JiangWenPL / FisherRF

FisherRF: Active View Selection and Uncertainty Quantification for Radiance Fields using Fisher Information

FisherRF: Active View Selection and Uncertainty Quantification for Radiance Fields using Fisher Informations

Wen Jiang, Boshu Lei, Kostas Daniilidis
| Project page | arxiv | full paper

Environment setup

This repo is heavely build on 3D Gaussian Splatting. Please follow their repo to set-up and compile most things about 3D Gaussian Splatting. Beside, this code was tested with the following dependencies:

einops==0.7.0
Pillow==10.2.0
pymeshlab==2023.12.post1
scipy==1.12.0
wandb==0.15.12

Clone our repo with submodules:

git clone [e-mail]:JiangWenPL/FisherRF.git --recursive

Install extensions:

pip install submodules/diff-gaussian-rasterization/
pip install submodules/simple-knn/
pip install -e ./diff/ -v

Download Datasets (Optional)

Please get the NeRF-synthetic from: https://drive.google.com/drive/folders/128yBriW1IG_3NJ5Rp7APSTZsJqdJdfc1 (nerf_synthetic.zip). The MipNeRF360 scenes are hosted by the paper authors here

Running our code

Please use scripts under ./scripts/ to run different experiments with different configurations. The first arguments of the script is the path to the scene that contains trainsforms_*.json and the second is the path which you would like to save your experiment. For example:

bash scripts/blender_seq1.sh /PATH/TO/YOUR/DATASET/lego YOUR_EXP_PATH

Citing

If you find this code useful for your research or the use data generated by our method, please consider citing our paper:

@article{Jiang2023FisherRF,
      title={FisherRF: Active View Selection and Uncertainty Quantification for Radiance Fields using Fisher Information},
      author={Wen Jiang and Boshu Lei and Kostas Daniilidis},
      journal={arXiv},
      year={2023}
  }

Acknowledgements

This project builds heavily on 3D Gaussian Splatting and Plenoxels. We thanks the authors for their excellent works! If you use our code, please also consider citing their papers as well.