multi-res consistent random shuffling
self.cameras_extent = scene_info.nerf_normalization["radius"]
for resolution_scale in resolution_scales:
print("Loading Training Cameras")
self.train_cameras[resolution_scale] = cameraList_from_camInfos(scene_info.train_cameras, resolution_scale, args, scene_info.is_nerf_synthetic, False)
print("Loading Test Cameras")
self.test_cameras[resolution_scale] = cameraList_from_camInfos(scene_info.test_cameras, resolution_scale, args, scene_info.is_nerf_synthetic, True)
if self.loaded_iter:
self.gaussians.load_ply(os.path.join(self.model_path,
"point_cloud",
"iteration_" + str(self.loaded_iter),
"point_cloud.ply"), args.train_test_exp)
else:
self.gaussians.create_from_pcd(scene_info.point_cloud, scene_info.train_cameras, self.cameras_extent)
def save(self, iteration):
point_cloud_path = os.path.join(self.model_path, "point_cloud/iteration_{}".format(iteration))
self.gaussians.save_ply(os.path.join(point_cloud_path, "point_cloud.ply"))
exposure_dict = {
image_name: self.gaussians.get_exposure_from_name(image_name).detach().cpu().numpy().tolist()
for image_name in self.gaussians.exposure_mapping
}
with open(os.path.join(self.model_path, "exposure.json"), "w") as f:
json.dump(exposure_dict, f, indent=2)
def getTrainCameras(self, scale=1.0):
return self.train_cameras[scale]
def getTestCameras(self, scale=1.0):
return self.test_cameras[scale]
================================================
FILE: scene/cameras.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import torch
from torch import nn
import numpy as np
from utils.graphics_utils import getWorld2View2, getProjectionMatrix
from utils.general_utils import PILtoTorch
import cv2
class Camera(nn.Module):
def __init__(self, resolution, colmap_id, R, T, FoVx, FoVy, depth_params, image, invdepthmap,
image_name, uid,
trans=np.array([0.0, 0.0, 0.0]), scale=1.0, data_device = "cuda",
train_test_exp = False, is_test_dataset = False, is_test_view = False
):
super(Camera, self).__init__()
self.uid = uid
self.colmap_id = colmap_id
self.R = R
self.T = T
self.FoVx = FoVx
self.FoVy = FoVy
self.image_name = image_name
try:
self.data_device = torch.device(data_device)
except Exception as e:
print(e)
print(f"[Warning] Custom device {data_device} failed, fallback to default cuda device" )
self.data_device = torch.device("cuda")
resized_image_rgb = PILtoTorch(image, resolution)
gt_image = resized_image_rgb[:3, ...]
self.alpha_mask = None
if resized_image_rgb.shape[0] == 4:
self.alpha_mask = resized_image_rgb[3:4, ...].to(self.data_device)
else:
self.alpha_mask = torch.ones_like(resized_image_rgb[0:1, ...].to(self.data_device))
if train_test_exp and is_test_view:
if is_test_dataset:
self.alpha_mask[..., :self.alpha_mask.shape[-1] // 2] = 0
else:
self.alpha_mask[..., self.alpha_mask.shape[-1] // 2:] = 0
self.original_image = gt_image.clamp(0.0, 1.0).to(self.data_device)
self.image_width = self.original_image.shape[2]
self.image_height = self.original_image.shape[1]
self.invdepthmap = None
self.depth_reliable = False
if invdepthmap is not None:
self.depth_mask = torch.ones_like(self.alpha_mask)
self.invdepthmap = cv2.resize(invdepthmap, resolution)
self.invdepthmap[self.invdepthmap < 0] = 0
self.depth_reliable = True
if depth_params is not None:
if depth_params["scale"] < 0.2 * depth_params["med_scale"] or depth_params["scale"] > 5 * depth_params["med_scale"]:
self.depth_reliable = False
self.depth_mask *= 0
if depth_params["scale"] > 0:
self.invdepthmap = self.invdepthmap * depth_params["scale"] + depth_params["offset"]
if self.invdepthmap.ndim != 2:
self.invdepthmap = self.invdepthmap[..., 0]
self.invdepthmap = torch.from_numpy(self.invdepthmap[None]).to(self.data_device)
self.zfar = 100.0
self.znear = 0.01
self.trans = trans
self.scale = scale
self.world_view_transform = torch.tensor(getWorld2View2(R, T, trans, scale)).transpose(0, 1).cuda()
self.projection_matrix = getProjectionMatrix(znear=self.znear, zfar=self.zfar, fovX=self.FoVx, fovY=self.FoVy).transpose(0,1).cuda()
self.full_proj_transform = (self.world_view_transform.unsqueeze(0).bmm(self.projection_matrix.unsqueeze(0))).squeeze(0)
self.camera_center = self.world_view_transform.inverse()[3, :3]
class MiniCam:
def __init__(self, width, height, fovy, fovx, znear, zfar, world_view_transform, full_proj_transform):
self.image_width = width
self.image_height = height
self.FoVy = fovy
self.FoVx = fovx
self.znear = znear
self.zfar = zfar
self.world_view_transform = world_view_transform
self.full_proj_transform = full_proj_transform
view_inv = torch.inverse(self.world_view_transform)
self.camera_center = view_inv[3][:3]
================================================
FILE: scene/colmap_loader.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import numpy as np
import collections
import struct
CameraModel = collections.namedtuple(
"CameraModel", ["model_id", "model_name", "num_params"])
Camera = collections.namedtuple(
"Camera", ["id", "model", "width", "height", "params"])
BaseImage = collections.namedtuple(
"Image", ["id", "qvec", "tvec", "camera_id", "name", "xys", "point3D_ids"])
Point3D = collections.namedtuple(
"Point3D", ["id", "xyz", "rgb", "error", "image_ids", "point2D_idxs"])
CAMERA_MODELS = {
CameraModel(model_id=0, model_name="SIMPLE_PINHOLE", num_params=3),
CameraModel(model_id=1, model_name="PINHOLE", num_params=4),
CameraModel(model_id=2, model_name="SIMPLE_RADIAL", num_params=4),
CameraModel(model_id=3, model_name="RADIAL", num_params=5),
CameraModel(model_id=4, model_name="OPENCV", num_params=8),
CameraModel(model_id=5, model_name="OPENCV_FISHEYE", num_params=8),
CameraModel(model_id=6, model_name="FULL_OPENCV", num_params=12),
CameraModel(model_id=7, model_name="FOV", num_params=5),
CameraModel(model_id=8, model_name="SIMPLE_RADIAL_FISHEYE", num_params=4),
CameraModel(model_id=9, model_name="RADIAL_FISHEYE", num_params=5),
CameraModel(model_id=10, model_name="THIN_PRISM_FISHEYE", num_params=12)
}
CAMERA_MODEL_IDS = dict([(camera_model.model_id, camera_model)
for camera_model in CAMERA_MODELS])
CAMERA_MODEL_NAMES = dict([(camera_model.model_name, camera_model)
for camera_model in CAMERA_MODELS])
def qvec2rotmat(qvec):
return np.array([
[1 - 2 * qvec[2]**2 - 2 * qvec[3]**2,
2 * qvec[1] * qvec[2] - 2 * qvec[0] * qvec[3],
2 * qvec[3] * qvec[1] + 2 * qvec[0] * qvec[2]],
[2 * qvec[1] * qvec[2] + 2 * qvec[0] * qvec[3],
1 - 2 * qvec[1]**2 - 2 * qvec[3]**2,
2 * qvec[2] * qvec[3] - 2 * qvec[0] * qvec[1]],
[2 * qvec[3] * qvec[1] - 2 * qvec[0] * qvec[2],
2 * qvec[2] * qvec[3] + 2 * qvec[0] * qvec[1],
1 - 2 * qvec[1]**2 - 2 * qvec[2]**2]])
def rotmat2qvec(R):
Rxx, Ryx, Rzx, Rxy, Ryy, Rzy, Rxz, Ryz, Rzz = R.flat
K = np.array([
[Rxx - Ryy - Rzz, 0, 0, 0],
[Ryx + Rxy, Ryy - Rxx - Rzz, 0, 0],
[Rzx + Rxz, Rzy + Ryz, Rzz - Rxx - Ryy, 0],
[Ryz - Rzy, Rzx - Rxz, Rxy - Ryx, Rxx + Ryy + Rzz]]) / 3.0
eigvals, eigvecs = np.linalg.eigh(K)
qvec = eigvecs[[3, 0, 1, 2], np.argmax(eigvals)]
if qvec[0] < 0:
qvec *= -1
return qvec
class Image(BaseImage):
def qvec2rotmat(self):
return qvec2rotmat(self.qvec)
def read_next_bytes(fid, num_bytes, format_char_sequence, endian_character="<"):
"""Read and unpack the next bytes from a binary file.
:param fid:
:param num_bytes: Sum of combination of {2, 4, 8}, e.g. 2, 6, 16, 30, etc.
:param format_char_sequence: List of {c, e, f, d, h, H, i, I, l, L, q, Q}.
:param endian_character: Any of {@, =, <, >, !}
:return: Tuple of read and unpacked values.
"""
data = fid.read(num_bytes)
return struct.unpack(endian_character + format_char_sequence, data)
def read_points3D_text(path):
"""
see: src/base/reconstruction.cc
void Reconstruction::ReadPoints3DText(const std::string& path)
void Reconstruction::WritePoints3DText(const std::string& path)
"""
xyzs = None
rgbs = None
errors = None
num_points = 0
with open(path, "r") as fid:
while True:
line = fid.readline()
if not line:
break
line = line.strip()
if len(line) > 0 and line[0] != "#":
num_points += 1
xyzs = np.empty((num_points, 3))
rgbs = np.empty((num_points, 3))
errors = np.empty((num_points, 1))
count = 0
with open(path, "r") as fid:
while True:
line = fid.readline()
if not line:
break
line = line.strip()
if len(line) > 0 and line[0] != "#":
elems = line.split()
xyz = np.array(tuple(map(float, elems[1:4])))
rgb = np.array(tuple(map(int, elems[4:7])))
error = np.array(float(elems[7]))
xyzs[count] = xyz
rgbs[count] = rgb
errors[count] = error
count += 1
return xyzs, rgbs, errors
def read_points3D_binary(path_to_model_file):
"""
see: src/base/reconstruction.cc
void Reconstruction::ReadPoints3DBinary(const std::string& path)
void Reconstruction::WritePoints3DBinary(const std::string& path)
"""
with open(path_to_model_file, "rb") as fid:
num_points = read_next_bytes(fid, 8, "Q")[0]
xyzs = np.empty((num_points, 3))
rgbs = np.empty((num_points, 3))
errors = np.empty((num_points, 1))
for p_id in range(num_points):
binary_point_line_properties = read_next_bytes(
fid, num_bytes=43, format_char_sequence="QdddBBBd")
xyz = np.array(binary_point_line_properties[1:4])
rgb = np.array(binary_point_line_properties[4:7])
error = np.array(binary_point_line_properties[7])
track_length = read_next_bytes(
fid, num_bytes=8, format_char_sequence="Q")[0]
track_elems = read_next_bytes(
fid, num_bytes=8*track_length,
format_char_sequence="ii"*track_length)
xyzs[p_id] = xyz
rgbs[p_id] = rgb
errors[p_id] = error
return xyzs, rgbs, errors
def read_intrinsics_text(path):
"""
Taken from https://github.com/colmap/colmap/blob/dev/scripts/python/read_write_model.py
"""
cameras = {}
with open(path, "r") as fid:
while True:
line = fid.readline()
if not line:
break
line = line.strip()
if len(line) > 0 and line[0] != "#":
elems = line.split()
camera_id = int(elems[0])
model = elems[1]
assert model == "PINHOLE", "While the loader support other types, the rest of the code assumes PINHOLE"
width = int(elems[2])
height = int(elems[3])
params = np.array(tuple(map(float, elems[4:])))
cameras[camera_id] = Camera(id=camera_id, model=model,
width=width, height=height,
params=params)
return cameras
def read_extrinsics_binary(path_to_model_file):
"""
see: src/base/reconstruction.cc
void Reconstruction::ReadImagesBinary(const std::string& path)
void Reconstruction::WriteImagesBinary(const std::string& path)
"""
images = {}
with open(path_to_model_file, "rb") as fid:
num_reg_images = read_next_bytes(fid, 8, "Q")[0]
for _ in range(num_reg_images):
binary_image_properties = read_next_bytes(
fid, num_bytes=64, format_char_sequence="idddddddi")
image_id = binary_image_properties[0]
qvec = np.array(binary_image_properties[1:5])
tvec = np.array(binary_image_properties[5:8])
camera_id = binary_image_properties[8]
image_name = ""
current_char = read_next_bytes(fid, 1, "c")[0]
while current_char != b"\x00": # look for the ASCII 0 entry
image_name += current_char.decode("utf-8")
current_char = read_next_bytes(fid, 1, "c")[0]
num_points2D = read_next_bytes(fid, num_bytes=8,
format_char_sequence="Q")[0]
x_y_id_s = read_next_bytes(fid, num_bytes=24*num_points2D,
format_char_sequence="ddq"*num_points2D)
xys = np.column_stack([tuple(map(float, x_y_id_s[0::3])),
tuple(map(float, x_y_id_s[1::3]))])
point3D_ids = np.array(tuple(map(int, x_y_id_s[2::3])))
images[image_id] = Image(
id=image_id, qvec=qvec, tvec=tvec,
camera_id=camera_id, name=image_name,
xys=xys, point3D_ids=point3D_ids)
return images
def read_intrinsics_binary(path_to_model_file):
"""
see: src/base/reconstruction.cc
void Reconstruction::WriteCamerasBinary(const std::string& path)
void Reconstruction::ReadCamerasBinary(const std::string& path)
"""
cameras = {}
with open(path_to_model_file, "rb") as fid:
num_cameras = read_next_bytes(fid, 8, "Q")[0]
for _ in range(num_cameras):
camera_properties = read_next_bytes(
fid, num_bytes=24, format_char_sequence="iiQQ")
camera_id = camera_properties[0]
model_id = camera_properties[1]
model_name = CAMERA_MODEL_IDS[camera_properties[1]].model_name
width = camera_properties[2]
height = camera_properties[3]
num_params = CAMERA_MODEL_IDS[model_id].num_params
params = read_next_bytes(fid, num_bytes=8*num_params,
format_char_sequence="d"*num_params)
cameras[camera_id] = Camera(id=camera_id,
model=model_name,
width=width,
height=height,
params=np.array(params))
assert len(cameras) == num_cameras
return cameras
def read_extrinsics_text(path):
"""
Taken from https://github.com/colmap/colmap/blob/dev/scripts/python/read_write_model.py
"""
images = {}
with open(path, "r") as fid:
while True:
line = fid.readline()
if not line:
break
line = line.strip()
if len(line) > 0 and line[0] != "#":
elems = line.split()
image_id = int(elems[0])
qvec = np.array(tuple(map(float, elems[1:5])))
tvec = np.array(tuple(map(float, elems[5:8])))
camera_id = int(elems[8])
image_name = elems[9]
elems = fid.readline().split()
xys = np.column_stack([tuple(map(float, elems[0::3])),
tuple(map(float, elems[1::3]))])
point3D_ids = np.array(tuple(map(int, elems[2::3])))
images[image_id] = Image(
id=image_id, qvec=qvec, tvec=tvec,
camera_id=camera_id, name=image_name,
xys=xys, point3D_ids=point3D_ids)
return images
def read_colmap_bin_array(path):
"""
Taken from https://github.com/colmap/colmap/blob/dev/scripts/python/read_dense.py
:param path: path to the colmap binary file.
