Skip to content

Latest commit

 

History

History
173 lines (139 loc) · 6.93 KB

README.md

File metadata and controls

173 lines (139 loc) · 6.93 KB
Raw

CorresNeRF: Image Correspondence Priors for Neural Radiance Fields

Yixing Lao · Xiaogang Xu · Zhipeng Cai · Xihui Liu · Hengshuang Zhao
(NeurIPS 2023)

Project Page Paper

For more details, please visit our project page and paper.

Installation

# Environment
conda create -n corresnerf python=3.10
conda activate corresnerf

# Install dependencies
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt
pip install git+https://github.com/Parskatt/DKM.git

Data

  • Download LLFF dataset for NeRF and extract to data/nerf_llff.
  • Download DTU dataset for NeuS and extract to data/neus_dtu.

After extraction, we shall have:

$ tree -lL 2 data
data
|-- nerf_llff
|   |-- fern
|   |-- flower
|   # ... more scenes
|   |-- room
|   `-- trex
`-- neus_dtu
    |-- dtu_scan105
    |-- dtu_scan106
    # ... more scenes
    |-- dtu_scan83
    `-- dtu_scan97

Pre-trained weights

We release pre-trained weights for the paper for each scene. Extract pretrained_weights.zip to logs/ and we shall have:

$ tree -lL 4 logs
logs
|-- nerf_llff
|   |-- fern
|   |   `-- 050000.tar
|   |-- flower
|   |   `-- 050000.tar
|   # ... more scenes
|   |-- room
|   |   `-- 050000.tar
|   `-- trex
|       `-- 050000.tar
`-- neus_dtu
    |-- dtu_scan105
    |   `-- checkpoints
    |       `-- ckpt_200000.pth
    |-- dtu_scan106
    |   `-- checkpoints
    |       `-- ckpt_200000.pth
    # ... more scenes
    |-- dtu_scan83
    |   `-- checkpoints
    |       `-- ckpt_200000.pth
    `-- dtu_scan97
        `-- checkpoints
            `-- ckpt_200000.pth

Inference with the pre-trained weights

We use the fern scene for NeRF on LLFF and the dtu_scan24 scene for NeuS on DTU as example.

# Inference with NeRF. This will load the latest checkpoint (050000.tar in this case).
# - The rendered images will be saved in logs/nerf_llff/fern/renderall_050000.
python -m src.tools.run_nerf --config configs/nerf_llff.txt --scene fern --factor 8 --render_all

# Inference with NeuS. This will load the specified checkpoint (ckpt_200000.pth in this case).
# - The rendered images will be saved in logs/neus_dtu/dtu_scan24/renders.
# - The extracted mesh will be saved in logs/neus_dtu/dtu_scan24/meshes.
python -m src.tools.run_neus --conf configs/neus_dtu.conf --case dtu_scan24 --from_checkpoint ckpt_200000.pth render --resolution_level 4
python -m src.tools.run_neus --conf configs/neus_dtu.conf --case dtu_scan24 --from_checkpoint ckpt_200000.pth extract_mesh

Correspondence matching and training

The full pipeline includes:

  • Extract correspondences.
  • Train NeRF / NeuS models.
  • Render with the trained NeRF / NeuS.
  • Extract mesh with the trained NeuS.

For NeRF on LLFF, we use the fern scene as example.

# Matcher
python -m src.tools.run_matcher --enable_matcher --enable_filter --matcher_name dkm --config_name default --num_views 3 --dataset llff --scene_dir data/nerf_llff/fern --corres_dir data_corres/nerf_llff/fern

# Train
python -m src.tools.run_nerf --config configs/nerf_llff.txt --scene fern

# Render
python -m src.tools.run_nerf --config configs/nerf_llff.txt --scene fern --factor 8 --render_all

For NeuS on DTU, we use the dtu_scan24 scene as example.

# Match
python -m src.tools.run_matcher --enable_matcher --enable_filter --matcher_name dkm --config_name default --num_views 3 --dataset dtu --scene_dir data/neus_dtu/dtu_scan24 --corres_dir data_corres/neus_dtu/dtu_scan24

# Train
python -m src.tools.run_neus --conf configs/neus_dtu.conf --case dtu_scan24 --is_continue train

# Render & extract mesh
python -m src.tools.run_neus --conf configs/neus_dtu.conf --case dtu_scan24 --from_checkpoint ckpt_200000.pth render --resolution_level 4
python -m src.tools.run_neus --conf configs/neus_dtu.conf --case dtu_scan24 --from_checkpoint ckpt_200000.pth extract_mesh

Citation

If you find our work useful in your research, please consider citing:

@inproceedings{lao2023corresnerf,
  title     = {{CorresNeRF}: Image Correspondence Priors for Neural Radiance Fields},
  author    = {Lao, Yixing and Xu, Xiaogang and Cai, Zhipeng and Liu, Xihui and Zhao, Hengshuang},
  booktitle = {NeurIPS},
  year      = {2023}
}

This project is built with CamTools.

Built with CamTools