Panagiotis Papantonakis Georgios Kopanas, Bernhard Kerbl, Alexandre Lanvin, George Drettakis
| Webpage | Full Paper | Datasets (TODO) | Video | Other GRAPHDECO Publications | FUNGRAPH project page |
This repository contains the code of the paper "Reducing the Memory Footprint of 3D Gaussian Splatting", which can be found here. We also provide the configurations to train the models mentioned in the paper, as well as the evaluation script that produces the results.
Abstract: 3D Gaussian splatting provides excellent visual quality for novel view synthesis, with fast training and real-time rendering; unfortunately, the memory requirements of this method for storing and transmission are unreasonably high. We first analyze the reasons for this, identifying three main areas where storage can be reduced: the number of 3D Gaussian primitives used to represent a scene, the number of coefficients for the spherical harmonics used to represent directional radiance, and the precision required to store Gaussian primitive attributes. We present a solution to each of these issues. First, we propose an efficient, resolution-aware primitive pruning approach, reducing the primitive count by half. Second, we introduce an adaptive adjustment method to choose the number of coefficients used to represent directional radiance for each Gaussian primitive, and finally a codebook-based quantization method, together with a half-float representation for further memory reduction. Taken together, these three components result in a ×27 reduction in overall size on disk on the standard datasets we tested, along with a ×1.7 speedup in rendering speed. We demonstrate our method on standard datasets and show how our solution results in significantly reduced download times when using the method on a mobile device
@Article{papantonakisReduced3DGS,
author = {Papantonakis, Panagiotis and Kopanas, Georgios and Kerbl, Bernhard and Lanvin, Alexandre and Drettakis, George},
title = {Reducing the Memory Footprint of 3D Gaussian Splatting},
journal = {Proceedings of the ACM on Computer Graphics and Interactive Techniques},
number = {1},
volume = {7},
month = {May},
year = {2024},
url = {https://repo-sam.inria.fr/fungraph/reduced_3dgs/}
}
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 repository contains submodules, thus please check it out with
# SSH
git clone [email protected]:graphdeco-inria/reduced-3dgs.git --recursive
or
# HTTPS
git clone https://github.com/graphdeco-inria/reduced-3dgs --recursive
The codebase consists an extension to the original 3DGS project that you can find here.
This project builts on top of the previous one,
so almost all of the functionality of the original code exists to this one,
with one major exception.
In this project the point cloud .ply
format is altered in a way that it is no longer compatible to the old one.
There is included a detailed explanation of the differences later on this README
and also there is provided a script that converts the point clouds from the old to the new format.
This README contains the instructions on how to setup and run the code, taken directly from the orinal project, modified accordingly to the needs of this project, wherever it is necessary.
The use of the train.py will result in a point cloud that is stored in a .ply
format.
Before the layout of this file was one row per primitive,
that contained a series of parameters, namely
- 3 floats for position (
x
,y
,z
) - 3 floats for normal (
nx
,ny
,nz
) - 3 floats for DC color (
f_dc_0
,f_dc_1
,f_dc_2
) - 45 floats for SH coefficients (
f_rest_0
, ...,f_rest_44
) - 1 float for opacity (
opacity
) - 3 float for scaling (
scale_0
,scale_1
,scale_2
) - 4 float for rotation (
rot_0
,rot_1
,rot_2
,rot_3
)
The normal parameters were a burden, as they were not used anywhere in the algorithm.
In this project, apart from removing these 3, unused normal parameters, we introduce these 3 changes:
With our SH culling technique we end up with sets of primitives that have a different number of SH coefficients.
So, we group primitives based on the SH bands they use and store only the coefficients that are used.
The SH coefficients are placed in the same place as before (after the DC colour and before opacity).
The point clouds are stored one after the other,
with an increasing number of SH band.
That is,
if we have a converged model with N_i = number of primitives that have i SH bands (i being 0, 1, ..., max_sh_degree),
the first N_0 rows of the .ply
file will contain no SH coefficients,
the next N_1 rows will contain 9,
the next N_2 rows will contain 9 + 15 = 24
and the final N_3 rows will contain 9 + 15 + 21 = 45.
We provide a script that converts the previous .ply
files to this new format,
by creating a file that has 0 primitives in all SH bands except the final one.