:return: nd array with the floating point values in the value
"""
with open(path, "rb") as fid:
width, height, channels = np.genfromtxt(fid, delimiter="&", max_rows=1,
usecols=(0, 1, 2), dtype=int)
fid.seek(0)
num_delimiter = 0
byte = fid.read(1)
while True:
if byte == b"&":
num_delimiter += 1
if num_delimiter >= 3:
break
byte = fid.read(1)
array = np.fromfile(fid, np.float32)
array = array.reshape((width, height, channels), order="F")
return np.transpose(array, (1, 0, 2)).squeeze()
================================================
FILE: scene/dataset_readers.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import os
import sys
from PIL import Image
from typing import NamedTuple
from scene.colmap_loader import read_extrinsics_text, read_intrinsics_text, qvec2rotmat, \
read_extrinsics_binary, read_intrinsics_binary, read_points3D_binary, read_points3D_text
from utils.graphics_utils import getWorld2View2, focal2fov, fov2focal
import numpy as np
import json
from pathlib import Path
from plyfile import PlyData, PlyElement
from utils.sh_utils import SH2RGB
from scene.gaussian_model import BasicPointCloud
class CameraInfo(NamedTuple):
uid: int
R: np.array
T: np.array
FovY: np.array
FovX: np.array
depth_params: dict
image_path: str
image_name: str
depth_path: str
width: int
height: int
is_test: bool
class SceneInfo(NamedTuple):
point_cloud: BasicPointCloud
train_cameras: list
test_cameras: list
nerf_normalization: dict
ply_path: str
is_nerf_synthetic: bool
def getNerfppNorm(cam_info):
def get_center_and_diag(cam_centers):
cam_centers = np.hstack(cam_centers)
avg_cam_center = np.mean(cam_centers, axis=1, keepdims=True)
center = avg_cam_center
dist = np.linalg.norm(cam_centers - center, axis=0, keepdims=True)
diagonal = np.max(dist)
return center.flatten(), diagonal
cam_centers = []
for cam in cam_info:
W2C = getWorld2View2(cam.R, cam.T)
C2W = np.linalg.inv(W2C)
cam_centers.append(C2W[:3, 3:4])
center, diagonal = get_center_and_diag(cam_centers)
radius = diagonal * 1.1
translate = -center
return {"translate": translate, "radius": radius}
def readColmapCameras(cam_extrinsics, cam_intrinsics, depths_params, images_folder, depths_folder, test_cam_names_list):
cam_infos = []
for idx, key in enumerate(cam_extrinsics):
sys.stdout.write('\r')
# the exact output you're looking for:
sys.stdout.write("Reading camera {}/{}".format(idx+1, len(cam_extrinsics)))
sys.stdout.flush()
extr = cam_extrinsics[key]
intr = cam_intrinsics[extr.camera_id]
height = intr.height
width = intr.width
uid = intr.id
R = np.transpose(qvec2rotmat(extr.qvec))
T = np.array(extr.tvec)
if intr.model=="SIMPLE_PINHOLE":
focal_length_x = intr.params[0]
FovY = focal2fov(focal_length_x, height)
FovX = focal2fov(focal_length_x, width)
elif intr.model=="PINHOLE":
focal_length_x = intr.params[0]
focal_length_y = intr.params[1]
FovY = focal2fov(focal_length_y, height)
FovX = focal2fov(focal_length_x, width)
else:
assert False, "Colmap camera model not handled: only undistorted datasets (PINHOLE or SIMPLE_PINHOLE cameras) supported!"
n_remove = len(extr.name.split('.')[-1]) + 1
depth_params = None
if depths_params is not None:
try:
depth_params = depths_params[extr.name[:-n_remove]]
except:
print("\n", key, "not found in depths_params")
image_path = os.path.join(images_folder, extr.name)
image_name = extr.name
depth_path = os.path.join(depths_folder, f"{extr.name[:-n_remove]}.png") if depths_folder != "" else ""
cam_info = CameraInfo(uid=uid, R=R, T=T, FovY=FovY, FovX=FovX, depth_params=depth_params,
image_path=image_path, image_name=image_name, depth_path=depth_path,
width=width, height=height, is_test=image_name in test_cam_names_list)
cam_infos.append(cam_info)
sys.stdout.write('\n')
return cam_infos
def fetchPly(path):
plydata = PlyData.read(path)
vertices = plydata['vertex']
positions = np.vstack([vertices['x'], vertices['y'], vertices['z']]).T
colors = np.vstack([vertices['red'], vertices['green'], vertices['blue']]).T / 255.0
normals = np.vstack([vertices['nx'], vertices['ny'], vertices['nz']]).T
return BasicPointCloud(points=positions, colors=colors, normals=normals)
def storePly(path, xyz, rgb):
# Define the dtype for the structured array
dtype = [('x', 'f4'), ('y', 'f4'), ('z', 'f4'),
('nx', 'f4'), ('ny', 'f4'), ('nz', 'f4'),
('red', 'u1'), ('green', 'u1'), ('blue', 'u1')]
normals = np.zeros_like(xyz)
elements = np.empty(xyz.shape[0], dtype=dtype)
attributes = np.concatenate((xyz, normals, rgb), axis=1)
elements[:] = list(map(tuple, attributes))
# Create the PlyData object and write to file
vertex_element = PlyElement.describe(elements, 'vertex')
ply_data = PlyData([vertex_element])
ply_data.write(path)
def readColmapSceneInfo(path, images, depths, eval, train_test_exp, llffhold=8):
try:
cameras_extrinsic_file = os.path.join(path, "sparse/0", "images.bin")
cameras_intrinsic_file = os.path.join(path, "sparse/0", "cameras.bin")
cam_extrinsics = read_extrinsics_binary(cameras_extrinsic_file)
cam_intrinsics = read_intrinsics_binary(cameras_intrinsic_file)
except:
cameras_extrinsic_file = os.path.join(path, "sparse/0", "images.txt")
cameras_intrinsic_file = os.path.join(path, "sparse/0", "cameras.txt")
cam_extrinsics = read_extrinsics_text(cameras_extrinsic_file)
cam_intrinsics = read_intrinsics_text(cameras_intrinsic_file)
depth_params_file = os.path.join(path, "sparse/0", "depth_params.json")
## if depth_params_file isnt there AND depths file is here -> throw error
depths_params = None
if depths != "":
try:
with open(depth_params_file, "r") as f:
depths_params = json.load(f)
all_scales = np.array([depths_params[key]["scale"] for key in depths_params])
if (all_scales > 0).sum():
med_scale = np.median(all_scales[all_scales > 0])
else:
med_scale = 0
for key in depths_params:
depths_params[key]["med_scale"] = med_scale
except FileNotFoundError:
print(f"Error: depth_params.json file not found at path '{depth_params_file}'.")
sys.exit(1)
except Exception as e:
print(f"An unexpected error occurred when trying to open depth_params.json file: {e}")
sys.exit(1)
if eval:
if "360" in path:
llffhold = 8
if llffhold:
print("------------LLFF HOLD-------------")
cam_names = [cam_extrinsics[cam_id].name for cam_id in cam_extrinsics]
cam_names = sorted(cam_names)
test_cam_names_list = [name for idx, name in enumerate(cam_names) if idx % llffhold == 0]
else:
with open(os.path.join(path, "sparse/0", "test.txt"), 'r') as file:
test_cam_names_list = [line.strip() for line in file]
else:
test_cam_names_list = []
reading_dir = "images" if images == None else images
cam_infos_unsorted = readColmapCameras(
cam_extrinsics=cam_extrinsics, cam_intrinsics=cam_intrinsics, depths_params=depths_params,
images_folder=os.path.join(path, reading_dir),
depths_folder=os.path.join(path, depths) if depths != "" else "", test_cam_names_list=test_cam_names_list)
cam_infos = sorted(cam_infos_unsorted.copy(), key = lambda x : x.image_name)
train_cam_infos = [c for c in cam_infos if train_test_exp or not c.is_test]
test_cam_infos = [c for c in cam_infos if c.is_test]
nerf_normalization = getNerfppNorm(train_cam_infos)
ply_path = os.path.join(path, "sparse/0/points3D.ply")
bin_path = os.path.join(path, "sparse/0/points3D.bin")
txt_path = os.path.join(path, "sparse/0/points3D.txt")
if not os.path.exists(ply_path):
print("Converting point3d.bin to .ply, will happen only the first time you open the scene.")
try:
xyz, rgb, _ = read_points3D_binary(bin_path)
except:
xyz, rgb, _ = read_points3D_text(txt_path)
storePly(ply_path, xyz, rgb)
try:
pcd = fetchPly(ply_path)
except:
pcd = None
scene_info = SceneInfo(point_cloud=pcd,
train_cameras=train_cam_infos,
test_cameras=test_cam_infos,
nerf_normalization=nerf_normalization,
ply_path=ply_path,
is_nerf_synthetic=False)
return scene_info
def readCamerasFromTransforms(path, transformsfile, depths_folder, white_background, is_test, extension=".png"):
cam_infos = []
with open(os.path.join(path, transformsfile)) as json_file:
contents = json.load(json_file)
fovx = contents["camera_angle_x"]
frames = contents["frames"]
for idx, frame in enumerate(frames):
cam_name = os.path.join(path, frame["file_path"] + extension)
# NeRF 'transform_matrix' is a camera-to-world transform
c2w = np.array(frame["transform_matrix"])
# change from OpenGL/Blender camera axes (Y up, Z back) to COLMAP (Y down, Z forward)
c2w[:3, 1:3] *= -1
# get the world-to-camera transform and set R, T
w2c = np.linalg.inv(c2w)
R = np.transpose(w2c[:3,:3]) # R is stored transposed due to 'glm' in CUDA code
T = w2c[:3, 3]
image_path = os.path.join(path, cam_name)
image_name = Path(cam_name).stem
image = Image.open(image_path)
im_data = np.array(image.convert("RGBA"))
bg = np.array([1,1,1]) if white_background else np.array([0, 0, 0])
norm_data = im_data / 255.0
arr = norm_data[:,:,:3] * norm_data[:, :, 3:4] + bg * (1 - norm_data[:, :, 3:4])
image = Image.fromarray(np.array(arr*255.0, dtype=np.byte), "RGB")
fovy = focal2fov(fov2focal(fovx, image.size[0]), image.size[1])
FovY = fovy
FovX = fovx
depth_path = os.path.join(depths_folder, f"{image_name}.png") if depths_folder != "" else ""
cam_infos.append(CameraInfo(uid=idx, R=R, T=T, FovY=FovY, FovX=FovX,
image_path=image_path, image_name=image_name,
width=image.size[0], height=image.size[1], depth_path=depth_path, depth_params=None, is_test=is_test))
return cam_infos
def readNerfSyntheticInfo(path, white_background, depths, eval, extension=".png"):
depths_folder=os.path.join(path, depths) if depths != "" else ""
print("Reading Training Transforms")
train_cam_infos = readCamerasFromTransforms(path, "transforms_train.json", depths_folder, white_background, False, extension)
print("Reading Test Transforms")
test_cam_infos = readCamerasFromTransforms(path, "transforms_test.json", depths_folder, white_background, True, extension)
if not eval:
train_cam_infos.extend(test_cam_infos)
test_cam_infos = []
nerf_normalization = getNerfppNorm(train_cam_infos)
ply_path = os.path.join(path, "points3d.ply")
if not os.path.exists(ply_path):
# Since this data set has no colmap data, we start with random points
num_pts = 100_000
print(f"Generating random point cloud ({num_pts})...")
