/nerfren

Code release for NeRFReN: Neural Radiance Fields with Reflections (CVPR 2022).

Primary LanguagePython

NeRFReN: Neural Radiance Fields with Reflections

This is the code release for our CVPR2022 paper, NeRFReN: Neural Radiance Fields with Reflections.

Update

  • 07/28/2022: Initial code release.
  • 08/01/2022: Pretrained models for all RFFR scenes are released.

Setup

  • Install PyTorch>=1.8
  • Install other dependencies: pip install -r requirements.txt
  • Download our Real Forward Facing with Reflections (RFFR) dataset from Google Drive, and extract to load/
  • (Optional) Download pretrained models from Google Drive, and extract to checkpoints/

The correct file structure should be like:

checkpoints/
  |
  -- art1_pretrain/
  |
  -- ...
load/
  |
  -- rffr/
    |
    -- art1/
    |
    -- ...

Training

We provide training scripts for all the 6 RFFR scenes in scripts/nerfren. Run the scripts to perform training:

sh scripts/nerfren/train_art1.sh

To train the NeRF baseline, run scripts/nerf/train.sh and specify the scene as arguments:

sh scripts/nerf/train.sh art1

The training process by default uses all available GPUs. Set CUDA_VISIBLE_DEVICES environment variable to specify the GPUs to be used.

The network checkpoints and visualizations are stored in checkpoints/ by default, and tensorboard logs can be found in runs/.

Testing

The testing process generates images from spiral poses for visualization. To test a pretrained model, run scripts/nerfren/test_pretrain.sh and specify the scene as arguments:

sh scripts/nerfren/test_pretrain.sh art1

To test on our pretrained models, please make sure you have downloaded the checkpoints and organized the files correctly as demonstrated in the Setup section.

The testing results are saved to results/ by default.

Citation

If you find our work useful, please cite:

@InProceedings{Guo_2022_CVPR,
    author    = {Guo, Yuan-Chen and Kang, Di and Bao, Linchao and He, Yu and Zhang, Song-Hai},
    title     = {NeRFReN: Neural Radiance Fields With Reflections},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {18409-18418}
}

Acknowledgement

Part of the code is borrowed or adapted from the following great codebases: