Awesome

StopThePop

Sorted Gaussian Splatting for View-Consistent Real-time Rendering

Lukas Radl^*, Michael Steiner^*, Mathias Parger, Alexander Weinrauch, Bernhard Kerbl, Markus Steinberger <br> ^* denotes equal contribution <br> | Webpage | Full Paper | Video | T&T+DB COLMAP (650MB) | Pre-trained Models (18.13 GB) <br>

This repository contains the official authors implementation associated with the paper "StopThePop: Sorted Gaussian Splatting for View-Consistent Real-time Rendering", which can be found here.

<a href="https://www.tugraz.at/en/home"><img height="100" src="assets/tugraz-logo.jpg"> </a> <img height="100" src="assets/huawei-logo.jpg"> <a href="https://www.tuwien.at/en/"><img height="100" src="assets/tuwien_logo.jpg"> </a>

Abstract: Gaussian Splatting has emerged as a prominent model for constructing 3D representations from images across diverse domains. However, the efficiency of the 3D Gaussian Splatting rendering pipeline relies on several simplifications. Notably, reducing Gaussian to 2D splats with a single view-space depth introduces popping and blending artifacts during view rotation. Addressing this issue requires accurate per-pixel depth computation, yet a full per-pixel sort proves excessively costly compared to a global sort operation. In this paper, we present a novel hierarchical rasterization approach that systematically resorts and culls splats with minimal processing overhead. Our software rasterizer <b>effectively eliminates popping artifacts and view inconsistencies</b>, as demonstrated through both quantitative and qualitative measurements. Simultaneously, our method mitigates the potential for cheating view-dependent effects with popping, ensuring a more authentic representation. Despite the elimination of cheating, our approach achieves comparable quantitative results for test images, while increasing the consistency for novel view synthesis in motion. Due to its design, our hierarchical approach is <b>only 4% slower</b> on average than the original Gaussian Splatting. Notably, enforcing consistency enables a reduction in the number of Gaussians by approximately half with nearly identical quality and view-consistency. Consequently, rendering performance is nearly doubled, making our approach 1.6x faster than the original Gaussian Splatting, with a 50% reduction in memory requirements.

Overview

Our repository is built on 3D Gaussian Splatting: For a full breakdown on how to get the code running, please consider 3DGS's Readme.

The project is split into submodules, each maintained in a separate github repository:

• StopThePop-Rasterization: A clone of the original diff-gaussian-rasterization that contains our CUDA hierarchical rasterizer implementation
• SIBR_StopThePop: A clone of the SIBR Core project, containing an adapted viewer with our additional settings and functionalities
• PoppingDetection: A self-contained module for our proposed popping detection metric

Licensing

The majority of the projects is licensed under the "Gaussian-Splatting License", with the exception of:

• PoppingDetection: MIT License
• StopThePop header files: MIT License
• FLIP: BSD-3 license

There are also several changes in the original source code. If you use any of our additional functionalities, please cite our paper and link to this repository.

Cloning the Repository

The repository contains submodules, thus please check it out with

# HTTPS
git clone https://github.com/r4dl/StopThePop --recursive

Setup

Local Setup

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 stopthepop

Note: This process assumes that you have CUDA SDK 11 installed. Optionally, you can use CUDA 12 and Pytorch 2.1, by using environment_cuda12.yml instead of environment.yml.

Subsequently, install the CUDA rasterizer:

pip install submodules/diff-gaussian-rasterization

Note: This can take several minutes. If you experience unreasonably long build times, consider using our fast build mode.

Running

Our implementation includes 4 flavors of Gaussian Splatting:

Note: Our hierarchical rasterizer is both faster and more view-consistent compared to the naïve KBuffer method.

The train.py script takes a .json config file as the argument --splatting_config, which should contain the following information (this example is also the default config.json, if none is provided):

{
  "sort_settings": 
  {
    "sort_mode": 0,    // Global (0), Per-Pixel Full (1), Per-Pixel K-Buffer (2), Hierarchical (3)
    "sort_order": 0,   /* Viewspace Z-Depth (0), Worldspace Distance (1), 
                          Per-Tile Depth at Tile Center (2), Per-Tile Depth at Max Contrib. Pos. (3) */
    "queue_sizes": 
    {
      "per_pixel": 4,  // Used for: Per-Pixel K-Buffer and Hierarchical
      "tile_2x2": 8,   // Used only for Hierarchical
      "tile_4x4": 64   // Used only for Hierarchical
    }
  },
  "culling_settings": 
  {
    "rect_bounding": false,            // Bound 2D Gaussians with a rectangle (instead of a square)
    "tight_opacity_bounding": false,   // Bound 2D Gaussians by considering their opacity value
    "tile_based_culling": false,       /* Cull Tiles where the max. contribution is below the alpha threshold;
                                          Recommended to be used together with Load Balancing*/
    "hierarchical_4x4_culling": false, // Used only for Hierarchical
  },
  "load_balancing": false,      // Use load balancing for per-tile calculations (culling, depth, and duplication)
  "proper_ewa_scaling": false,  /* Proper dilation of opacity, as proposed by Yu et al. ("Mip-Splatting");
                                   Model also needs to be trained with this setting */
}

These values can be overwritten through the command line. Call python train.py --help to see all available options. At the start of training, the provided arguments will be written into the output directory. The render.py script uses the config.json in the model directory per default, with the option to overwrite through the command line.