# We create random points inside the bounds of the synthetic Blender scenes
xyz = np.random.random((num_pts, 3)) * 2.6 - 1.3
shs = np.random.random((num_pts, 3)) / 255.0
pcd = BasicPointCloud(points=xyz, colors=SH2RGB(shs), normals=np.zeros((num_pts, 3)))
storePly(ply_path, xyz, SH2RGB(shs) * 255)
try:
pcd = fetchPly(ply_path)
except:
pcd = None
scene_info = SceneInfo(point_cloud=pcd,
train_cameras=train_cam_infos,
test_cameras=test_cam_infos,
nerf_normalization=nerf_normalization,
ply_path=ply_path,
is_nerf_synthetic=True)
return scene_info
sceneLoadTypeCallbacks = {
"Colmap": readColmapSceneInfo,
"Blender" : readNerfSyntheticInfo
}
================================================
FILE: scene/gaussian_model.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import torch
import numpy as np
from utils.general_utils import inverse_sigmoid, get_expon_lr_func, build_rotation
from torch import nn
import os
import json
from utils.system_utils import mkdir_p
from plyfile import PlyData, PlyElement
from utils.sh_utils import RGB2SH
from simple_knn._C import distCUDA2
from utils.graphics_utils import BasicPointCloud
from utils.general_utils import strip_symmetric, build_scaling_rotation
try:
from diff_gaussian_rasterization import SparseGaussianAdam
except:
pass
class GaussianModel:
def setup_functions(self):
def build_covariance_from_scaling_rotation(scaling, scaling_modifier, rotation):
L = build_scaling_rotation(scaling_modifier * scaling, rotation)
actual_covariance = L @ L.transpose(1, 2)
symm = strip_symmetric(actual_covariance)
return symm
self.scaling_activation = torch.exp
self.scaling_inverse_activation = torch.log
self.covariance_activation = build_covariance_from_scaling_rotation
self.opacity_activation = torch.sigmoid
self.inverse_opacity_activation = inverse_sigmoid
self.rotation_activation = torch.nn.functional.normalize
def __init__(self, sh_degree, optimizer_type="default"):
self.active_sh_degree = 0
self.optimizer_type = optimizer_type
self.max_sh_degree = sh_degree
self._xyz = torch.empty(0)
self._features_dc = torch.empty(0)
self._features_rest = torch.empty(0)
self._scaling = torch.empty(0)
self._rotation = torch.empty(0)
self._opacity = torch.empty(0)
self.max_radii2D = torch.empty(0)
self.xyz_gradient_accum = torch.empty(0)
self.denom = torch.empty(0)
self.optimizer = None
self.percent_dense = 0
self.spatial_lr_scale = 0
self.setup_functions()
def capture(self):
return (
self.active_sh_degree,
self._xyz,
self._features_dc,
self._features_rest,
self._scaling,
self._rotation,
self._opacity,
self.max_radii2D,
self.xyz_gradient_accum,
self.denom,
self.optimizer.state_dict(),
self.spatial_lr_scale,
)
def restore(self, model_args, training_args):
(self.active_sh_degree,
self._xyz,
self._features_dc,
self._features_rest,
self._scaling,
self._rotation,
self._opacity,
self.max_radii2D,
xyz_gradient_accum,
denom,
opt_dict,
self.spatial_lr_scale) = model_args
self.training_setup(training_args)
self.xyz_gradient_accum = xyz_gradient_accum
self.denom = denom
self.optimizer.load_state_dict(opt_dict)
@property
def get_scaling(self):
return self.scaling_activation(self._scaling)
@property
def get_rotation(self):
return self.rotation_activation(self._rotation)
@property
def get_xyz(self):
return self._xyz
@property
def get_features(self):
features_dc = self._features_dc
features_rest = self._features_rest
return torch.cat((features_dc, features_rest), dim=1)
@property
def get_features_dc(self):
return self._features_dc
@property
def get_features_rest(self):
return self._features_rest
@property
def get_opacity(self):
return self.opacity_activation(self._opacity)
@property
def get_exposure(self):
return self._exposure
def get_exposure_from_name(self, image_name):
if self.pretrained_exposures is None:
return self._exposure[self.exposure_mapping[image_name]]
else:
return self.pretrained_exposures[image_name]
def get_covariance(self, scaling_modifier = 1):
return self.covariance_activation(self.get_scaling, scaling_modifier, self._rotation)
def oneupSHdegree(self):
if self.active_sh_degree < self.max_sh_degree:
self.active_sh_degree += 1
def create_from_pcd(self, pcd : BasicPointCloud, cam_infos : int, spatial_lr_scale : float):
self.spatial_lr_scale = spatial_lr_scale
fused_point_cloud = torch.tensor(np.asarray(pcd.points)).float().cuda()
fused_color = RGB2SH(torch.tensor(np.asarray(pcd.colors)).float().cuda())
features = torch.zeros((fused_color.shape[0], 3, (self.max_sh_degree + 1) ** 2)).float().cuda()
features[:, :3, 0 ] = fused_color
features[:, 3:, 1:] = 0.0
print("Number of points at initialisation : ", fused_point_cloud.shape[0])
dist2 = torch.clamp_min(distCUDA2(torch.from_numpy(np.asarray(pcd.points)).float().cuda()), 0.0000001)
scales = torch.log(torch.sqrt(dist2))[...,None].repeat(1, 3)
rots = torch.zeros((fused_point_cloud.shape[0], 4), device="cuda")
rots[:, 0] = 1
opacities = self.inverse_opacity_activation(0.1 * torch.ones((fused_point_cloud.shape[0], 1), dtype=torch.float, device="cuda"))
self._xyz = nn.Parameter(fused_point_cloud.requires_grad_(True))
self._features_dc = nn.Parameter(features[:,:,0:1].transpose(1, 2).contiguous().requires_grad_(True))
self._features_rest = nn.Parameter(features[:,:,1:].transpose(1, 2).contiguous().requires_grad_(True))
self._scaling = nn.Parameter(scales.requires_grad_(True))
self._rotation = nn.Parameter(rots.requires_grad_(True))
self._opacity = nn.Parameter(opacities.requires_grad_(True))
self.max_radii2D = torch.zeros((self.get_xyz.shape[0]), device="cuda")
self.exposure_mapping = {cam_info.image_name: idx for idx, cam_info in enumerate(cam_infos)}
self.pretrained_exposures = None
exposure = torch.eye(3, 4, device="cuda")[None].repeat(len(cam_infos), 1, 1)
self._exposure = nn.Parameter(exposure.requires_grad_(True))
def training_setup(self, training_args):
self.percent_dense = training_args.percent_dense
self.xyz_gradient_accum = torch.zeros((self.get_xyz.shape[0], 1), device="cuda")
self.denom = torch.zeros((self.get_xyz.shape[0], 1), device="cuda")
l = [
{'params': [self._xyz], 'lr': training_args.position_lr_init * self.spatial_lr_scale, "name": "xyz"},
{'params': [self._features_dc], 'lr': training_args.feature_lr, "name": "f_dc"},
{'params': [self._features_rest], 'lr': training_args.feature_lr / 20.0, "name": "f_rest"},
{'params': [self._opacity], 'lr': training_args.opacity_lr, "name": "opacity"},
{'params': [self._scaling], 'lr': training_args.scaling_lr, "name": "scaling"},
{'params': [self._rotation], 'lr': training_args.rotation_lr, "name": "rotation"}
]
if self.optimizer_type == "default":
self.optimizer = torch.optim.Adam(l, lr=0.0, eps=1e-15)
elif self.optimizer_type == "sparse_adam":
try:
self.optimizer = SparseGaussianAdam(l, lr=0.0, eps=1e-15)
except:
# A special version of the rasterizer is required to enable sparse adam
self.optimizer = torch.optim.Adam(l, lr=0.0, eps=1e-15)
self.exposure_optimizer = torch.optim.Adam([self._exposure])
self.xyz_scheduler_args = get_expon_lr_func(lr_init=training_args.position_lr_init*self.spatial_lr_scale,
lr_final=training_args.position_lr_final*self.spatial_lr_scale,
lr_delay_mult=training_args.position_lr_delay_mult,
max_steps=training_args.position_lr_max_steps)
self.exposure_scheduler_args = get_expon_lr_func(training_args.exposure_lr_init, training_args.exposure_lr_final,
lr_delay_steps=training_args.exposure_lr_delay_steps,
lr_delay_mult=training_args.exposure_lr_delay_mult,
max_steps=training_args.iterations)
def update_learning_rate(self, iteration):
''' Learning rate scheduling per step '''
if self.pretrained_exposures is None:
for param_group in self.exposure_optimizer.param_groups:
param_group['lr'] = self.exposure_scheduler_args(iteration)
for param_group in self.optimizer.param_groups:
if param_group["name"] == "xyz":
lr = self.xyz_scheduler_args(iteration)
param_group['lr'] = lr
return lr
def construct_list_of_attributes(self):
l = ['x', 'y', 'z', 'nx', 'ny', 'nz']
# All channels except the 3 DC
for i in range(self._features_dc.shape[1]*self._features_dc.shape[2]):
l.append('f_dc_{}'.format(i))
for i in range(self._features_rest.shape[1]*self._features_rest.shape[2]):
l.append('f_rest_{}'.format(i))
l.append('opacity')
for i in range(self._scaling.shape[1]):
l.append('scale_{}'.format(i))
for i in range(self._rotation.shape[1]):
l.append('rot_{}'.format(i))
return l
def save_ply(self, path):
mkdir_p(os.path.dirname(path))
xyz = self._xyz.detach().cpu().numpy()
normals = np.zeros_like(xyz)
f_dc = self._features_dc.detach().transpose(1, 2).flatten(start_dim=1).contiguous().cpu().numpy()
f_rest = self._features_rest.detach().transpose(1, 2).flatten(start_dim=1).contiguous().cpu().numpy()
opacities = self._opacity.detach().cpu().numpy()
scale = self._scaling.detach().cpu().numpy()
rotation = self._rotation.detach().cpu().numpy()
dtype_full = [(attribute, 'f4') for attribute in self.construct_list_of_attributes()]
elements = np.empty(xyz.shape[0], dtype=dtype_full)
attributes = np.concatenate((xyz, normals, f_dc, f_rest, opacities, scale, rotation), axis=1)
elements[:] = list(map(tuple, attributes))
el = PlyElement.describe(elements, 'vertex')
PlyData([el]).write(path)
def reset_opacity(self):
opacities_new = self.inverse_opacity_activation(torch.min(self.get_opacity, torch.ones_like(self.get_opacity)*0.01))
optimizable_tensors = self.replace_tensor_to_optimizer(opacities_new, "opacity")
self._opacity = optimizable_tensors["opacity"]
def load_ply(self, path, use_train_test_exp = False):
plydata = PlyData.read(path)
if use_train_test_exp:
exposure_file = os.path.join(os.path.dirname(path), os.pardir, os.pardir, "exposure.json")
if os.path.exists(exposure_file):
with open(exposure_file, "r") as f:
exposures = json.load(f)
self.pretrained_exposures = {image_name: torch.FloatTensor(exposures[image_name]).requires_grad_(False).cuda() for image_name in exposures}
print(f"Pretrained exposures loaded.")
else:
print(f"No exposure to be loaded at {exposure_file}")
self.pretrained_exposures = None
xyz = np.stack((np.asarray(plydata.elements[0]["x"]),
np.asarray(plydata.elements[0]["y"]),
np.asarray(plydata.elements[0]["z"])), axis=1)
opacities = np.asarray(plydata.elements[0]["opacity"])[..., np.newaxis]
features_dc = np.zeros((xyz.shape[0], 3, 1))
features_dc[:, 0, 0] = np.asarray(plydata.elements[0]["f_dc_0"])
features_dc[:, 1, 0] = np.asarray(plydata.elements[0]["f_dc_1"])
features_dc[:, 2, 0] = np.asarray(plydata.elements[0]["f_dc_2"])
extra_f_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("f_rest_")]
extra_f_names = sorted(extra_f_names, key = lambda x: int(x.split('_')[-1]))
assert len(extra_f_names)==3*(self.max_sh_degree + 1) ** 2 - 3
features_extra = np.zeros((xyz.shape[0], len(extra_f_names)))
for idx, attr_name in enumerate(extra_f_names):
features_extra[:, idx] = np.asarray(plydata.elements[0][attr_name])
# Reshape (P,F*SH_coeffs) to (P, F, SH_coeffs except DC)
features_extra = features_extra.reshape((features_extra.shape[0], 3, (self.max_sh_degree + 1) ** 2 - 1))
scale_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("scale_")]
scale_names = sorted(scale_names, key = lambda x: int(x.split('_')[-1]))
scales = np.zeros((xyz.shape[0], len(scale_names)))
for idx, attr_name in enumerate(scale_names):
scales[:, idx] = np.asarray(plydata.elements[0][attr_name])
rot_names = [p.name for p in plydata.elements[0].properties if p.name.startswith("rot")]
rot_names = sorted(rot_names, key = lambda x: int(x.split('_')[-1]))
rots = np.zeros((xyz.shape[0], len(rot_names)))
for idx, attr_name in enumerate(rot_names):
rots[:, idx] = np.asarray(plydata.elements[0][attr_name])
self._xyz = nn.Parameter(torch.tensor(xyz, dtype=torch.float, device="cuda").requires_grad_(True))
self._features_dc = nn.Parameter(torch.tensor(features_dc, dtype=torch.float, device="cuda").transpose(1, 2).contiguous().requires_grad_(True))
self._features_rest = nn.Parameter(torch.tensor(features_extra, dtype=torch.float, device="cuda").transpose(1, 2).contiguous().requires_grad_(True))
self._opacity = nn.Parameter(torch.tensor(opacities, dtype=torch.float, device="cuda").requires_grad_(True))
self._scaling = nn.Parameter(torch.tensor(scales, dtype=torch.float, device="cuda").requires_grad_(True))
self._rotation = nn.Parameter(torch.tensor(rots, dtype=torch.float, device="cuda").requires_grad_(True))
self.active_sh_degree = self.max_sh_degree
def replace_tensor_to_optimizer(self, tensor, name):
optimizable_tensors = {}
for group in self.optimizer.param_groups:
if group["name"] == name:
stored_state = self.optimizer.state.get(group['params'][0], None)
stored_state["exp_avg"] = torch.zeros_like(tensor)
stored_state["exp_avg_sq"] = torch.zeros_like(tensor)
del self.optimizer.state[group['params'][0]]
group["params"][0] = nn.Parameter(tensor.requires_grad_(True))
self.optimizer.state[group['params'][0]] = stored_state
optimizable_tensors[group["name"]] = group["params"][0]
return optimizable_tensors
def _prune_optimizer(self, mask):
optimizable_tensors = {}
for group in self.optimizer.param_groups:
stored_state = self.optimizer.state.get(group['params'][0], None)
if stored_state is not None:
stored_state["exp_avg"] = stored_state["exp_avg"][mask]
stored_state["exp_avg_sq"] = stored_state["exp_avg_sq"][mask]
del self.optimizer.state[group['params'][0]]
group["params"][0] = nn.Parameter((group["params"][0][mask].requires_grad_(True)))
self.optimizer.state[group['params'][0]] = stored_state
optimizable_tensors[group["name"]] = group["params"][0]
else:
group["params"][0] = nn.Parameter(group["params"][0][mask].requires_grad_(True))
optimizable_tensors[group["name"]] = group["params"][0]
return optimizable_tensors
def prune_points(self, mask):
valid_points_mask = ~mask
optimizable_tensors = self._prune_optimizer(valid_points_mask)
self._xyz = optimizable_tensors["xyz"]
self._features_dc = optimizable_tensors["f_dc"]
self._features_rest = optimizable_tensors["f_rest"]
self._opacity = optimizable_tensors["opacity"]
self._scaling = optimizable_tensors["scaling"]
self._rotation = optimizable_tensors["rotation"]
self.xyz_gradient_accum = self.xyz_gradient_accum[valid_points_mask]
self.denom = self.denom[valid_points_mask]
self.max_radii2D = self.max_radii2D[valid_points_mask]
self.tmp_radii = self.tmp_radii[valid_points_mask]
def cat_tensors_to_optimizer(self, tensors_dict):
optimizable_tensors = {}
for group in self.optimizer.param_groups:
assert len(group["params"]) == 1
extension_tensor = tensors_dict[group["name"]]
stored_state = self.optimizer.state.get(group['params'][0], None)
if stored_state is not None:
stored_state["exp_avg"] = torch.cat((stored_state["exp_avg"], torch.zeros_like(extension_tensor)), dim=0)
stored_state["exp_avg_sq"] = torch.cat((stored_state["exp_avg_sq"], torch.zeros_like(extension_tensor)), dim=0)
del self.optimizer.state[group['params'][0]]
group["params"][0] = nn.Parameter(torch.cat((group["params"][0], extension_tensor), dim=0).requires_grad_(True))
self.optimizer.state[group['params'][0]] = stored_state
optimizable_tensors[group["name"]] = group["params"][0]
else:
group["params"][0] = nn.Parameter(torch.cat((group["params"][0], extension_tensor), dim=0).requires_grad_(True))
optimizable_tensors[group["name"]] = group["params"][0]
