DynaVol-S

Project page | DynaVol | DynaVol-S

Code repository for this paper:
DynaVol-S: Dynamic Scene Understanding through Object-Centric Voxelization and Neural Rendering
Yanpeng Zhao, Yiwei Hao, Siyu Gao, Yunbo Wang^†, Xiaokang Yang

News🎉

-[2024.7.31] DynaVol-S has been integrated into this repo, which significantly improve the model performance in real-world scenes by incorporating DINOv2 features. For the original version aligned with ICLR24 paper, please check the dynavol branch.

-[2024.1.17] DynaVol got accepted by ICLR2024!

Preparation

Installation

git clone -b main https://github.com/zyp123494/DynaVol.git
cd DynaVol
conda create -n dynavol python=3.8
conda activate dynavol

#install pytorch
pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu121

#install Featup
git clone https://github.com/mhamilton723/FeatUp.git
cd FeatUp
pip install -e .

#install requirements
cd ..
pip install -r requirements.txt

Install the correct version of torch_scatter, for torch=2.1.0+cuda12.1, you can simply download the corresponding version from here and run:

pip install torch_scatter-2.1.2+pt21cu121-cp38-cp38-linux_x86_64.whl

Dataset

In our paper, we use:

• synthetic dataset from DynaVol.
• real-world dataset from Hyper-NeRF, NeRF-DS, and D2NeRF.

Experiment

For real-world scenes, first extract DINOv2 features with FeatUp, modify the "img_dir" in extract_dinov2.py then run:

python extract_dinov2.py

Training

Stage 1: Warmup

Cofig files are under the config directory

$ cd warmup
#Synthetic dataset
$ bash run.sh

#Real-world dataset
$ bash run_hyper.sh

Stage 2: CRF postprocess

Modify the "base_path" and "data_dir" in crf_postprocess.py, then run:

python crf_postprocess.py

Stage 3: Joint-optimization

$ cd ../joint_optim
#Synthetic dataset
$ bash run.sh

#Real-world dataset
$ bash run_hyper.sh

Known issues

Semantic probability results can vary slightly even with the same configuration settings and random seed. To achieve optimal results, consider running for multiple times or adjusting the weight entropy.

Citation

If you find our work helps, please cite our paper.

@inproceedings{
      zhao2024dynavol,
      title={DynaVol: Unsupervised Learning for Dynamic Scenes through Object-Centric Voxelization},
      author={Yanpeng Zhao and Siyu Gao and Yunbo Wang and Xiaokang Yang},
      booktitle={The Twelfth International Conference on Learning Representations},
      year={2024},
      url={https://openreview.net/forum?id=koYsgfEwCQ}
}

@misc{zhao2024dynamicsceneunderstandingobjectcentric,
      title={Dynamic Scene Understanding through Object-Centric Voxelization and
      Neural Rendering}, 
      author={Yanpeng Zhao and Yiwei Hao and Siyu Gao and Yunbo Wang and Xiaokang Yang},
      year={2024},
      eprint={2407.20908},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.20908}, 
}