To train different example models (see the corresponding config files for the used settings), run:

# Our Hierarchical Rasterizer, as proposed in StopThePop
python train.py --splatting_config configs/hierarchical.json -s <path to COLMAP or NeRF Synthetic dataset>
# Vanilla 3DGS
python train.py --splatting_config configs/vanilla.json -s <path to COLMAP or NeRF Synthetic dataset>
# Per-Pixel K-Buffer Sort (Queue Size 16)
python train.py --splatting_config configs/kbuffer.json -s <path to COLMAP or NeRF Synthetic dataset>

--opacity_decay

Train with Opacity Decay - this results in comparable image metrics with significantly fewer Gaussians. We used --opacity_decay 0.9995 for the reported results in our paper.

--splatting_config

Full config to specify the flavor of Gaussian Splatting. See configs/ for pre-defined configs.

--sort_mode

Specify Sort Mode - must be one of {GLOBAL,PPX_FULL,PPX_KBUFFER,HIER}

--sort_order

Specify Sort Order - must be one of {Z_DEPTH,DISTANCE,PTD_CENTER,PTD_MAX}

--tile_4x4

Specify size of 4x4 tile queue - only needed if using sort_mode HIER, only 64 supported.

--tile_2x2

Specify size of 2x2 tile queue - only needed if using sort_mode HIER, only {8,12,20} supported.

--per_pixel {1,2,4,8,12,16,20,24}

Specify size of per-pixel queue: If using sort_mode HIER, only {4,8,16} supported. If using sort_mode KBUFFER, all values are supported.

--rect_bounding

Bound 2D Gaussians with a rectangle instead of a circle - must be one of {True,False}

--tight_opacity_bounding

Bound 2D Gaussians by considering their opacity - must be one of {True,False}

--tile_based_culling

Cull complete tiles based on opacity - must be one of {True,False} (recommended with Load Balancing)

--hierarchical_4x4_culling

Cull Gaussians for 4x4 subtiles - must be one of {True,False}, only when using sort_mode HIER

--load_balancing {True,False}

Perform per-tile computations cooperatively (e.g. duplication) - must be one of {True,False}

--proper_ewa_scaling

Dilation of 2D Gaussians as proposed by Yu et al. ("Mip-Splatting") - must be one of {True,False}

--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.

--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.

--compute_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.

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:

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 and gaussian count
python metrics.py -m <path to trained model> # Compute error metrics on renderings

Our repository additionally permits rendering of Depth, visualized with the Turbo colormap. To render depth, run

python render.py -m <path to trained model> --render_depth

<a href="https://www.inria.fr/"><img width="49%" src="assets/00008.png"> </a> <a href="https://univ-cotedazur.eu/"><img width="49%" src="assets/00007.png"> </a>

--render_depth

Flag to enable depth rendering.

--skip_train

Flag to skip rendering the training set.

--skip_test

Flag to skip rendering the test set.

--model_path / -m

Path to the trained model directory you want to create renderings for.

--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.

--compute_cov3D_python

Flag to make pipeline render with computed 3D covariance from PyTorch instead of ours.

--model_paths / -m

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> --config <splatting config file>

Pre-Trained Models

If you want to evaluate our pre-trained models, you have to download the source datsets and indicate their location to render.py, just as done here:

python render.py -m <path to pre-trained model> -s <path to COLMAP dataset>

Alternatively, you can modify the source_path with the cfg_args-file and manually insert the correct path.

Note: We included our models for StopThePop and 3DGS, which were used in our evaluation: to minimize file size, we only include the final checkpoint. We also include the final, rendered images, hence you can reproduce our results easily.

Download Results

--opacity_decay (float)

Train with Opacity Decay - this results in comparable image metrics with significantly fewer Gaussians. We used --opacity_decay 0.9995 for the reported results in our paper.

--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.

Performance

Our proposed optimization imply a significant performance improvement, even for vanilla 3DGS. Here is a framerate comparison for two exemplary scenes in Full HD resolution, run on an NVIDIA RTX 4090 with CUDA 11.8.

Note: For 3DGS, 4x4 Tile Culling is not an option.

Interactive Viewers

Following 3DGS, we provide 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. Our modified viewer contains additional debug modes, and options to disable several of our proposed optimizations. The settings on startup are based on the config.json file in the model directory (if it exists). The implementation is hosted here. Hardware requirements and setup steps are identical to 3DGS, hence, refer to the corresponding README for details.

Quick Setup: Installation from Source

If you cloned with submodules (e.g., using --recursive), the source code for the viewers is found in SIBR_viewers.

Windows

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.

Ubuntu 22.04

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

STOPTHEPOP_FASTBUILD

For performance reasons, we use templates for several of our options, causing very long compile times for our submodule and SIBR. Hence, we provide a STOPTHEPOP_FASTBUILD option in submodules/diff-gaussian-rasterization/cuda_rasterizer/rasterizer.h. Simply uncomment

// #define STOPTHEPOP_FASTBUILD

This solely compiles the default options for our method, which should be sufficient. If you further want to reduce the compile time, simply specify the exact CUDA_ARCHITECTURE in the submodules/diff-gaussian-rasterization/CMakeLists.txt.

Note: For SIBR, the corresponding CMakeLists.txt is located in SIBR_viewers/extlibs/CudaRasterizer/CudaRasterizer/CMakeLists.txt, and rasterizer.h is located in SIBR_viewers/extlibs/CudaRasterizer/CudaRasterizer/cuda_rasterizer/rasterizer.h

Running the Real-Time Viewer

https://github.com/r4dl/StopThePop/assets/45897040/5e763600-c0d9-4055-b664-0b9ea342a248

Popping Detection

Our popping detection method is a self-contained module, hosted here, and is included as a submodule.<br> For more information on how to run the method, consult the submodules README.

FAQ

Please consider 3DGS's FAQ, contained in their README. In addition, several issues are also covered on 3DGS's issues page. We will update this FAQ as issues arise.