return optimizable_tensors
def densification_postfix(self, new_xyz, new_features_dc, new_features_rest, new_opacities, new_scaling, new_rotation, new_tmp_radii):
d = {"xyz": new_xyz,
"f_dc": new_features_dc,
"f_rest": new_features_rest,
"opacity": new_opacities,
"scaling" : new_scaling,
"rotation" : new_rotation}
optimizable_tensors = self.cat_tensors_to_optimizer(d)
self._xyz = optimizable_tensors["xyz"]
self._features_dc = optimizable_tensors["f_dc"]
self._features_rest = optimizable_tensors["f_rest"]
self._opacity = optimizable_tensors["opacity"]
self._scaling = optimizable_tensors["scaling"]
self._rotation = optimizable_tensors["rotation"]
self.tmp_radii = torch.cat((self.tmp_radii, new_tmp_radii))
self.xyz_gradient_accum = torch.zeros((self.get_xyz.shape[0], 1), device="cuda")
self.denom = torch.zeros((self.get_xyz.shape[0], 1), device="cuda")
self.max_radii2D = torch.zeros((self.get_xyz.shape[0]), device="cuda")
def densify_and_split(self, grads, grad_threshold, scene_extent, N=2):
n_init_points = self.get_xyz.shape[0]
# Extract points that satisfy the gradient condition
padded_grad = torch.zeros((n_init_points), device="cuda")
padded_grad[:grads.shape[0]] = grads.squeeze()
selected_pts_mask = torch.where(padded_grad >= grad_threshold, True, False)
selected_pts_mask = torch.logical_and(selected_pts_mask,
torch.max(self.get_scaling, dim=1).values > self.percent_dense*scene_extent)
stds = self.get_scaling[selected_pts_mask].repeat(N,1)
means =torch.zeros((stds.size(0), 3),device="cuda")
samples = torch.normal(mean=means, std=stds)
rots = build_rotation(self._rotation[selected_pts_mask]).repeat(N,1,1)
new_xyz = torch.bmm(rots, samples.unsqueeze(-1)).squeeze(-1) + self.get_xyz[selected_pts_mask].repeat(N, 1)
new_scaling = self.scaling_inverse_activation(self.get_scaling[selected_pts_mask].repeat(N,1) / (0.8*N))
new_rotation = self._rotation[selected_pts_mask].repeat(N,1)
new_features_dc = self._features_dc[selected_pts_mask].repeat(N,1,1)
new_features_rest = self._features_rest[selected_pts_mask].repeat(N,1,1)
new_opacity = self._opacity[selected_pts_mask].repeat(N,1)
new_tmp_radii = self.tmp_radii[selected_pts_mask].repeat(N)
self.densification_postfix(new_xyz, new_features_dc, new_features_rest, new_opacity, new_scaling, new_rotation, new_tmp_radii)
prune_filter = torch.cat((selected_pts_mask, torch.zeros(N * selected_pts_mask.sum(), device="cuda", dtype=bool)))
self.prune_points(prune_filter)
def densify_and_clone(self, grads, grad_threshold, scene_extent):
# Extract points that satisfy the gradient condition
selected_pts_mask = torch.where(torch.norm(grads, dim=-1) >= grad_threshold, True, False)
selected_pts_mask = torch.logical_and(selected_pts_mask,
torch.max(self.get_scaling, dim=1).values <= self.percent_dense*scene_extent)
new_xyz = self._xyz[selected_pts_mask]
new_features_dc = self._features_dc[selected_pts_mask]
new_features_rest = self._features_rest[selected_pts_mask]
new_opacities = self._opacity[selected_pts_mask]
new_scaling = self._scaling[selected_pts_mask]
new_rotation = self._rotation[selected_pts_mask]
new_tmp_radii = self.tmp_radii[selected_pts_mask]
self.densification_postfix(new_xyz, new_features_dc, new_features_rest, new_opacities, new_scaling, new_rotation, new_tmp_radii)
def densify_and_prune(self, max_grad, min_opacity, extent, max_screen_size, radii):
grads = self.xyz_gradient_accum / self.denom
grads[grads.isnan()] = 0.0
self.tmp_radii = radii
self.densify_and_clone(grads, max_grad, extent)
self.densify_and_split(grads, max_grad, extent)
prune_mask = (self.get_opacity < min_opacity).squeeze()
if max_screen_size:
big_points_vs = self.max_radii2D > max_screen_size
big_points_ws = self.get_scaling.max(dim=1).values > 0.1 * extent
prune_mask = torch.logical_or(torch.logical_or(prune_mask, big_points_vs), big_points_ws)
self.prune_points(prune_mask)
tmp_radii = self.tmp_radii
self.tmp_radii = None
torch.cuda.empty_cache()
def add_densification_stats(self, viewspace_point_tensor, update_filter):
self.xyz_gradient_accum[update_filter] += torch.norm(viewspace_point_tensor.grad[update_filter,:2], dim=-1, keepdim=True)
self.denom[update_filter] += 1
================================================
FILE: train.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import os
import torch
from random import randint
from utils.loss_utils import l1_loss, ssim
from gaussian_renderer import render, network_gui
import sys
from scene import Scene, GaussianModel
from utils.general_utils import safe_state, get_expon_lr_func
import uuid
from tqdm import tqdm
from utils.image_utils import psnr
from argparse import ArgumentParser, Namespace
from arguments import ModelParams, PipelineParams, OptimizationParams
try:
from torch.utils.tensorboard import SummaryWriter
TENSORBOARD_FOUND = True
except ImportError:
TENSORBOARD_FOUND = False
try:
from fused_ssim import fused_ssim
FUSED_SSIM_AVAILABLE = True
except:
FUSED_SSIM_AVAILABLE = False
try:
from diff_gaussian_rasterization import SparseGaussianAdam
SPARSE_ADAM_AVAILABLE = True
except:
SPARSE_ADAM_AVAILABLE = False
def training(dataset, opt, pipe, testing_iterations, saving_iterations, checkpoint_iterations, checkpoint, debug_from):
if not SPARSE_ADAM_AVAILABLE and opt.optimizer_type == "sparse_adam":
sys.exit(f"Trying to use sparse adam but it is not installed, please install the correct rasterizer using pip install [3dgs_accel].")
first_iter = 0
tb_writer = prepare_output_and_logger(dataset)
gaussians = GaussianModel(dataset.sh_degree, opt.optimizer_type)
scene = Scene(dataset, gaussians)
gaussians.training_setup(opt)
if checkpoint:
(model_params, first_iter) = torch.load(checkpoint)
gaussians.restore(model_params, opt)
bg_color = [1, 1, 1] if dataset.white_background else [0, 0, 0]
background = torch.tensor(bg_color, dtype=torch.float32, device="cuda")
iter_start = torch.cuda.Event(enable_timing = True)
iter_end = torch.cuda.Event(enable_timing = True)
use_sparse_adam = opt.optimizer_type == "sparse_adam" and SPARSE_ADAM_AVAILABLE
depth_l1_weight = get_expon_lr_func(opt.depth_l1_weight_init, opt.depth_l1_weight_final, max_steps=opt.iterations)
viewpoint_stack = scene.getTrainCameras().copy()
viewpoint_indices = list(range(len(viewpoint_stack)))
ema_loss_for_log = 0.0
ema_Ll1depth_for_log = 0.0
progress_bar = tqdm(range(first_iter, opt.iterations), desc="Training progress")
first_iter += 1
for iteration in range(first_iter, opt.iterations + 1):
if network_gui.conn == None:
network_gui.try_connect()
while network_gui.conn != None:
try:
net_image_bytes = None
custom_cam, do_training, pipe.convert_SHs_python, pipe.compute_cov3D_python, keep_alive, scaling_modifer = network_gui.receive()
if custom_cam != None:
net_image = render(custom_cam, gaussians, pipe, background, scaling_modifier=scaling_modifer, use_trained_exp=dataset.train_test_exp, separate_sh=SPARSE_ADAM_AVAILABLE)["render"]
net_image_bytes = memoryview((torch.clamp(net_image, min=0, max=1.0) * 255).byte().permute(1, 2, 0).contiguous().cpu().numpy())
network_gui.send(net_image_bytes, dataset.source_path)
if do_training and ((iteration < int(opt.iterations)) or not keep_alive):
break
except Exception as e:
network_gui.conn = None
iter_start.record()
gaussians.update_learning_rate(iteration)
# Every 1000 its we increase the levels of SH up to a maximum degree
if iteration % 1000 == 0:
gaussians.oneupSHdegree()
# Pick a random Camera
if not viewpoint_stack:
viewpoint_stack = scene.getTrainCameras().copy()
viewpoint_indices = list(range(len(viewpoint_stack)))
rand_idx = randint(0, len(viewpoint_indices) - 1)
viewpoint_cam = viewpoint_stack.pop(rand_idx)
vind = viewpoint_indices.pop(rand_idx)
# Render
if (iteration - 1) == debug_from:
pipe.debug = True
bg = torch.rand((3), device="cuda") if opt.random_background else background
render_pkg = render(viewpoint_cam, gaussians, pipe, bg, use_trained_exp=dataset.train_test_exp, separate_sh=SPARSE_ADAM_AVAILABLE)
image, viewspace_point_tensor, visibility_filter, radii = render_pkg["render"], render_pkg["viewspace_points"], render_pkg["visibility_filter"], render_pkg["radii"]
if viewpoint_cam.alpha_mask is not None:
alpha_mask = viewpoint_cam.alpha_mask.cuda()
image *= alpha_mask
# Loss
gt_image = viewpoint_cam.original_image.cuda()
Ll1 = l1_loss(image, gt_image)
if FUSED_SSIM_AVAILABLE:
ssim_value = fused_ssim(image.unsqueeze(0), gt_image.unsqueeze(0))
else:
ssim_value = ssim(image, gt_image)
loss = (1.0 - opt.lambda_dssim) * Ll1 + opt.lambda_dssim * (1.0 - ssim_value)
# Depth regularization
Ll1depth_pure = 0.0
if depth_l1_weight(iteration) > 0 and viewpoint_cam.depth_reliable:
invDepth = render_pkg["depth"]
mono_invdepth = viewpoint_cam.invdepthmap.cuda()
depth_mask = viewpoint_cam.depth_mask.cuda()
Ll1depth_pure = torch.abs((invDepth - mono_invdepth) * depth_mask).mean()
Ll1depth = depth_l1_weight(iteration) * Ll1depth_pure
loss += Ll1depth
Ll1depth = Ll1depth.item()
else:
Ll1depth = 0
loss.backward()
iter_end.record()
with torch.no_grad():
# Progress bar
ema_loss_for_log = 0.4 * loss.item() + 0.6 * ema_loss_for_log
ema_Ll1depth_for_log = 0.4 * Ll1depth + 0.6 * ema_Ll1depth_for_log
if iteration % 10 == 0:
progress_bar.set_postfix({"Loss": f"{ema_loss_for_log:.{7}f}", "Depth Loss": f"{ema_Ll1depth_for_log:.{7}f}"})
progress_bar.update(10)
if iteration == opt.iterations:
progress_bar.close()
# Log and save
training_report(tb_writer, iteration, Ll1, loss, l1_loss, iter_start.elapsed_time(iter_end), testing_iterations, scene, render, (pipe, background, 1., SPARSE_ADAM_AVAILABLE, None, dataset.train_test_exp), dataset.train_test_exp)
if (iteration in saving_iterations):
print("\n[ITER {}] Saving Gaussians".format(iteration))
scene.save(iteration)
# Densification
if iteration < opt.densify_until_iter:
# Keep track of max radii in image-space for pruning
gaussians.max_radii2D[visibility_filter] = torch.max(gaussians.max_radii2D[visibility_filter], radii[visibility_filter])
gaussians.add_densification_stats(viewspace_point_tensor, visibility_filter)
if iteration > opt.densify_from_iter and iteration % opt.densification_interval == 0:
size_threshold = 20 if iteration > opt.opacity_reset_interval else None
gaussians.densify_and_prune(opt.densify_grad_threshold, 0.005, scene.cameras_extent, size_threshold, radii)
if iteration % opt.opacity_reset_interval == 0 or (dataset.white_background and iteration == opt.densify_from_iter):
gaussians.reset_opacity()
# Optimizer step
if iteration < opt.iterations:
gaussians.exposure_optimizer.step()
gaussians.exposure_optimizer.zero_grad(set_to_none = True)
if use_sparse_adam:
visible = radii > 0
gaussians.optimizer.step(visible, radii.shape[0])
gaussians.optimizer.zero_grad(set_to_none = True)
else:
gaussians.optimizer.step()
gaussians.optimizer.zero_grad(set_to_none = True)
if (iteration in checkpoint_iterations):
print("\n[ITER {}] Saving Checkpoint".format(iteration))
torch.save((gaussians.capture(), iteration), scene.model_path + "/chkpnt" + str(iteration) + ".pth")
def prepare_output_and_logger(args):
if not args.model_path:
if os.getenv('OAR_JOB_ID'):
unique_str=os.getenv('OAR_JOB_ID')
else:
unique_str = str(uuid.uuid4())
args.model_path = os.path.join("./output/", unique_str[0:10])
# Set up output folder
print("Output folder: {}".format(args.model_path))
os.makedirs(args.model_path, exist_ok = True)
with open(os.path.join(args.model_path, "cfg_args"), 'w') as cfg_log_f:
cfg_log_f.write(str(Namespace(**vars(args))))
# Create Tensorboard writer
tb_writer = None
if TENSORBOARD_FOUND:
tb_writer = SummaryWriter(args.model_path)
else:
print("Tensorboard not available: not logging progress")
return tb_writer
def training_report(tb_writer, iteration, Ll1, loss, l1_loss, elapsed, testing_iterations, scene : Scene, renderFunc, renderArgs, train_test_exp):
if tb_writer:
tb_writer.add_scalar('train_loss_patches/l1_loss', Ll1.item(), iteration)
tb_writer.add_scalar('train_loss_patches/total_loss', loss.item(), iteration)
tb_writer.add_scalar('iter_time', elapsed, iteration)
# Report test and samples of training set
if iteration in testing_iterations:
torch.cuda.empty_cache()
validation_configs = ({'name': 'test', 'cameras' : scene.getTestCameras()},
{'name': 'train', 'cameras' : [scene.getTrainCameras()[idx % len(scene.getTrainCameras())] for idx in range(5, 30, 5)]})
for config in validation_configs:
if config['cameras'] and len(config['cameras']) > 0:
l1_test = 0.0
psnr_test = 0.0
for idx, viewpoint in enumerate(config['cameras']):
image = torch.clamp(renderFunc(viewpoint, scene.gaussians, *renderArgs)["render"], 0.0, 1.0)
gt_image = torch.clamp(viewpoint.original_image.to("cuda"), 0.0, 1.0)
if train_test_exp:
image = image[..., image.shape[-1] // 2:]
gt_image = gt_image[..., gt_image.shape[-1] // 2:]
if tb_writer and (idx < 5):
tb_writer.add_images(config['name'] + "_view_{}/render".format(viewpoint.image_name), image[None], global_step=iteration)
if iteration == testing_iterations[0]:
tb_writer.add_images(config['name'] + "_view_{}/ground_truth".format(viewpoint.image_name), gt_image[None], global_step=iteration)
l1_test += l1_loss(image, gt_image).mean().double()
psnr_test += psnr(image, gt_image).mean().double()
psnr_test /= len(config['cameras'])
l1_test /= len(config['cameras'])
print("\n[ITER {}] Evaluating {}: L1 {} PSNR {}".format(iteration, config['name'], l1_test, psnr_test))
if tb_writer:
tb_writer.add_scalar(config['name'] + '/loss_viewpoint - l1_loss', l1_test, iteration)
tb_writer.add_scalar(config['name'] + '/loss_viewpoint - psnr', psnr_test, iteration)
if tb_writer:
tb_writer.add_histogram("scene/opacity_histogram", scene.gaussians.get_opacity, iteration)
tb_writer.add_scalar('total_points', scene.gaussians.get_xyz.shape[0], iteration)
torch.cuda.empty_cache()
if __name__ == "__main__":
# Set up command line argument parser
parser = ArgumentParser(description="Training script parameters")
lp = ModelParams(parser)
op = OptimizationParams(parser)
pp = PipelineParams(parser)
parser.add_argument('--ip', type=str, default="127.0.0.1")
parser.add_argument('--port', type=int, default=6009)
parser.add_argument('--debug_from', type=int, default=-1)
parser.add_argument('--detect_anomaly', action='store_true', default=False)
parser.add_argument("--test_iterations", nargs="+", type=int, default=[7_000, 30_000])
parser.add_argument("--save_iterations", nargs="+", type=int, default=[7_000, 30_000])
parser.add_argument("--quiet", action="store_true")
parser.add_argument('--disable_viewer', action='store_true', default=False)
parser.add_argument("--checkpoint_iterations", nargs="+", type=int, default=[])
parser.add_argument("--start_checkpoint", type=str, default = None)
args = parser.parse_args(sys.argv[1:])
args.save_iterations.append(args.iterations)
print("Optimizing " + args.model_path)
# Initialize system state (RNG)
safe_state(args.quiet)
# Start GUI server, configure and run training
if not args.disable_viewer:
network_gui.init(args.ip, args.port)
torch.autograd.set_detect_anomaly(args.detect_anomaly)
training(lp.extract(args), op.extract(args), pp.extract(args), args.test_iterations, args.save_iterations, args.checkpoint_iterations, args.start_checkpoint, args.debug_from)
# All done
print("\nTraining complete.")
