Awesome

Compact 3D Gaussian Splatting for Static and Dynamic Radiance Fields

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

[Project Page] [Extended Paper]

This is an extended version of C3DGS (CVPR24) for dynamic scenes.

Our code is based on STG@d9833e6.

Setup

Installation

git clone https://github.com/maincold2/Dynamic_C3DGS.git
cd Dynamic_C3DGS
bash script/setup.sh

Dataset Preprocessing

We used Neural 3D and Technicolor datasets.

conda activate colmapenv
# Neural 3D
python script/pre_n3d.py --videopath <location>/<scene>
# Technicolor
python script/pre_technicolor.py --videopath <location>/<scene>

Running

Training

conda activate dynamic_c3dgs
python train.py --quiet --eval --config config/n3d_ours/<scene>.json --model_path <path to save model> --source_path <location>/<scene>/colmap_0

--comp

Applying post-processings for compression.

--store_npz

Storing npz file reflecting the actual storage.

For example, we can apply post-processings and save the compressed results in .npz format by the following command.

python train.py --quiet --eval --config configs/n3d_ours/cook_spinach.json --model_path log/ours_cook_spinach --source_path <location>/cook_spinach/colmap_0 --comp --store_npz

Command line arguments can also set these.

lambda_mask

Weight of masking loss to control the number of Gaussians, 0.0005 by default

mask_lr

Learning rate of the masking parameter, 0.01 by default

net_lr

Learning rate for the neural field, 0.001 by default

net_lr_step

Step schedule for training the neural field

max_hashmap

Maximum hashmap size (log) of the neural field

rvq_size_geo

Codebook size in each R-VQ stage for geometric attributes

rvq_num_geo

The number of R-VQ stages for geometric attributes

rvq_size_temp

Codebook size in each R-VQ stage for temporal attributes

rvq_num_temp

The number of R-VQ stages for temporal attributes

mask_prune_iter

Pruning inteval after densification, 1000 by default

rvq_iter

The iteration at which R-VQ is implemented

Evaluation

# Neural 3D
python test.py --quiet --eval --skip_train --valloader colmapvalid --configpath config/n3d_ours/<scene>.json --model_path <path to model>
# Technicolor
python test.py --quiet --eval --skip_train --valloader technicolorvalid --configpath config/techni_ours/<scene>.json --model_path <path to model>

To run the original STG, use stg's config files (e.g., configs/n3d_stg/<scene>.json).

Real-time Viewer

Without --comp and --store_npz options, our code saves the models in the original STG format, which can be used for the STG's viewer.

Acknowledgements

A great thanks to the authors of 3DGS and STG for their amazing work. For more details, please check out their repos.

BibTeX

@article{Lee_2024_C3DGS,
  title={Compact 3D Gaussian Splatting for Static and Dynamic Radiance Fields},
  author={Lee, Joo Chan and Rho, Daniel and Sun, Xiangyu and Ko, Jong Hwan and Park, Eunbyung},
  journal={arXiv preprint arXiv:2408.03822},
  year={2024}
}
@InProceedings{Lee_2024_CVPR,
  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}
}