This repository contains a reproduction of the Gaussian-Splatting software, originally developed by Inria and the Max Planck Institut for Informatik (MPII). The reproduction is written in C++ and CUDA. I have used the source code from the original repo as blueprint for my first implementation. The original code is written in Python and PyTorch.
I embarked on this project to deepen my understanding of the groundbreaking paper on 3D Gaussian splatting, by reimplementing everything from scratch.
The materials, code, and assets within this repository are intended solely for educational, training, or research purposes. They may not be utilized for commercial activities without explicit authorization. Any unauthorized commercial usage, distribution, or licensing of this repository's content is strictly forbidden. I am not the license holder for the original implementation. This is derived work. For detailed terms, please refer to the license section.
- [2023-09-11]:
- Resolution parameter -r [2,4,8] introduced. For instance, truck scene trains now in about 52 seconds.
- [2023-09-11]:
- Optimization parameters can be now configured in the
opitimization_params.json
located in parameter folder. - Still struggling with more floaters without autograd. Its harder to get it right than I thought.
- Optimization parameters can be now configured in the
- [2023-09-07]:
- Introduced the PSNR metric for a more accurate comparison of image quality relative to the reference implementation.
- There hasn't been much activity on the master branch lately. I've been focusing intently on the cuda-optimizer branch to eliminate autograd. I faced challenges with disrupted training and encountered several bugs. However, I've now successfully removed autograd, and everything appears to be in order. Additionally, the image quality has improved. Yet, these changesācoupled with optimizations removed for debugging purposesāhave significantly impacted performance. I won't merge back until the performance matches or surpasses that of the master branch.
- [2023-08-29]:
- Better Image loading error description by paulmelis
- I've spent some time working on manual loss derivatives with the aim of completely eliminating the need for autograd. The derivation appears to be accurate, as confirmed by a comparison with a Python implementation. The transition to our custom Adam implementation is still underway, but I hope to complete it by tomorrow.
- Bug fixing :(
This project is a derivative of the original Gaussian-Splatting software and is governed by the Gaussian-Splatting License, which can be found in the LICENSE file in this repository. The original software was developed by Inria and MPII.
Please be advised that the software in this repository cannot be used for commercial purposes without explicit consent from the original licensors, Inria and MPII.
NVIDIA GeForce RTX 4090
tandt/truck:
~87 seconds for 7000 iterations (my implementation 2023-08-18)
~90 seconds for 7000 iterations (my implementation 2023-08-17)
~100 seconds for 7000 iterations (my implementation 2023-08-16)
~120 seconds for 7000 iterations (my implementation 2023-08-16)
~122 seconds for 7000 iterations (original PyTorch implementation)
NVIDIA GeForce RTX 3090
tandt/truck:
~180 seconds for 7000 iterations (Latest 2023-08-17)
~200 seconds for 7000 iterations (2023-08-16)
NVIDIA GeForce RTX 3050 (Ubuntu 20.04)
tandt/truck:
~725.400sec seconds for 7000 iterations (Latest 2023-12-26)
While completely unoptimized, the gains in performance, though modest, are noteworthy.
=> Next Goal: Achieve 60 seconds for 7000 iterations in my implementation
- Linux (tested with Ubuntu 22.04), windows probably won't work.
- CMake 3.24 or higher.
- CUDA 11.7 or higher (might work with a lower version, has to be manually set and tested).
- Python with development headers.
- libtorch: You can find the setup instructions in the libtorch section of this README.
- Other dependencies will be handled by the CMake script.
- NVIDIA GPU with CUDA support. Successfully tested so far are RTX 4090, RTX A5000, 3090Ti and A100. With 3080Ti there is an outstanding issue (#21) with larger datasets.
- So far, the lowest compute capability tested was 8.0.
It might work with other NVIDIA GPUs as well, but these are mostly untested. If you do successfully run on such hardware please post a message in the Discussions section of the repo.
git clone --recursive https://github.com/MrNeRF/gaussian-splatting-cuda
cd gaussian-splatting-cuda
wget https://download.pytorch.org/libtorch/cu118/libtorch-cxx11-abi-shared-with-deps-2.0.1%2Bcu118.zip
unzip libtorch-cxx11-abi-shared-with-deps-2.0.1+cu118.zip -d external/
rm libtorch-cxx11-abi-shared-with-deps-2.0.1+cu118.zip
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -- -j
The dataset is not included in this repository. You can download it from the original repository under the following link: tanks & trains. Then unzip it in the data folder.