================================================
FILE: utils/camera_utils.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
from scene.cameras import Camera
import numpy as np
from utils.graphics_utils import fov2focal
from PIL import Image
import cv2
WARNED = False
def loadCam(args, id, cam_info, resolution_scale, is_nerf_synthetic, is_test_dataset):
image = Image.open(cam_info.image_path)
if cam_info.depth_path != "":
try:
if is_nerf_synthetic:
invdepthmap = cv2.imread(cam_info.depth_path, -1).astype(np.float32) / 512
else:
invdepthmap = cv2.imread(cam_info.depth_path, -1).astype(np.float32) / float(2**16)
except FileNotFoundError:
print(f"Error: The depth file at path '{cam_info.depth_path}' was not found.")
raise
except IOError:
print(f"Error: Unable to open the image file '{cam_info.depth_path}'. It may be corrupted or an unsupported format.")
raise
except Exception as e:
print(f"An unexpected error occurred when trying to read depth at {cam_info.depth_path}: {e}")
raise
else:
invdepthmap = None
orig_w, orig_h = image.size
if args.resolution in [1, 2, 4, 8]:
resolution = round(orig_w/(resolution_scale * args.resolution)), round(orig_h/(resolution_scale * args.resolution))
else: # should be a type that converts to float
if args.resolution == -1:
if orig_w > 1600:
global WARNED
if not WARNED:
print("[ INFO ] Encountered quite large input images (>1.6K pixels width), rescaling to 1.6K.\n "
"If this is not desired, please explicitly specify '--resolution/-r' as 1")
WARNED = True
global_down = orig_w / 1600
else:
global_down = 1
else:
global_down = orig_w / args.resolution
scale = float(global_down) * float(resolution_scale)
resolution = (int(orig_w / scale), int(orig_h / scale))
return Camera(resolution, colmap_id=cam_info.uid, R=cam_info.R, T=cam_info.T,
FoVx=cam_info.FovX, FoVy=cam_info.FovY, depth_params=cam_info.depth_params,
image=image, invdepthmap=invdepthmap,
image_name=cam_info.image_name, uid=id, data_device=args.data_device,
train_test_exp=args.train_test_exp, is_test_dataset=is_test_dataset, is_test_view=cam_info.is_test)
def cameraList_from_camInfos(cam_infos, resolution_scale, args, is_nerf_synthetic, is_test_dataset):
camera_list = []
for id, c in enumerate(cam_infos):
camera_list.append(loadCam(args, id, c, resolution_scale, is_nerf_synthetic, is_test_dataset))
return camera_list
def camera_to_JSON(id, camera : Camera):
Rt = np.zeros((4, 4))
Rt[:3, :3] = camera.R.transpose()
Rt[:3, 3] = camera.T
Rt[3, 3] = 1.0
W2C = np.linalg.inv(Rt)
pos = W2C[:3, 3]
rot = W2C[:3, :3]
serializable_array_2d = [x.tolist() for x in rot]
camera_entry = {
'id' : id,
'img_name' : camera.image_name,
'width' : camera.width,
'height' : camera.height,
'position': pos.tolist(),
'rotation': serializable_array_2d,
'fy' : fov2focal(camera.FovY, camera.height),
'fx' : fov2focal(camera.FovX, camera.width)
}
return camera_entry
================================================
FILE: utils/general_utils.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import torch
import sys
from datetime import datetime
import numpy as np
import random
def inverse_sigmoid(x):
return torch.log(x/(1-x))
def PILtoTorch(pil_image, resolution):
resized_image_PIL = pil_image.resize(resolution)
resized_image = torch.from_numpy(np.array(resized_image_PIL)) / 255.0
if len(resized_image.shape) == 3:
return resized_image.permute(2, 0, 1)
else:
return resized_image.unsqueeze(dim=-1).permute(2, 0, 1)
def get_expon_lr_func(
lr_init, lr_final, lr_delay_steps=0, lr_delay_mult=1.0, max_steps=1000000
):
"""
Copied from Plenoxels
Continuous learning rate decay function. Adapted from JaxNeRF
The returned rate is lr_init when step=0 and lr_final when step=max_steps, and
is log-linearly interpolated elsewhere (equivalent to exponential decay).
If lr_delay_steps>0 then the learning rate will be scaled by some smooth
function of lr_delay_mult, such that the initial learning rate is
lr_init*lr_delay_mult at the beginning of optimization but will be eased back
to the normal learning rate when steps>lr_delay_steps.
:param conf: config subtree 'lr' or similar
:param max_steps: int, the number of steps during optimization.
:return HoF which takes step as input
"""
def helper(step):
if step < 0 or (lr_init == 0.0 and lr_final == 0.0):
# Disable this parameter
return 0.0
if lr_delay_steps > 0:
# A kind of reverse cosine decay.
delay_rate = lr_delay_mult + (1 - lr_delay_mult) * np.sin(
0.5 * np.pi * np.clip(step / lr_delay_steps, 0, 1)
)
else:
delay_rate = 1.0
t = np.clip(step / max_steps, 0, 1)
log_lerp = np.exp(np.log(lr_init) * (1 - t) + np.log(lr_final) * t)
return delay_rate * log_lerp
return helper
def strip_lowerdiag(L):
uncertainty = torch.zeros((L.shape[0], 6), dtype=torch.float, device="cuda")
uncertainty[:, 0] = L[:, 0, 0]
uncertainty[:, 1] = L[:, 0, 1]
uncertainty[:, 2] = L[:, 0, 2]
uncertainty[:, 3] = L[:, 1, 1]
uncertainty[:, 4] = L[:, 1, 2]
uncertainty[:, 5] = L[:, 2, 2]
return uncertainty
def strip_symmetric(sym):
return strip_lowerdiag(sym)
def build_rotation(r):
norm = torch.sqrt(r[:,0]*r[:,0] + r[:,1]*r[:,1] + r[:,2]*r[:,2] + r[:,3]*r[:,3])
q = r / norm[:, None]
R = torch.zeros((q.size(0), 3, 3), device='cuda')
r = q[:, 0]
x = q[:, 1]
y = q[:, 2]
z = q[:, 3]
R[:, 0, 0] = 1 - 2 * (y*y + z*z)
R[:, 0, 1] = 2 * (x*y - r*z)
R[:, 0, 2] = 2 * (x*z + r*y)
R[:, 1, 0] = 2 * (x*y + r*z)
R[:, 1, 1] = 1 - 2 * (x*x + z*z)
R[:, 1, 2] = 2 * (y*z - r*x)
R[:, 2, 0] = 2 * (x*z - r*y)
R[:, 2, 1] = 2 * (y*z + r*x)
R[:, 2, 2] = 1 - 2 * (x*x + y*y)
return R
def build_scaling_rotation(s, r):
L = torch.zeros((s.shape[0], 3, 3), dtype=torch.float, device="cuda")
R = build_rotation(r)
L[:,0,0] = s[:,0]
L[:,1,1] = s[:,1]
L[:,2,2] = s[:,2]
L = R @ L
return L
def safe_state(silent):
old_f = sys.stdout
class F:
def __init__(self, silent):
self.silent = silent
def write(self, x):
if not self.silent:
if x.endswith("\n"):
old_f.write(x.replace("\n", " [{}]\n".format(str(datetime.now().strftime("%d/%m %H:%M:%S")))))
else:
old_f.write(x)
def flush(self):
old_f.flush()
sys.stdout = F(silent)
random.seed(0)
np.random.seed(0)
torch.manual_seed(0)
torch.cuda.set_device(torch.device("cuda:0"))
================================================
FILE: utils/graphics_utils.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import torch
import math
import numpy as np
from typing import NamedTuple
class BasicPointCloud(NamedTuple):
points : np.array
colors : np.array
normals : np.array
def geom_transform_points(points, transf_matrix):
P, _ = points.shape
ones = torch.ones(P, 1, dtype=points.dtype, device=points.device)
points_hom = torch.cat([points, ones], dim=1)
points_out = torch.matmul(points_hom, transf_matrix.unsqueeze(0))
denom = points_out[..., 3:] + 0.0000001
return (points_out[..., :3] / denom).squeeze(dim=0)
def getWorld2View(R, t):
Rt = np.zeros((4, 4))
Rt[:3, :3] = R.transpose()
Rt[:3, 3] = t
Rt[3, 3] = 1.0
return np.float32(Rt)
def getWorld2View2(R, t, translate=np.array([.0, .0, .0]), scale=1.0):
Rt = np.zeros((4, 4))
Rt[:3, :3] = R.transpose()
Rt[:3, 3] = t
Rt[3, 3] = 1.0
C2W = np.linalg.inv(Rt)
cam_center = C2W[:3, 3]
cam_center = (cam_center + translate) * scale
C2W[:3, 3] = cam_center
Rt = np.linalg.inv(C2W)
return np.float32(Rt)
def getProjectionMatrix(znear, zfar, fovX, fovY):
tanHalfFovY = math.tan((fovY / 2))
tanHalfFovX = math.tan((fovX / 2))
top = tanHalfFovY * znear
bottom = -top
right = tanHalfFovX * znear
left = -right
P = torch.zeros(4, 4)
z_sign = 1.0
P[0, 0] = 2.0 * znear / (right - left)
P[1, 1] = 2.0 * znear / (top - bottom)
P[0, 2] = (right + left) / (right - left)
P[1, 2] = (top + bottom) / (top - bottom)
P[3, 2] = z_sign
P[2, 2] = z_sign * zfar / (zfar - znear)
P[2, 3] = -(zfar * znear) / (zfar - znear)
return P
def fov2focal(fov, pixels):
return pixels / (2 * math.tan(fov / 2))
def focal2fov(focal, pixels):
return 2*math.atan(pixels/(2*focal))
================================================
FILE: utils/image_utils.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import torch
def mse(img1, img2):
return (((img1 - img2)) ** 2).view(img1.shape[0], -1).mean(1, keepdim=True)
def psnr(img1, img2):
mse = (((img1 - img2)) ** 2).view(img1.shape[0], -1).mean(1, keepdim=True)
return 20 * torch.log10(1.0 / torch.sqrt(mse))
