i-am-future / scoda

Implementation of "SCoDA: Domain Adaptive Shape Completion for Real Scans", published in CVPR2023

SCoDA

SCoDA: Domain Adaptive Shape Completion for Real Scans
GAP Lab, Yushuang Wu

Paper - Project Website - Arxiv - Published in CVPR 2023.

Citation

If you find our code or paper useful for your project, please consider citing:

@inproceedings{wu2023scoda,
  title={SCoDA: Domain Adaptive Shape Completion for Real Scans},
  author={Yushuang, Wu and Zizheng, Yan and Ce, Chen and Lai, Wei and Xiao, Li and Guanbin, Li and Yihao, Li and Shuguang, Cui and Xiaoguang, Han},
  booktitle={The IEEE/CVF Computer Vision and Pattern Recognition Conference (CVPR)},
  year={2023},
}

ScanSalon Dataset

We build a new dataset, ScanSalon, for the shape completion of real scans, with 800 mesh-scan pairs in 6 classes: chair, desk, sofa, bed, lamp, car. We provide videos in our project page for data sample visualization.

ScanSalon Data: At Google Drive (Updated on 9/22/2023) (paired data only).
ShapeNet Data: Turn to ShapeNet for synthetic data download.
Further details about ScanSalon: Please refer to the README and scripts in the ScanSalon zip package.
Refer to here to find out how do we extract point clouds for this dataset.

Installation

Our implementation is based on IF-Net as the basic framework for reconstruction. Please refer to the "Install" part of IF-Net for the installation of our method.

Running

1. Following the steps in data_processing/mesh-fusion to get the water-tight ScanSalon meshes.
   
2. Following the steps in Mesh2PC to get simulated scans from ShapeNet meshes.
   
3. Following the steps in data_processing/process.sh to preprocess all data.
   
4. Run train_ddp.sh to train the model in a parrallel way.
   
5. After training by around 30-50 epochs, run generate_ddp.sh to generate meshes in the test set.