yahskapar / mvpsnet

MVPSNet: Fast Generalizable Multi-view Photometric Stereo

Project Page | Paper

Getting the data

sMVPS

The dataset is ~ 200 GB on OneDrive. Please fill out this form to request access. You'll have to be manually added to the viewer list.

Thanks to Daniel Lichy for generating the data.

DiLiGenT-MV

You can download the DiLiGenT-MV dataset from their website.

Usage

Setup

We highly recommend using conda to setup a virual environment. Then run pip install -r requirements.txt to add dependencies.

You can download the pre-trained model here.

Generate meshes

Run the following command to generate 3D point clouds.

export DILIGENT_MV=...  # update the path of DiLiGenT_MV dataset
export CKPT_FILE=...  # update the path of checkpoint
export MESH_FOLDER=...  # path to save the meshes
python run.py --dataset=diligent_mv --testpath=$DILIGENT_MV --loadckpt=$CKPT_FILE --save_folder=$MESH_FOLDER --numlights=10 --numviews=5 $@ 

Then use MeshLab to reconstruct meshes from point clouds using Screened Poisson method.

Evaluation

Our evaluation metrics were implemented based on pytorch3d. You can install it following their instructions.

You also need to install trimesh by running

pip install trimesh

Then run the following commands:

export MESH_PATH={your mesh path}
export GT_PATH={ground truth meshes path}
python evaluate.py --reconstruction_path=$MESH_PATH --ground_truth=$GT_PATH

Citation

@inproceedings{zhao2023mvpsnet,
  title={MVPSNet: Fast Generalizable Multi-view Photometric Stereo},
  author={Zhao, Dongxu and Lichy, Daniel and Perrin, Pierre-Nicolas and Frahm, Jan-Michael and Sengupta, Soumyadip},
  booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision},
  pages={12525--12536},
  year={2023}
}

Acknowledgements

Parts of the code were based on CasMVSNet: https://github.com/hz-ants/cascade-mvsnet/tree/master/CasMVSNet, SDPS-Net https://github.com/guanyingc/SDPS-Net and ShapeAndMaterial https://github.com/dlichy/ShapeAndMaterial.