================================================
FILE: utils/loss_utils.py
================================================
#
# Copyright (C) 2023, Inria
# GRAPHDECO research group, https://team.inria.fr/graphdeco
# All rights reserved.
#
# This software is free for non-commercial, research and evaluation use
# under the terms of the LICENSE.md file.
#
# For inquiries contact george.drettakis@inria.fr
#
import torch
import torch.nn.functional as F
from torch.autograd import Variable
from math import exp
try:
from diff_gaussian_rasterization._C import fusedssim, fusedssim_backward
except:
pass
C1 = 0.01 ** 2
C2 = 0.03 ** 2
class FusedSSIMMap(torch.autograd.Function):
@staticmethod
def forward(ctx, C1, C2, img1, img2):
ssim_map = fusedssim(C1, C2, img1, img2)
ctx.save_for_backward(img1.detach(), img2)
ctx.C1 = C1
ctx.C2 = C2
return ssim_map
@staticmethod
def backward(ctx, opt_grad):
img1, img2 = ctx.saved_tensors
C1, C2 = ctx.C1, ctx.C2
grad = fusedssim_backward(C1, C2, img1, img2, opt_grad)
return None, None, grad, None
def l1_loss(network_output, gt):
return torch.abs((network_output - gt)).mean()
def l2_loss(network_output, gt):
return ((network_output - gt) ** 2).mean()
def gaussian(window_size, sigma):
gauss = torch.Tensor([exp(-(x - window_size // 2) ** 2 / float(2 * sigma ** 2)) for x in range(window_size)])
return gauss / gauss.sum()
def create_window(window_size, channel):
_1D_window = gaussian(window_size, 1.5).unsqueeze(1)
_2D_window = _1D_window.mm(_1D_window.t()).float().unsqueeze(0).unsqueeze(0)
window = Variable(_2D_window.expand(channel, 1, window_size, window_size).contiguous())
return window
def ssim(img1, img2, window_size=11, size_average=True):
channel = img1.size(-3)
window = create_window(window_size, channel)
if img1.is_cuda:
window = window.cuda(img1.get_device())
window = window.type_as(img1)
return _ssim(img1, img2, window, window_size, channel, size_average)
def _ssim(img1, img2, window, window_size, channel, size_average=True):
mu1 = F.conv2d(img1, window, padding=window_size // 2, groups=channel)
mu2 = F.conv2d(img2, window, padding=window_size // 2, groups=channel)
mu1_sq = mu1.pow(2)
mu2_sq = mu2.pow(2)
mu1_mu2 = mu1 * mu2
sigma1_sq = F.conv2d(img1 * img1, window, padding=window_size // 2, groups=channel) - mu1_sq
sigma2_sq = F.conv2d(img2 * img2, window, padding=window_size // 2, groups=channel) - mu2_sq
sigma12 = F.conv2d(img1 * img2, window, padding=window_size // 2, groups=channel) - mu1_mu2
C1 = 0.01 ** 2
C2 = 0.03 ** 2
ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2))
if size_average:
return ssim_map.mean()
else:
return ssim_map.mean(1).mean(1).mean(1)
def fast_ssim(img1, img2):
ssim_map = FusedSSIMMap.apply(C1, C2, img1, img2)
return ssim_map.mean()
================================================
FILE: utils/make_depth_scale.py
================================================
import numpy as np
import argparse
import cv2
from joblib import delayed, Parallel
import json
from read_write_model import *
def get_scales(key, cameras, images, points3d_ordered, args):
image_meta = images[key]
cam_intrinsic = cameras[image_meta.camera_id]
pts_idx = images_metas[key].point3D_ids
mask = pts_idx >= 0
mask *= pts_idx < len(points3d_ordered)
pts_idx = pts_idx[mask]
valid_xys = image_meta.xys[mask]
if len(pts_idx) > 0:
pts = points3d_ordered[pts_idx]
else:
pts = np.array([0, 0, 0])
R = qvec2rotmat(image_meta.qvec)
pts = np.dot(pts, R.T) + image_meta.tvec
invcolmapdepth = 1. / pts[..., 2]
n_remove = len(image_meta.name.split('.')[-1]) + 1
invmonodepthmap = cv2.imread(f"{args.depths_dir}/{image_meta.name[:-n_remove]}.png", cv2.IMREAD_UNCHANGED)
if invmonodepthmap is None:
return None
if invmonodepthmap.ndim != 2:
invmonodepthmap = invmonodepthmap[..., 0]
invmonodepthmap = invmonodepthmap.astype(np.float32) / (2**16)
s = invmonodepthmap.shape[0] / cam_intrinsic.height
maps = (valid_xys * s).astype(np.float32)
valid = (
(maps[..., 0] >= 0) *
(maps[..., 1] >= 0) *
(maps[..., 0] < cam_intrinsic.width * s) *
(maps[..., 1] < cam_intrinsic.height * s) * (invcolmapdepth > 0))
if valid.sum() > 10 and (invcolmapdepth.max() - invcolmapdepth.min()) > 1e-3:
maps = maps[valid, :]
invcolmapdepth = invcolmapdepth[valid]
invmonodepth = cv2.remap(invmonodepthmap, maps[..., 0], maps[..., 1], interpolation=cv2.INTER_LINEAR, borderMode=cv2.BORDER_REPLICATE)[..., 0]
## Median / dev
t_colmap = np.median(invcolmapdepth)
s_colmap = np.mean(np.abs(invcolmapdepth - t_colmap))
t_mono = np.median(invmonodepth)
s_mono = np.mean(np.abs(invmonodepth - t_mono))
scale = s_colmap / s_mono
offset = t_colmap - t_mono * scale
else:
scale = 0
offset = 0
return {"image_name": image_meta.name[:-n_remove], "scale": scale, "offset": offset}
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--base_dir', default="../data/big_gaussians/standalone_chunks/campus")
parser.add_argument('--depths_dir', default="../data/big_gaussians/standalone_chunks/campus/depths_any")
parser.add_argument('--model_type', default="bin")
args = parser.parse_args()
cam_intrinsics, images_metas, points3d = read_model(os.path.join(args.base_dir, "sparse", "0"), ext=f".{args.model_type}")
pts_indices = np.array([points3d[key].id for key in points3d])
pts_xyzs = np.array([points3d[key].xyz for key in points3d])
points3d_ordered = np.zeros([pts_indices.max()+1, 3])
points3d_ordered[pts_indices] = pts_xyzs
# depth_param_list = [get_scales(key, cam_intrinsics, images_metas, points3d_ordered, args) for key in images_metas]
depth_param_list = Parallel(n_jobs=-1, backend="threading")(
delayed(get_scales)(key, cam_intrinsics, images_metas, points3d_ordered, args) for key in images_metas
)
depth_params = {
depth_param["image_name"]: {"scale": depth_param["scale"], "offset": depth_param["offset"]}
for depth_param in depth_param_list if depth_param != None
}
with open(f"{args.base_dir}/sparse/0/depth_params.json", "w") as f:
json.dump(depth_params, f, indent=2)
print(0)
================================================
FILE: utils/read_write_model.py
================================================
# Copyright (c) 2023, ETH Zurich and UNC Chapel Hill.
# All rights reserved.
#
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# * Redistributions of source code must retain the above copyright
# notice, this list of conditions and the following disclaimer.
#
# * Redistributions in binary form must reproduce the above copyright
# notice, this list of conditions and the following disclaimer in the
# documentation and/or other materials provided with the distribution.
#
# * Neither the name of ETH Zurich and UNC Chapel Hill nor the names of
# its contributors may be used to endorse or promote products derived
# from this software without specific prior written permission.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDERS OR CONTRIBUTORS BE
# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
# POSSIBILITY OF SUCH DAMAGE.
import os
import collections
import numpy as np
import struct
import argparse
CameraModel = collections.namedtuple(
"CameraModel", ["model_id", "model_name", "num_params"]
)
Camera = collections.namedtuple(
"Camera", ["id", "model", "width", "height", "params"]
)
BaseImage = collections.namedtuple(
"Image", ["id", "qvec", "tvec", "camera_id", "name", "xys", "point3D_ids"]
)
Point3D = collections.namedtuple(
"Point3D", ["id", "xyz", "rgb", "error", "image_ids", "point2D_idxs"]
)
class Image(BaseImage):
def qvec2rotmat(self):
return qvec2rotmat(self.qvec)
CAMERA_MODELS = {
CameraModel(model_id=0, model_name="SIMPLE_PINHOLE", num_params=3),
CameraModel(model_id=1, model_name="PINHOLE", num_params=4),
CameraModel(model_id=2, model_name="SIMPLE_RADIAL", num_params=4),
CameraModel(model_id=3, model_name="RADIAL", num_params=5),
CameraModel(model_id=4, model_name="OPENCV", num_params=8),
CameraModel(model_id=5, model_name="OPENCV_FISHEYE", num_params=8),
CameraModel(model_id=6, model_name="FULL_OPENCV", num_params=12),
CameraModel(model_id=7, model_name="FOV", num_params=5),
CameraModel(model_id=8, model_name="SIMPLE_RADIAL_FISHEYE", num_params=4),
CameraModel(model_id=9, model_name="RADIAL_FISHEYE", num_params=5),
CameraModel(model_id=10, model_name="THIN_PRISM_FISHEYE", num_params=12),
}
CAMERA_MODEL_IDS = dict(
[(camera_model.model_id, camera_model) for camera_model in CAMERA_MODELS]
)
CAMERA_MODEL_NAMES = dict(
[(camera_model.model_name, camera_model) for camera_model in CAMERA_MODELS]
)
def read_next_bytes(fid, num_bytes, format_char_sequence, endian_character="<"):
"""Read and unpack the next bytes from a binary file.
:param fid:
:param num_bytes: Sum of combination of {2, 4, 8}, e.g. 2, 6, 16, 30, etc.
:param format_char_sequence: List of {c, e, f, d, h, H, i, I, l, L, q, Q}.
:param endian_character: Any of {@, =, <, >, !}
:return: Tuple of read and unpacked values.
"""
data = fid.read(num_bytes)
return struct.unpack(endian_character + format_char_sequence, data)
def write_next_bytes(fid, data, format_char_sequence, endian_character="<"):
"""pack and write to a binary file.
:param fid:
:param data: data to send, if multiple elements are sent at the same time,
they should be encapsuled either in a list or a tuple
:param format_char_sequence: List of {c, e, f, d, h, H, i, I, l, L, q, Q}.
should be the same length as the data list or tuple
:param endian_character: Any of {@, =, <, >, !}
"""
if isinstance(data, (list, tuple)):
bytes = struct.pack(endian_character + format_char_sequence, *data)
else:
bytes = struct.pack(endian_character + format_char_sequence, data)
fid.write(bytes)
def read_cameras_text(path):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::WriteCamerasText(const std::string& path)
void Reconstruction::ReadCamerasText(const std::string& path)
"""
cameras = {}
with open(path, "r") as fid:
while True:
line = fid.readline()
if not line:
break
line = line.strip()
if len(line) > 0 and line[0] != "#":
elems = line.split()
camera_id = int(elems[0])
model = elems[1]
width = int(elems[2])
height = int(elems[3])
params = np.array(tuple(map(float, elems[4:])))
cameras[camera_id] = Camera(
id=camera_id,
model=model,
width=width,
height=height,
params=params,
)
return cameras
def read_cameras_binary(path_to_model_file):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::WriteCamerasBinary(const std::string& path)
void Reconstruction::ReadCamerasBinary(const std::string& path)
"""
cameras = {}
with open(path_to_model_file, "rb") as fid:
num_cameras = read_next_bytes(fid, 8, "Q")[0]
for _ in range(num_cameras):
camera_properties = read_next_bytes(
fid, num_bytes=24, format_char_sequence="iiQQ"
)
camera_id = camera_properties[0]
model_id = camera_properties[1]
model_name = CAMERA_MODEL_IDS[camera_properties[1]].model_name
width = camera_properties[2]
height = camera_properties[3]
num_params = CAMERA_MODEL_IDS[model_id].num_params
params = read_next_bytes(
fid,
num_bytes=8 * num_params,
format_char_sequence="d" * num_params,
)
cameras[camera_id] = Camera(
id=camera_id,
model=model_name,
width=width,
height=height,
params=np.array(params),
)
assert len(cameras) == num_cameras
return cameras
def write_cameras_text(cameras, path):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::WriteCamerasText(const std::string& path)
void Reconstruction::ReadCamerasText(const std::string& path)
"""
HEADER = (
"# Camera list with one line of data per camera:\n"
+ "# CAMERA_ID, MODEL, WIDTH, HEIGHT, PARAMS[]\n"
+ "# Number of cameras: {}\n".format(len(cameras))
)
with open(path, "w") as fid:
fid.write(HEADER)
for _, cam in cameras.items():
to_write = [cam.id, cam.model, cam.width, cam.height, *cam.params]
line = " ".join([str(elem) for elem in to_write])
fid.write(line + "\n")
def write_cameras_binary(cameras, path_to_model_file):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::WriteCamerasBinary(const std::string& path)
void Reconstruction::ReadCamerasBinary(const std::string& path)
"""
with open(path_to_model_file, "wb") as fid:
write_next_bytes(fid, len(cameras), "Q")
for _, cam in cameras.items():
model_id = CAMERA_MODEL_NAMES[cam.model].model_id
camera_properties = [cam.id, model_id, cam.width, cam.height]
write_next_bytes(fid, camera_properties, "iiQQ")
for p in cam.params:
write_next_bytes(fid, float(p), "d")
return cameras
def read_images_text(path):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::ReadImagesText(const std::string& path)
void Reconstruction::WriteImagesText(const std::string& path)
"""
images = {}
with open(path, "r") as fid:
while True:
line = fid.readline()
if not line:
break
line = line.strip()
if len(line) > 0 and line[0] != "#":
elems = line.split()
image_id = int(elems[0])
qvec = np.array(tuple(map(float, elems[1:5])))
tvec = np.array(tuple(map(float, elems[5:8])))
camera_id = int(elems[8])
image_name = elems[9]
elems = fid.readline().split()
xys = np.column_stack(
[
tuple(map(float, elems[0::3])),
tuple(map(float, elems[1::3])),
]
)
point3D_ids = np.array(tuple(map(int, elems[2::3])))
images[image_id] = Image(
id=image_id,
qvec=qvec,
tvec=tvec,
camera_id=camera_id,
name=image_name,
xys=xys,
point3D_ids=point3D_ids,
)
return images
def read_images_binary(path_to_model_file):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::ReadImagesBinary(const std::string& path)
void Reconstruction::WriteImagesBinary(const std::string& path)
"""
images = {}
with open(path_to_model_file, "rb") as fid:
num_reg_images = read_next_bytes(fid, 8, "Q")[0]
for _ in range(num_reg_images):
binary_image_properties = read_next_bytes(
fid, num_bytes=64, format_char_sequence="idddddddi"
)
image_id = binary_image_properties[0]
qvec = np.array(binary_image_properties[1:5])
tvec = np.array(binary_image_properties[5:8])
camera_id = binary_image_properties[8]
image_name = ""
current_char = read_next_bytes(fid, 1, "c")[0]
while current_char != b"\x00": # look for the ASCII 0 entry
image_name += current_char.decode("utf-8")
current_char = read_next_bytes(fid, 1, "c")[0]
num_points2D = read_next_bytes(
fid, num_bytes=8, format_char_sequence="Q"
)[0]
x_y_id_s = read_next_bytes(
fid,
num_bytes=24 * num_points2D,
format_char_sequence="ddq" * num_points2D,
)
xys = np.column_stack(
[
tuple(map(float, x_y_id_s[0::3])),
tuple(map(float, x_y_id_s[1::3])),
]
)
point3D_ids = np.array(tuple(map(int, x_y_id_s[2::3])))
images[image_id] = Image(
id=image_id,
qvec=qvec,
tvec=tvec,
camera_id=camera_id,
name=image_name,
xys=xys,
point3D_ids=point3D_ids,
)
return images
def write_images_text(images, path):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::ReadImagesText(const std::string& path)
void Reconstruction::WriteImagesText(const std::string& path)
"""
if len(images) == 0:
mean_observations = 0
else:
mean_observations = sum(
(len(img.point3D_ids) for _, img in images.items())
) / len(images)
HEADER = (
"# Image list with two lines of data per image:\n"
