Guanjun Wu ^1*, Taoran Yi ^2*, Jiemin Fang ^2‡, Wenyu Liu ², Xinggang Wang ^2‡✉

¹ School of CS, HUST   ² School of EIC, HUST  

^* Equal Contributions. ^$\ddagger$ Project Lead. ^✉ Corresponding Author.

We propose a dynamic HDR NeRF framework, named as HDR-HexPlane, which can learn 3D scenes from dynamic 2D images captured with various exposures. We further construct a dataset containing multiple dynamic scenes captured with diverse exposures for evaluation.

# create conda environment
conda create --name hdrhexplane python=3.9

# pip install 
pip install -r requirements.txt

Please download all the data from the link:

Make the dataset format like: dataset/lego.

Please change the "datadir" in config based on the locations of downloaded datasets.

python main.py config=config/dnerf_slim_tank.yaml

We provide several config files under config folder for different datasets and models.

With render_test=True, render_path=True, results at test viewpoint are automatically evaluated and validation viewpoints are generated after reconstruction.

python main.py config=config/dnerf_slim_airplane.yaml systems.ckpt="checkpoint/path" render_only=True

Some insights about neural voxel grids and dynamic scenes reconstruction originate from TiNeuVox and hexplane. If you find this repository/work helpful in your research, welcome to cite these papers and give a ⭐.

@inproceedings{wu2024fast,
  title={Fast High Dynamic Range Radiance Fields for Dynamic Scenes},
  author={Wu, Guanjun and Yi, Taoran and Fang, Jiemin and Liu, Wenyu and Wang, Xinggang},
  booktitle={2024 International Conference on 3D Vision (3DV)},
  pages={862--872},
  year={2024},
  organization={IEEE}
}

Our code is hugely influenced by hexplane and many other projects. We would like to acknowledge them for making great code openly available for us to use. All the datasets are provided for academic studies only.