Decoupling Dynamic Monocular Videos for Dynamic View Synthesis (TVCG 2024)

Demo

Left to right: input video, fix time and synthesize from novel viewpoints, fix viewpoint and synthesize from novel timestamps, synthesize from both novel viewpoints and timestamps. Please find more results in Demo.

Getting Started

Dependencies

Linux
Anaconda 3
Python 3.8
CUDA 11.1
RTX 3090

Installation

git clone https://github.com/mengyou2/DecoulpingNeRF.git
cd DecoulpingNeRF
conda create -n denerf python=3.8
conda activate denerf
pip install -r requirements.txt

Training

To train the model by running

python run_nerf.py --config ./configs/config_xxxx.txt

Testing

To train the model by running

python run_nerf.py --config ./configs/config_xxxx.txt --render_only --ft_path ./logs/xxxx/200000.tar

Acknowledgments

Our code is build upon NeRF, NeRF-pytorch, NSFF and DynamicNeRF.

Citation

@article{you2024decoupling,
  title={Decoupling dynamic monocular videos for dynamic view synthesis},
  author={You, Meng and Hou, Junhui},
  journal={IEEE Transactions on Visualization and Computer Graphics},
  year={2024},
  publisher={IEEE}
}