+ "# IMAGE_ID, QW, QX, QY, QZ, TX, TY, TZ, CAMERA_ID, NAME\n"
+ "# POINTS2D[] as (X, Y, POINT3D_ID)\n"
+ "# Number of images: {}, mean observations per image: {}\n".format(
len(images), mean_observations
)
)
with open(path, "w") as fid:
fid.write(HEADER)
for _, img in images.items():
image_header = [
img.id,
*img.qvec,
*img.tvec,
img.camera_id,
img.name,
]
first_line = " ".join(map(str, image_header))
fid.write(first_line + "\n")
points_strings = []
for xy, point3D_id in zip(img.xys, img.point3D_ids):
points_strings.append(" ".join(map(str, [*xy, point3D_id])))
fid.write(" ".join(points_strings) + "\n")
def write_images_binary(images, path_to_model_file):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::ReadImagesBinary(const std::string& path)
void Reconstruction::WriteImagesBinary(const std::string& path)
"""
with open(path_to_model_file, "wb") as fid:
write_next_bytes(fid, len(images), "Q")
for _, img in images.items():
write_next_bytes(fid, img.id, "i")
write_next_bytes(fid, img.qvec.tolist(), "dddd")
write_next_bytes(fid, img.tvec.tolist(), "ddd")
write_next_bytes(fid, img.camera_id, "i")
for char in img.name:
write_next_bytes(fid, char.encode("utf-8"), "c")
write_next_bytes(fid, b"\x00", "c")
write_next_bytes(fid, len(img.point3D_ids), "Q")
for xy, p3d_id in zip(img.xys, img.point3D_ids):
write_next_bytes(fid, [*xy, p3d_id], "ddq")
def read_points3D_text(path):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::ReadPoints3DText(const std::string& path)
void Reconstruction::WritePoints3DText(const std::string& path)
"""
points3D = {}
with open(path, "r") as fid:
while True:
line = fid.readline()
if not line:
break
line = line.strip()
if len(line) > 0 and line[0] != "#":
elems = line.split()
point3D_id = int(elems[0])
xyz = np.array(tuple(map(float, elems[1:4])))
rgb = np.array(tuple(map(int, elems[4:7])))
error = float(elems[7])
image_ids = np.array(tuple(map(int, elems[8::2])))
point2D_idxs = np.array(tuple(map(int, elems[9::2])))
points3D[point3D_id] = Point3D(
id=point3D_id,
xyz=xyz,
rgb=rgb,
error=error,
image_ids=image_ids,
point2D_idxs=point2D_idxs,
)
return points3D
def read_points3D_binary(path_to_model_file):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::ReadPoints3DBinary(const std::string& path)
void Reconstruction::WritePoints3DBinary(const std::string& path)
"""
points3D = {}
with open(path_to_model_file, "rb") as fid:
num_points = read_next_bytes(fid, 8, "Q")[0]
for _ in range(num_points):
binary_point_line_properties = read_next_bytes(
fid, num_bytes=43, format_char_sequence="QdddBBBd"
)
point3D_id = binary_point_line_properties[0]
xyz = np.array(binary_point_line_properties[1:4])
rgb = np.array(binary_point_line_properties[4:7])
error = np.array(binary_point_line_properties[7])
track_length = read_next_bytes(
fid, num_bytes=8, format_char_sequence="Q"
)[0]
track_elems = read_next_bytes(
fid,
num_bytes=8 * track_length,
format_char_sequence="ii" * track_length,
)
image_ids = np.array(tuple(map(int, track_elems[0::2])))
point2D_idxs = np.array(tuple(map(int, track_elems[1::2])))
points3D[point3D_id] = Point3D(
id=point3D_id,
xyz=xyz,
rgb=rgb,
error=error,
image_ids=image_ids,
point2D_idxs=point2D_idxs,
)
return points3D
def write_points3D_text(points3D, path):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::ReadPoints3DText(const std::string& path)
void Reconstruction::WritePoints3DText(const std::string& path)
"""
if len(points3D) == 0:
mean_track_length = 0
else:
mean_track_length = sum(
(len(pt.image_ids) for _, pt in points3D.items())
) / len(points3D)
HEADER = (
"# 3D point list with one line of data per point:\n"
+ "# POINT3D_ID, X, Y, Z, R, G, B, ERROR, TRACK[] as (IMAGE_ID, POINT2D_IDX)\n"
+ "# Number of points: {}, mean track length: {}\n".format(
len(points3D), mean_track_length
)
)
with open(path, "w") as fid:
fid.write(HEADER)
for _, pt in points3D.items():
point_header = [pt.id, *pt.xyz, *pt.rgb, pt.error]
fid.write(" ".join(map(str, point_header)) + " ")
track_strings = []
for image_id, point2D in zip(pt.image_ids, pt.point2D_idxs):
track_strings.append(" ".join(map(str, [image_id, point2D])))
fid.write(" ".join(track_strings) + "\n")
def write_points3D_binary(points3D, path_to_model_file):
"""
see: src/colmap/scene/reconstruction.cc
void Reconstruction::ReadPoints3DBinary(const std::string& path)
void Reconstruction::WritePoints3DBinary(const std::string& path)
"""
with open(path_to_model_file, "wb") as fid:
write_next_bytes(fid, len(points3D), "Q")
for _, pt in points3D.items():
write_next_bytes(fid, pt.id, "Q")
write_next_bytes(fid, pt.xyz.tolist(), "ddd")
write_next_bytes(fid, pt.rgb.tolist(), "BBB")
write_next_bytes(fid, pt.error, "d")
track_length = pt.image_ids.shape[0]
write_next_bytes(fid, track_length, "Q")
for image_id, point2D_id in zip(pt.image_ids, pt.point2D_idxs):
write_next_bytes(fid, [image_id, point2D_id], "ii")
def detect_model_format(path, ext):
if (
os.path.isfile(os.path.join(path, "cameras" + ext))
and os.path.isfile(os.path.join(path, "images" + ext))
and os.path.isfile(os.path.join(path, "points3D" + ext))
):
print("Detected model format: '" + ext + "'")
return True
return False
def read_model(path, ext=""):
# try to detect the extension automatically
if ext == "":
if detect_model_format(path, ".bin"):
ext = ".bin"
elif detect_model_format(path, ".txt"):
ext = ".txt"
else:
print("Provide model format: '.bin' or '.txt'")
return
if ext == ".txt":
cameras = read_cameras_text(os.path.join(path, "cameras" + ext))
images = read_images_text(os.path.join(path, "images" + ext))
points3D = read_points3D_text(os.path.join(path, "points3D") + ext)
else:
cameras = read_cameras_binary(os.path.join(path, "cameras" + ext))
images = read_images_binary(os.path.join(path, "images" + ext))
points3D = read_points3D_binary(os.path.join(path, "points3D") + ext)
return cameras, images, points3D
def write_model(cameras, images, points3D, path, ext=".bin"):
if ext == ".txt":
write_cameras_text(cameras, os.path.join(path, "cameras" + ext))
write_images_text(images, os.path.join(path, "images" + ext))
write_points3D_text(points3D, os.path.join(path, "points3D") + ext)
else:
write_cameras_binary(cameras, os.path.join(path, "cameras" + ext))
write_images_binary(images, os.path.join(path, "images" + ext))
write_points3D_binary(points3D, os.path.join(path, "points3D") + ext)
return cameras, images, points3D
def qvec2rotmat(qvec):
return np.array(
[
[
1 - 2 * qvec[2] ** 2 - 2 * qvec[3] ** 2,
2 * qvec[1] * qvec[2] - 2 * qvec[0] * qvec[3],
2 * qvec[3] * qvec[1] + 2 * qvec[0] * qvec[2],
],
[
2 * qvec[1] * qvec[2] + 2 * qvec[0] * qvec[3],
1 - 2 * qvec[1] ** 2 - 2 * qvec[3] ** 2,
2 * qvec[2] * qvec[3] - 2 * qvec[0] * qvec[1],
],
[
2 * qvec[3] * qvec[1] - 2 * qvec[0] * qvec[2],
2 * qvec[2] * qvec[3] + 2 * qvec[0] * qvec[1],
1 - 2 * qvec[1] ** 2 - 2 * qvec[2] ** 2,
],
]
)
def rotmat2qvec(R):
Rxx, Ryx, Rzx, Rxy, Ryy, Rzy, Rxz, Ryz, Rzz = R.flat
K = (
np.array(
[
[Rxx - Ryy - Rzz, 0, 0, 0],
[Ryx + Rxy, Ryy - Rxx - Rzz, 0, 0],
[Rzx + Rxz, Rzy + Ryz, Rzz - Rxx - Ryy, 0],
[Ryz - Rzy, Rzx - Rxz, Rxy - Ryx, Rxx + Ryy + Rzz],
]
)
/ 3.0
)
eigvals, eigvecs = np.linalg.eigh(K)
qvec = eigvecs[[3, 0, 1, 2], np.argmax(eigvals)]
if qvec[0] < 0:
qvec *= -1
return qvec
# def main():
# parser = argparse.ArgumentParser(
# description="Read and write COLMAP binary and text models"
# )
# parser.add_argument("--input_model", help="path to input model folder")
# parser.add_argument(
# "--input_format",
# choices=[".bin", ".txt"],
# help="input model format",
# default="",
# )
# parser.add_argument("--output_model", help="path to output model folder")
# parser.add_argument(
# "--output_format",
# choices=[".bin", ".txt"],
# help="outut model format",
# default=".txt",
# )
# args = parser.parse_args()
# cameras, images, points3D = read_model(
# path=args.input_model, ext=args.input_format
# )
# print("num_cameras:", len(cameras))
# print("num_images:", len(images))
# print("num_points3D:", len(points3D))
# if args.output_model is not None:
# write_model(
# cameras,
# images,
# points3D,
# path=args.output_model,
# ext=args.output_format,
# )
# if __name__ == "__main__":
# main()
================================================
FILE: utils/sh_utils.py
================================================
# Copyright 2021 The PlenOctree Authors.
# Redistribution and use in source and binary forms, with or without
# modification, are permitted provided that the following conditions are met:
#
# 1. Redistributions of source code must retain the above copyright notice,
# this list of conditions and the following disclaimer.
#
# 2. Redistributions in binary form must reproduce the above copyright notice,
# this list of conditions and the following disclaimer in the documentation
# and/or other materials provided with the distribution.
#
# THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
# AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
# IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
# ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE
# LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
# CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
# SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
# INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
# CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
# ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
# POSSIBILITY OF SUCH DAMAGE.
import torch
C0 = 0.28209479177387814
C1 = 0.4886025119029199
C2 = [
1.0925484305920792,
-1.0925484305920792,
0.31539156525252005,
-1.0925484305920792,
0.5462742152960396
]
C3 = [
-0.5900435899266435,
2.890611442640554,
-0.4570457994644658,
0.3731763325901154,
-0.4570457994644658,
1.445305721320277,
-0.5900435899266435
]
C4 = [
2.5033429417967046,
-1.7701307697799304,
0.9461746957575601,
-0.6690465435572892,
0.10578554691520431,
-0.6690465435572892,
0.47308734787878004,
-1.7701307697799304,
0.6258357354491761,
]
def eval_sh(deg, sh, dirs):
"""
Evaluate spherical harmonics at unit directions
using hardcoded SH polynomials.
Works with torch/np/jnp.
... Can be 0 or more batch dimensions.
Args:
deg: int SH deg. Currently, 0-3 supported
sh: jnp.ndarray SH coeffs [..., C, (deg + 1) ** 2]
dirs: jnp.ndarray unit directions [..., 3]
Returns:
[..., C]
"""
assert deg <= 4 and deg >= 0
coeff = (deg + 1) ** 2
assert sh.shape[-1] >= coeff
result = C0 * sh[..., 0]
if deg > 0:
x, y, z = dirs[..., 0:1], dirs[..., 1:2], dirs[..., 2:3]
result = (result -
C1 * y * sh[..., 1] +
C1 * z * sh[..., 2] -
C1 * x * sh[..., 3])
if deg > 1:
xx, yy, zz = x * x, y * y, z * z
xy, yz, xz = x * y, y * z, x * z
result = (result +
C2[0] * xy * sh[..., 4] +
C2[1] * yz * sh[..., 5] +
C2[2] * (2.0 * zz - xx - yy) * sh[..., 6] +
C2[3] * xz * sh[..., 7] +
C2[4] * (xx - yy) * sh[..., 8])
if deg > 2:
result = (result +
C3[0] * y * (3 * xx - yy) * sh[..., 9] +
C3[1] * xy * z * sh[..., 10] +
C3[2] * y * (4 * zz - xx - yy)* sh[..., 11] +
C3[3] * z * (2 * zz - 3 * xx - 3 * yy) * sh[..., 12] +
C3[4] * x * (4 * zz - xx
gitextract_uhkue7yy/
├── .gitignore
├── .gitmodules
├── LICENSE.md
├── README.md
├── arguments/
│ └── __init__.py
├── convert.py
├── environment.yml
├── full_eval.py
├── gaussian_renderer/
│ ├── __init__.py
│ └── network_gui.py
├── lpipsPyTorch/
│ ├── __init__.py
│ └── modules/
│ ├── lpips.py
│ ├── networks.py
│ └── utils.py
├── metrics.py
├── render.py
├── results.md
├── scene/
│ ├── __init__.py
│ ├── cameras.py
│ ├── colmap_loader.py
│ ├── dataset_readers.py
│ └── gaussian_model.py
├── train.py
└── utils/
├── camera_utils.py
├── general_utils.py
├── graphics_utils.py
├── image_utils.py
├── loss_utils.py
├── make_depth_scale.py
├── read_write_model.py
├── sh_utils.py
└── system_utils.py
SYMBOL INDEX (164 symbols across 24 files)
FILE: arguments/__init__.py
class GroupParams (line 16) | class GroupParams:
class ParamGroup (line 19) | class ParamGroup:
method __init__ (line 20) | def __init__(self, parser: ArgumentParser, name : str, fill_none = Fal...
method extract (line 40) | def extract(self, args):
class ModelParams (line 47) | class ModelParams(ParamGroup):
method __init__ (line 48) | def __init__(self, parser, sentinel=False):
method extract (line 61) | def extract(self, args):
class PipelineParams (line 66) | class PipelineParams(ParamGroup):
method __init__ (line 67) | def __init__(self, parser):
class OptimizationParams (line 74) | class OptimizationParams(ParamGroup):
method __init__ (line 75) | def __init__(self, parser):
function get_combined_args (line 102) | def get_combined_args(parser : ArgumentParser):
FILE: gaussian_renderer/__init__.py
function render (line 18) | def render(viewpoint_camera, pc : GaussianModel, pipe, bg_color : torch....
FILE: gaussian_renderer/network_gui.py
function init (line 26) | def init(wish_host, wish_port):
function try_connect (line 34) | def try_connect():
function read (line 43) | def read():
function send (line 50) | def send(message_bytes, verify):
function receive (line 57) | def receive():
FILE: lpipsPyTorch/__init__.py
function lpips (line 6) | def lpips(x: torch.Tensor,
FILE: lpipsPyTorch/modules/lpips.py
class LPIPS (line 8) | class LPIPS(nn.Module):
method __init__ (line 17) | def __init__(self, net_type: str = 'alex', version: str = '0.1'):
method forward (line 30) | def forward(self, x: torch.Tensor, y: torch.Tensor):
FILE: lpipsPyTorch/modules/networks.py
function get_network (line 12) | def get_network(net_type: str):
class LinLayers (line 23) | class LinLayers(nn.ModuleList):
method __init__ (line 24) | def __init__(self, n_channels_list: Sequence[int]):
class BaseNet (line 36) | class BaseNet(nn.Module):
method __init__ (line 37) | def __init__(self):
method set_requires_grad (line 46) | def set_requires_grad(self, state: bool):
method z_score (line 50) | def z_score(self, x: torch.Tensor):
method forward (line 53) | def forward(self, x: torch.Tensor):
class SqueezeNet (line 66) | class SqueezeNet(BaseNet):
method __init__ (line 67) | def __init__(self):
class AlexNet (line 77) | class AlexNet(BaseNet):
method __init__ (line 78) | def __init__(self):
class VGG16 (line 88) | class VGG16(BaseNet):
method __init__ (line 89) | def __init__(self):
FILE: lpipsPyTorch/modules/utils.py
function normalize_activation (line 6) | def normalize_activation(x, eps=1e-10):
function get_state_dict (line 11) | def get_state_dict(net_type: str = 'alex', version: str = '0.1'):
FILE: metrics.py
function readImages (line 24) | def readImages(renders_dir, gt_dir):
function evaluate (line 36) | def evaluate(model_paths):
FILE: render.py
function render_set (line 30) | def render_set(model_path, name, iteration, views, gaussians, pipeline, ...
function render_sets (line 48) | def render_sets(dataset : ModelParams, iteration : int, pipeline : Pipel...