To run the optimizer, simply use
python update_old_ply_format.py -p <path of the ply file to convert> -n <name of the new, converted ply file>
Command Line Arguments for update_old_ply_format.py
Path of the ply file to convert #### --model_path / -m
Name of the newly created point cloud
Max order of SH bands. 0 if just DC colour, 1 if 3 coefficients per channel etc.
The codebook quantization induces some additional changes.
First of all,
since we are using 256 entries for each quantized parameter,
we store only an 8-bit unsigned int for each one of them.
Secondly,
we add the codebook entries at the bottom of the .ply
file
as rows containing 1 entry of each codebook (256 rows in total).
The codebooks are ordered as follows:
features_dc
for the 3 DC color parametersfeatures_rest_0
, ...features_rest_<max SH coeffiecient group>
. The per channel coefficients are encoded together so for a 3 SH band model the total number of codebooks would be 15.opacity
for the opacityscaling
for the 3 parameters of scale.rotation_re
for the real part of the rotationrotation_im
for the 3 imaginary parts of the rotation quaternion
So, the nth row will contain the nth entry of the above codebooks, in the given order.
If the half-float quantization has been employed,
the codebook entries as well as the position parameters
are stored in half precision.
That means that 16 bits are used instead of 32,
and subsequently,
float16 are stored instead of float32.
However,
as the .ply
format does not allow for float16 typed numbers,
the parameters are pointer casted to int16 and stored as such.
The optimizer uses PyTorch and CUDA extensions in a Python environment to produce trained models.
- CUDA-ready GPU with Compute Capability 7.0+
- 24 GB VRAM (to train to paper evaluation quality)
- Please see FAQ for smaller VRAM configurations
- 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
Our default, provided install method is based on Conda package and environment management:
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:
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
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.
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.
To run the optimizer, simply use
python train.py -s <path to COLMAP or NeRF Synthetic dataset>
Command Line Arguments for train.py
Path to the source directory containing a COLMAP or Synthetic NeRF data set.
Path where the trained model should be stored (output/<random>
by default).
Alternative subdirectory for COLMAP images (images
by default).
Add this flag to use a MipNeRF360-style training/test split for evaluation.
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.
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.
Add this flag to use white background instead of black (default), e.g., for evaluation of NeRF Synthetic dataset.
Order of spherical harmonics to be used (no larger than 3). 3
by default.
Flag to make pipeline compute forward and backward of SHs with PyTorch instead of ours.
Flag to make pipeline compute forward and backward of the 3D covariance with PyTorch instead of ours.
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.
Debugging is slow. You may specify an iteration (starting from 0) after which the above debugging becomes active.
Number of total iterations to train for, 30_000
by default.
IP to start GUI server on, 127.0.0.1
by default.
Port to use for GUI server, 6009
by default.
Space-separated iterations at which the training script computes L1 and PSNR over test set, 7000 30000
by default.
Space-separated iterations at which the training script saves the Gaussian model, 7000 30000 <iterations>
by default.
Space-separated iterations at which to store a checkpoint for continuing later, saved in the model directory.
Path to a saved checkpoint to continue training from.
Flag to omit any text written to standard out pipe.
Spherical harmonics features learning rate, 0.0025
by default.
Opacity learning rate, 0.05
by default.
Scaling learning rate, 0.005
by default.
Rotation learning rate, 0.001
by default.
Number of steps (from 0) where position learning rate goes from initial
to final
. 30_000
by default.
Initial 3D position learning rate, 0.00016
by default.
Final 3D position learning rate, 0.0000016
by default.
Position learning rate multiplier (cf. Plenoxels), 0.01
by default.
Iteration where densification starts, 500
by default.
Iteration where densification stops, 15_000
by default.
Limit that decides if points should be densified based on 2D position gradient, 0.0002
by default.
How frequently to densify, 100
(every 100 iterations) by default.
How frequently to reset opacity, 3_000
by default.
Influence of SSIM on total loss from 0 to 1, 0.2
by default.
Percentage of scene extent (0--1) a point must exceed to be forcibly densified, 0.01
by default.
