Awesome

Compact 3D Gaussian Representation for Radiance Field (CVPR 2024 Highlight)

Joo Chan Lee, Daniel Rho, Xiangyu Sun, Jong Hwan Ko, and Eunbyung Park

[Project Page] [Paper(arxiv)] [Extended Paper]

Our code is based on 3D Gaussian Splatting.

Method Overview

We place a specific emphasis on two key objectives: reducing the number of Gaussian points without sacrificing performance and compressing the Gaussian attributes. To this end, we propose a learnable mask strategy that significantly reduces the number of Gaussians while preserving high performance. In addition, we propose a compact but effective representation of view-dependent color by employing a grid-based neural field rather than relying on spherical harmonics. Finally, we learn codebooks to compactly represent the geometric attributes of Gaussian by vector quantization.

Update

Aug. 2024

• We added another post-processing technique: sorting Gaussians in morton order like in Compressed 3DGS. With --comp and --store_npz options, we can achieve a further 10% compression (the size of point_cloud_pp.npz file).
• We extended the compact 3D Gaussian splatting for dynamic scene representation.

Please check out our extended paper and codes for dynamic scenes.

Apr. 2024

We updated saving codes for the result file (.npz) reflecting the actual storage. With --store_npz option, the code will generate point_cloud.npz or point_cloud_pp.npz file according to --comp option, rather than the .ply file. The storage of these files would be almost the same as the estimations in the 'Storage' file (slightly reduced storage for point_cloud_pp.npz due to the npz compression).

Feb. 2024

We additionally implement straightforward post-processing techniques on the model attributes: 1) Applying 8-bit min-max quantization to opacity and hash grid parameters. 2) Pruning hash grid parameters with values below 0.1. 3) Applying Huffman encoding on the quantized opacity and hash parameters, and R-VQ indices.

As a result, our model is further downsized by over 40 % regardless of dataset, achieving more than 25x compression from 3DGS, while maintaining high performance.

Setup

For installation:

git clone https://github.com/maincold2/Compact-3DGS.git --recursive
conda env create --file environment.yml
conda activate c3dgs

We used Mip-NeRF 360, Tanks & Temples, Deep Blending, and NeRF synthetic datasets.

Running

Real-world scenes (e.g., 360, T&T, and DB)

python train.py -s <path to COLMAP> --eval

--comp

Applying post-processings for compression.

--store_npz

Storing npz file reflecting the actual storage.

--lambda_mask

Weight of masking loss to control ma the number of Gaussians masking control factor, 0.01 by default

--mask_lr

Learning rate of masking parameter, 0.01 by default

--net_lr

Learning rate for the neural field, 0.01 by default

--net_lr_step

Step schedule for training the neural field, [5000, 15000, 25000] by default

--max_hashmap

Maximum hashmap size (log) of the neural field, 19 by default

--rvq_size

Codebook size in each R-VQ stage, 64 by default

--rvq_num

The number of R-VQ stages, 6 by default

Refer to other arguments of 3DGS.

NeRF-synthetic scenes

Some different hyper-parameters are required for synthetic scenes.

python train.py -s <path to NeRF Synthetic dataset> --eval --max_hashmap 16 --lambda_mask 4e-3 --mask_lr 1e-3 --net_lr 1e-3 --net_lr_step 25000

Evaluation

python render.py -m <path to trained model> --max_hashmap <max hash size of the model>
python metrics.py -m <path to trained model> 

3DGS Viewer

The original SIBR interactive viewer of 3DGS can not support neural fields for view-dependent color. We would like to support and update this shortly if possible.

Currently, to use the viewer, you have two options: either bypass the neural field for view-dependent color by only applying masking and the geometry codebook, or train neural fields to represent spherical harmonics without inputting view direction (slightly lower performance). After this, you can save the output in a PLY format, similar to 3DGS.

BibTeX

@InProceedings{lee2024c3dgs,
    author    = {Lee, Joo Chan and Rho, Daniel and Sun, Xiangyu and Ko, Jong Hwan and Park, Eunbyung},
    title     = {Compact 3D Gaussian Representation for Radiance Field},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    year      = {2024},
    pages     = {21719-21728}
}