FILE: scene/__init__.py
class Scene (line 21) | class Scene:
method __init__ (line 25) | def __init__(self, args : ModelParams, gaussians : GaussianModel, load...
method save (line 85) | def save(self, iteration):
method getTrainCameras (line 96) | def getTrainCameras(self, scale=1.0):
method getTestCameras (line 99) | def getTestCameras(self, scale=1.0):
FILE: scene/cameras.py
class Camera (line 19) | class Camera(nn.Module):
method __init__ (line 20) | def __init__(self, resolution, colmap_id, R, T, FoVx, FoVy, depth_para...
class MiniCam (line 91) | class MiniCam:
method __init__ (line 92) | def __init__(self, width, height, fovy, fovx, znear, zfar, world_view_...
FILE: scene/colmap_loader.py
function qvec2rotmat (line 43) | def qvec2rotmat(qvec):
function rotmat2qvec (line 55) | def rotmat2qvec(R):
class Image (line 68) | class Image(BaseImage):
method qvec2rotmat (line 69) | def qvec2rotmat(self):
function read_next_bytes (line 72) | def read_next_bytes(fid, num_bytes, format_char_sequence, endian_charact...
function read_points3D_text (line 83) | def read_points3D_text(path):
function read_points3D_binary (line 125) | def read_points3D_binary(path_to_model_file):
function read_intrinsics_text (line 156) | def read_intrinsics_text(path):
function read_extrinsics_binary (line 180) | def read_extrinsics_binary(path_to_model_file):
function read_intrinsics_binary (line 215) | def read_intrinsics_binary(path_to_model_file):
function read_extrinsics_text (line 244) | def read_extrinsics_text(path):
function read_colmap_bin_array (line 273) | def read_colmap_bin_array(path):
FILE: scene/dataset_readers.py
class CameraInfo (line 26) | class CameraInfo(NamedTuple):
class SceneInfo (line 40) | class SceneInfo(NamedTuple):
function getNerfppNorm (line 48) | def getNerfppNorm(cam_info):
function readColmapCameras (line 71) | def readColmapCameras(cam_extrinsics, cam_intrinsics, depths_params, ima...
function fetchPly (line 120) | def fetchPly(path):
function storePly (line 128) | def storePly(path, xyz, rgb):
function readColmapSceneInfo (line 145) | def readColmapSceneInfo(path, images, depths, eval, train_test_exp, llff...
function readCamerasFromTransforms (line 228) | def readCamerasFromTransforms(path, transformsfile, depths_folder, white...
function readNerfSyntheticInfo (line 273) | def readNerfSyntheticInfo(path, white_background, depths, eval, extensio...
FILE: scene/gaussian_model.py
class GaussianModel (line 30) | class GaussianModel:
method setup_functions (line 32) | def setup_functions(self):
method __init__ (line 50) | def __init__(self, sh_degree, optimizer_type="default"):
method capture (line 68) | def capture(self):
method restore (line 84) | def restore(self, model_args, training_args):
method get_scaling (line 103) | def get_scaling(self):
method get_rotation (line 107) | def get_rotation(self):
method get_xyz (line 111) | def get_xyz(self):
method get_features (line 115) | def get_features(self):
method get_features_dc (line 121) | def get_features_dc(self):
method get_features_rest (line 125) | def get_features_rest(self):
method get_opacity (line 129) | def get_opacity(self):
method get_exposure (line 133) | def get_exposure(self):
method get_exposure_from_name (line 136) | def get_exposure_from_name(self, image_name):
method get_covariance (line 142) | def get_covariance(self, scaling_modifier = 1):
method oneupSHdegree (line 145) | def oneupSHdegree(self):
method create_from_pcd (line 149) | def create_from_pcd(self, pcd : BasicPointCloud, cam_infos : int, spat...
method training_setup (line 178) | def training_setup(self, training_args):
method update_learning_rate (line 213) | def update_learning_rate(self, iteration):
method construct_list_of_attributes (line 225) | def construct_list_of_attributes(self):
method save_ply (line 239) | def save_ply(self, path):
method reset_opacity (line 258) | def reset_opacity(self):
method load_ply (line 263) | def load_ply(self, path, use_train_test_exp = False):
method replace_tensor_to_optimizer (line 316) | def replace_tensor_to_optimizer(self, tensor, name):
method _prune_optimizer (line 331) | def _prune_optimizer(self, mask):
method prune_points (line 349) | def prune_points(self, mask):
method cat_tensors_to_optimizer (line 366) | def cat_tensors_to_optimizer(self, tensors_dict):
method densification_postfix (line 388) | def densification_postfix(self, new_xyz, new_features_dc, new_features...
method densify_and_split (line 409) | def densify_and_split(self, grads, grad_threshold, scene_extent, N=2):
method densify_and_clone (line 435) | def densify_and_clone(self, grads, grad_threshold, scene_extent):
method densify_and_prune (line 452) | def densify_and_prune(self, max_grad, min_opacity, extent, max_screen_...
method add_densification_stats (line 471) | def add_densification_stats(self, viewspace_point_tensor, update_filter):
FILE: train.py
function training (line 43) | def training(dataset, opt, pipe, testing_iterations, saving_iterations, ...
function prepare_output_and_logger (line 192) | def prepare_output_and_logger(args):
function training_report (line 214) | def training_report(tb_writer, iteration, Ll1, loss, l1_loss, elapsed, t...
FILE: utils/camera_utils.py
function loadCam (line 20) | def loadCam(args, id, cam_info, resolution_scale, is_nerf_synthetic, is_...
function cameraList_from_camInfos (line 69) | def cameraList_from_camInfos(cam_infos, resolution_scale, args, is_nerf_...
function camera_to_JSON (line 77) | def camera_to_JSON(id, camera : Camera):
FILE: utils/general_utils.py
function inverse_sigmoid (line 18) | def inverse_sigmoid(x):
function PILtoTorch (line 21) | def PILtoTorch(pil_image, resolution):
function get_expon_lr_func (line 29) | def get_expon_lr_func(
function strip_lowerdiag (line 64) | def strip_lowerdiag(L):
function strip_symmetric (line 75) | def strip_symmetric(sym):
function build_rotation (line 78) | def build_rotation(r):
function build_scaling_rotation (line 101) | def build_scaling_rotation(s, r):
function safe_state (line 112) | def safe_state(silent):
FILE: utils/graphics_utils.py
class BasicPointCloud (line 17) | class BasicPointCloud(NamedTuple):
function geom_transform_points (line 22) | def geom_transform_points(points, transf_matrix):
function getWorld2View (line 31) | def getWorld2View(R, t):
function getWorld2View2 (line 38) | def getWorld2View2(R, t, translate=np.array([.0, .0, .0]), scale=1.0):
function getProjectionMatrix (line 51) | def getProjectionMatrix(znear, zfar, fovX, fovY):
function fov2focal (line 73) | def fov2focal(fov, pixels):
function focal2fov (line 76) | def focal2fov(focal, pixels):
FILE: utils/image_utils.py
function mse (line 14) | def mse(img1, img2):
function psnr (line 17) | def psnr(img1, img2):
FILE: utils/loss_utils.py
class FusedSSIMMap (line 24) | class FusedSSIMMap(torch.autograd.Function):
method forward (line 26) | def forward(ctx, C1, C2, img1, img2):
method backward (line 34) | def backward(ctx, opt_grad):
function l1_loss (line 40) | def l1_loss(network_output, gt):
function l2_loss (line 43) | def l2_loss(network_output, gt):
function gaussian (line 46) | def gaussian(window_size, sigma):
function create_window (line 50) | def create_window(window_size, channel):
function ssim (line 56) | def ssim(img1, img2, window_size=11, size_average=True):
function _ssim (line 66) | def _ssim(img1, img2, window, window_size, channel, size_average=True):
function fast_ssim (line 89) | def fast_ssim(img1, img2):
FILE: utils/make_depth_scale.py
function get_scales (line 8) | def get_scales(key, cameras, images, points3d_ordered, args):
FILE: utils/read_write_model.py
class Image (line 52) | class Image(BaseImage):
method qvec2rotmat (line 53) | def qvec2rotmat(self):
function read_next_bytes (line 78) | def read_next_bytes(fid, num_bytes, format_char_sequence, endian_charact...
function write_next_bytes (line 90) | def write_next_bytes(fid, data, format_char_sequence, endian_character="...
function read_cameras_text (line 106) | def read_cameras_text(path):
function read_cameras_binary (line 136) | def read_cameras_binary(path_to_model_file):
function write_cameras_text (line 171) | def write_cameras_text(cameras, path):
function write_cameras_binary (line 190) | def write_cameras_binary(cameras, path_to_model_file):
function read_images_text (line 207) | def read_images_text(path):
function read_images_binary (line 247) | def read_images_binary(path_to_model_file):
function write_images_text (line 296) | def write_images_text(images, path):
function write_images_binary (line 336) | def write_images_binary(images, path_to_model_file):
function read_points3D_text (line 357) | def read_points3D_text(path):
function read_points3D_binary (line 389) | def read_points3D_binary(path_to_model_file):
function write_points3D_text (line 427) | def write_points3D_text(points3D, path):
function write_points3D_binary (line 458) | def write_points3D_binary(points3D, path_to_model_file):
function detect_model_format (line 477) | def detect_model_format(path, ext):
function read_model (line 489) | def read_model(path, ext=""):
function write_model (line 511) | def write_model(cameras, images, points3D, path, ext=".bin"):
function qvec2rotmat (line 523) | def qvec2rotmat(qvec):
function rotmat2qvec (line 545) | def rotmat2qvec(R):
FILE: utils/sh_utils.py
function eval_sh (line 57) | def eval_sh(deg, sh, dirs):
function RGB2SH (line 114) | def RGB2SH(rgb):
function SH2RGB (line 117) | def SH2RGB(sh):
FILE: utils/system_utils.py
function mkdir_p (line 16) | def mkdir_p(folder_path):
function searchForMaxIteration (line 26) | def searchForMaxIteration(folder):
Condensed preview — 32 files, each showing path, character count, and a content snippet. Download the .json file or copy for the full structured content (206K chars).
[
{
"path": ".gitignore",
"chars": 115,
"preview": "*.pyc\n.vscode\noutput\nbuild\ndiff_rasterization/diff_rast.egg-info\ndiff_rasterization/dist\ntensorboard_3d\nscreenshots"
},
{
"path": ".gitmodules",
"chars": 525,
"preview": "[submodule \"submodules/simple-knn\"]\n\tpath = submodules/simple-knn\n\turl = https://gitlab.inria.fr/bkerbl/simple-knn.git\n["
},
{
"path": "LICENSE.md",
"chars": 4662,
"preview": "Gaussian-Splatting License \n=========================== \n\n**Inria** and **the Max Planck Institut for Informatik (MPII"
},
{
"path": "README.md",
"chars": 40764,
"preview": "# 3D Gaussian Splatting for Real-Time Radiance Field Rendering\nBernhard Kerbl*, Georgios Kopanas*, Thomas Leimkühler, Ge"
},
{
"path": "arguments/__init__.py",
"chars": 4199,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "convert.py",
"chars": 5349,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "environment.yml",
"chars": 372,
"preview": "name: gaussian_splatting\nchannels:\n - pytorch\n - conda-forge\n - defaults\ndependencies:\n - cudatoolkit=11.6\n - plyfi"
},
{
"path": "full_eval.py",
"chars": 4526,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "gaussian_renderer/__init__.py",
"chars": 4763,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "gaussian_renderer/network_gui.py",
"chars": 2716,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "lpipsPyTorch/__init__.py",
"chars": 635,
"preview": "import torch\n\nfrom .modules.lpips import LPIPS\n\n\ndef lpips(x: torch.Tensor,\n y: torch.Tensor,\n net_typ"
},
{
"path": "lpipsPyTorch/modules/lpips.py",
"chars": 1151,
"preview": "import torch\nimport torch.nn as nn\n\nfrom .networks import get_network, LinLayers\nfrom .utils import get_state_dict\n\n\ncla"
},
{
"path": "lpipsPyTorch/modules/networks.py",
"chars": 2692,
"preview": "from typing import Sequence\n\nfrom itertools import chain\n\nimport torch\nimport torch.nn as nn\nfrom torchvision import mod"
},
{
"path": "lpipsPyTorch/modules/utils.py",
"chars": 885,
"preview": "from collections import OrderedDict\n\nimport torch\n\n\ndef normalize_activation(x, eps=1e-10):\n norm_factor = torch.sqrt"
},
{
"path": "metrics.py",
"chars": 4143,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "render.py",
"chars": 3314,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "results.md",
"chars": 3451,
"preview": "# Evaluations\nWe evaluated the impact of the features we added on MipNeRF360, Tanks&Temples and Deep Blending datasets. "
},
{
"path": "scene/__init__.py",
"chars": 4550,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "scene/cameras.py",
"chars": 4151,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "scene/colmap_loader.py",
"chars": 11859,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "scene/dataset_readers.py",
"chars": 12521,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "scene/gaussian_model.py",
"chars": 22740,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "train.py",
"chars": 13404,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "utils/camera_utils.py",
"chars": 3651,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "utils/general_utils.py",
"chars": 3971,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "utils/graphics_utils.py",
"chars": 2052,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "utils/image_utils.py",
"chars": 554,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "utils/loss_utils.py",
"chars": 2925,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
},
{
"path": "utils/make_depth_scale.py",
"chars": 3486,
"preview": "import numpy as np\nimport argparse\nimport cv2\nfrom joblib import delayed, Parallel\nimport json\nfrom read_write_model imp"
},
{
"path": "utils/read_write_model.py",
"chars": 22550,
"preview": "# Copyright (c) 2023, ETH Zurich and UNC Chapel Hill.\n# All rights reserved.\n#\n# Redistribution and use in source and bi"
},
{
"path": "utils/sh_utils.py",
"chars": 4371,
"preview": "# Copyright 2021 The PlenOctree Authors.\n# Redistribution and use in source and binary forms, with or without\n# modif"
},
{
"path": "utils/system_utils.py",
"chars": 785,
"preview": "#\n# Copyright (C) 2023, Inria\n# GRAPHDECO research group, https://team.inria.fr/graphdeco\n# All rights reserved.\n#\n# Thi"
}
]
About this extraction
This page contains the full source code of the graphdeco-inria/gaussian-splatting GitHub repository, extracted and formatted as plain text for AI agents and large language models (LLMs). The extraction includes 32 files (193.2 KB), approximately 50.5k tokens, and a symbol index with 164 extracted functions, classes, methods, constants, and types. Use this with OpenClaw, Claude, ChatGPT, Cursor, Windsurf, or any other AI tool that accepts text input. You can copy the full output to your clipboard or download it as a .txt file.
Extracted by GitExtract — free GitHub repo to text converter for AI. Built by Nikandr Surkov.