Factor that controls the alpha regularization
Factor that controls the spherical harmonics regularization
Enables the additional culling scheme
Determines the culling scheme used
Factors that controls how many standard deviations from the mean should be filetered for culling
Factor that determines the size of the spherical region around each point to calculate the redundancy score
Value of redundancy score under which culling stops
Enables the removal of points that are no longer splatted due to low opacity
Prevents the zero-out of gradients for densified points in densification iterations
How many densification cycles pass before culling is performed
Threshold for the colour distance for SH culling
Threshold on the standard deviation for the SH culling
Uses the variable SH bands variant rasterizer. USED ONLY FOR INFERENCE
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. You can find our SfM data sets for Tanks&Temples and Deep Blending here. 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).
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:
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
Command Line Arguments for render.py
Path to the trained model directory you want to create renderings for.
Flag to skip rendering the training set.
Flag to skip rendering the test set.
Flag to skip calculating fps
Flag to omit any text written to standard out pipe.
Types of models to test. Choices between "baseline" and "quantised_half"
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.
Path to the source directory containing a COLMAP or Synthetic NeRF data set.
Alternative subdirectory for COLMAP images (images
by default).
Add this flag to use a MipNeRF360-style training/test split for evaluation.
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.
Add this flag to use white background instead of black (default), e.g., for evaluation of NeRF Synthetic dataset.
Flag to make pipeline render with computed SHs from PyTorch instead of ours.
Flag to make pipeline render with computed 3D covariance from PyTorch instead of ours.
Command Line Arguments for metrics.py
Space-separated list of model paths for which metrics should be computed.
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:
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.
python full_eval.py -o <directory with pretrained model> --skip_training -m360 <mipnerf360 folder> -tat <tanks and temples folder> -db <deep blending folder>
Command Line Arguments for full_eval.py
Flag to skip training stage.
Flag to skip rendering stage.
Flag to skip metrics calculation stage.
Argument passed to render.py to skip FPS measurement
Directory to put renderings and results in, ./eval
by default, set to pre-trained model location if evaluating them.
Path to MipNeRF360 source datasets, required if training or rendering.
Path to Tanks&Temples source datasets, required if training or rendering.
Path to Deep Blending source datasets, required if training or rendering.
Names of experiments that the script will use. Defaults to "full_final"
Names of the scenes that the script will use.
We provide two interactive viewers for our method: remote and real-time. Our viewing solutions are based on the SIBR framework, developed by the GRAPHDECO group for several novel-view synthesis projects.
- 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)
- 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)
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.
CMake should take care of your dependencies.
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.
You will need to install a few dependencies before running the project setup.
# 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
Backwards compatibility with Focal Fossa is not fully tested, but building SIBR with CMake should still work after invoking
git checkout fossa_compatibility
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.
remoteviewer.mp4
After extracting or installing the viewers, you may run the compiled SIBR_remoteGaussian_app[_config]
app in <SIBR install dir>/bin
, e.g.:
./<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>
.
Primary Command Line Arguments for Network Viewer
Argument to override model's path to source dataset.
IP to use for connection to a running training script.
Port to use for connection to a running training script.
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.
Flag to load source dataset images to be displayed in the top view for each camera.
realtimeviewer.mp4
After extracting or installing the viewers, you may run the compiled SIBR_gaussianViewer_app[_config]
app in <SIBR install dir>/bin
, e.g.:
./<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 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.
Primary Command Line Arguments for Real-Time Viewer
Path to trained model.
Specifies which of state to load if multiple are available. Defaults to latest available iteration.
Argument to override model's path to source dataset.
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.
Flag to load source dataset images to be displayed in the top view for each camera.
Index of CUDA device to use for rasterization if multiple are available, 0
by default.
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).
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
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
python convert.py -s <location> --skip_matching [--resize] #If not resizing, ImageMagick is not needed
Command Line Arguments for convert.py
Flag to avoid using GPU in COLMAP.
Flag to indicate that COLMAP info is available for images.
Location of the inputs.
Which camera model to use for the early matching steps, OPENCV
by default.
Flag for creating resized versions of input images.
Path to the COLMAP executable (.bat
on Windows).
Path to the ImageMagick executable.
OpenXR is supported in the branch gaussian_code_release_openxr Within that branch, you can find documentation for VR support here.
-
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 -
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, 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.