The 3D Gaussian Splatting CUDA Implementation
provides a suite of command-line options to facilitate easy and customizable execution. Below are the available options:
-
-h, --help
Display this help menu. -
-d, --data_path [PATH]
Specify the path to the training data. -
-f, --force
Force overwriting of output folder. If not set, the program will exit if the output folder already exists. -
-o, --output_path [PATH]
Specify the path to save the trained model. If this option is not specified, the trained model will be saved to the "output" folder located in the root directory of the project. -
-i, --iter [NUM]
Specify the number of iterations to train the model. Although the paper sets the maximum number of iterations at 30k, you'll likely need far fewer. Starting with 6k or 7k iterations should yield preliminary results. Outputs are saved every 7k iterations and also at the end of the training. Therefore, even if you set it to 5k iterations, an output will be generated upon completion.
-
--empty-gpu-cache Empty CUDA memory after ever 100 iterations. Attention! This has a considerable performance impact
-
--enable-cr-monitoring
Enable monitoring of the average convergence rate throughout training. If done, it will stop optimizing when the average convergence rate is below 0.008 per default after 15k iterations. This is useful for speeding up the training process when the gain starts to dimish. If not enabled, the training will stop after the specified number of iterations--iter
. Otherwise its stops when max 30k iterations are reached. -
-c, --convergence_rate [RATE]
Set custom average onvergence rate for the training process. Requires the flag--enable-cr-monitoring
to be set.
To run the 3D Gaussian Splatting CUDA Implementation
with specified data path, output path, and iterations, use the following command:
$ ./build/gaussian_splatting_cuda -d /path/to/data -o /path/to/output -i 1000
For now, you will need the SIBR view
git clone --recursive https://gitlab.inria.fr/sibr/sibr_core SIBR_core
cd SIBR_viewers
cmake -B build .
cmake --build build --target install --config Release -- -j
cd ..
Then, you can view the results with:
./SIBR_viewers/install/bin/SIBR_gaussianViewer_app -m output
Contributions are welcome! I want to make this a community project.
Some ideas for relative straight forward contributions:
- Revamp the README.
- Add a proper config file or cmd line config.
I want to get rid of some heavy dependencies:
- Replace glm with custom matrix operations
- Replace the few Eigen with some custom matrix operations
Advanced contributions or long term goals:
- Build a renderer to view training output in real time and to replace SIBR viewer.
- Look into gtsfm to replace colmap dependency
- CUDA optimization
- Build a proper viewer for the training output (maybe with XR support?).
Own ideas are welcome as well!
Below are some guidelines to help ensure our project remains effective and consistent.
-
Getting Started with Contributions:
- I've marked several beginner-friendly issues as good first issues. If you're new to the project, these are great places to start.
- For those looking to contribute something not currently listed as an issue or propose something in the discussion section. You can direct message me on Twitter for a quick chat. Since there are not many contributors at the moment, I'm happy to discuss your ideas and help you get started.
-
Before Submitting Your Pull Request:
- Ensure you've applied
clang-format
to maintain consistent coding style. There is in tools folder a git pre-commit hook. You can just copy it to .git/hooks/pre-commit. It will run clang-format before every commit. - We aim to minimize dependencies. If you're introducing a new one, it's essential to raise an issue for discussion first. There are ongoing efforts to reduce the number of dependencies, and your understanding in this area is appreciated.
- Ensure you've applied
-
Key Principles for Contributions:
- Speed: We want to be the fastest 3D gaussian splatting implementation on this planet. Being lightning fast is key! I want instant training!
- Quality: Always prioritize high-quality rendering output. Never compromise quality for speed.
- Usability: We want to have a nice user experience. We're still perfecting this, and your contribution can make a difference!
-
Dataset Contributions:
- If you have a unique dataset that you believe will be an excellent addition and that is eye popping, we'd love to see it! Remember, we're aiming to showcase exceptional datasets. We want to show off the best of the best. If you're unsure, feel free to raise an issue for discussion first.
Together, with your contributions, we can make this project stand out. Thank you for being a part of this journey!
Initially, I utilized libtorch to simplify the development process. Once the implementation is stable with libtorch, I will begin replacing torch elements with my custom CUDA implementation.
Here is random collection of things that have to be described in README later on
- Needed for simple-knn:
bash sudo apt-get install python3-dev
If you utilize this software or present results obtained using it, please reference the original work:
Kerbl, Bernhard; Kopanas, Georgios; LeimkĆ¼hler, Thomas; Drettakis, George (2023). 3D Gaussian Splatting for Real-Time Radiance Field Rendering. ACM Transactions on Graphics, 42(4).
This will ensure the original authors receive the recognition they deserve.
This project is licensed under the Gaussian-Splatting License - see the LICENSE file for details.
Follow me on Twitter if you want to know more about the latest development: https://twitter.com/janusch_patas