NerfingMVS: Guided Optimization of Neural Radiance Fields for Indoor Multi-view Stereo
Yi Wei, Shaohui Liu, Yongming Rao, Wang Zhao, Jiwen Lu, Jie Zhou
ICCV 2021 (Oral Presentation)

• Pull NerfingMVS repo.

  git clone --recursive git@github.com:weiyithu/NerfingMVS.git
  

• Install python packages with anaconda.

  conda create -n NerfingMVS python=3.7
  conda activate NerfingMVS
  conda install pytorch==1.7.1 torchvision==0.8.2 torchaudio==0.7.2 -c pytorch
  pip install -r requirements.txt
  

• We use COLMAP to calculate poses and sparse depths. However, original COLMAP does not have fusion mask for each view. Thus, we add masks to COLMAP and denote it as a submodule. Please follow https://colmap.github.io/install.html to install COLMAP in ./colmap folder (Note that do not cover colmap folder with the original version).

• Download 8 ScanNet scene data used in the paper here and put them under ./data folder. We also upload final results and checkpoints of each scene here.
• Run NerfingMVS

  sh run.sh $scene_name
  

  The whole procedure takes about 3.5 hours on one NVIDIA GeForce RTX 2080 GPU, including COLMAP, depth priors training, NeRF training, filtering and evaluation. COLMAP can be accelerated with multiple GPUs.You will get per-view depth maps in ./logs/$scene_name/filter. Note that these depth maps have been aligned with COLMAP poses. COLMAP results will be saved in ./data/$scene_name while others will be preserved in ./logs/$scene_name

• Place your data with the following structure:

  NerfingMVS
  |───data
  |    |──────$scene_name
  |    |   |   train.txt
  |    |   |──────images
  |    |   |    |    001.jpg
  |    |   |    |    002.jpg
  |    |   |    |    ...
  |───configs
  |    $scene_name.txt
  |     ...
  

  train.txt contains names of all the images. Images can be renamed arbitrarily and '001.jpg' is just an example. You also need to imitate ScanNet scenes to create a config file in ./configs. Note that factor parameter controls the resolution of output depth maps. You also should adjust depth_N_iters, depth_H, depth_W in options.py accordingly.
• Run NerfingMVS without evaluation

  sh demo.sh $scene_name
  

  Since our work currently relies on COLMAP, the results are dependent on the quality of the acquired poses and sparse reconstruction from COLMAP.

Our code is based on the pytorch implementation of NeRF: NeRF-pytorch. We also refer to mannequin challenge.

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

@inproceedings{wei2021nerfingmvs,
  author    = {Wei, Yi and Liu, Shaohui and Rao, Yongming and Zhao, Wang and Lu, Jiwen and Zhou, Jie},
  title     = {NerfingMVS: Guided Optimization of Neural Radiance Fields for Indoor Multi-view Stereo},
  booktitle = {ICCV},
  year = {2021}
}