Cat-3D

Code repository for our paper "Planes vs. Chairs: Category-guided 3D shape learning without any 3D cues"

[Project Page] [Paper]

The repository currently includes training and evaluation code for ShapeNet-13 experiments.

Dependencies

Please install the dependencies by running

conda env create --file requirements.yaml
cd external/chamfer3D
python3 setup.py install
cd ../..

You may need to modify CUDA_HOME accordingly for the compilation.

Dataset

ShapeNet

Please download the required data by running

cd data
bash download_data.sh

Make sure your data/NMR_Dataset folder is structured as follows:

├── 02691156/
|   ├── 1a04e3eab45ca15dd86060f189eb133/
|   |   ├── image/
|   |   |   ├── 0000.png
|   |   |   ├── ...
|   |   |   ├── 0023.png
|   |   ├── mask/
|   |   |   ├── 0000.png
|   |   |   ├── ...
|   |   |   ├── 0023.png
|   |   ├── cameras.npz
|   |   ├── pointcloud.npz
|   |   ├── pointcloud3.npz
|   ├── softras_train.lst
|   ├── ...
├── ...

Training

Please first pretrain the model with a spherical SDF similar to SDF-SRN

python train.py --yaml=options/shapenet13.yaml --name=pretrain --pretrain

Then please run

python train.py --yaml=options/shapenet13.yaml

The training logs and visualizations are saved at the output directory.

Evaluating

To evaluate the model for Chamfer Distance and F-score, Please run

python evaluate.py --yaml=options/shapenet13.yaml --eval.vox_res=128 --resume

The evaluation results are saved at the output directory.

References

If you are using our code, please consider citing our paper.

@article{huang2022planes,
  title={Planes vs. Chairs: Category-guided 3D shape learning without any 3D cues},
  author={Huang, Zixuan and Stojanov, Stefan and Thai, Anh and Jampani, Varun and Rehg, James M},
  journal={arXiv preprint arXiv:2204.10235},
  year={2022}
}

This project contains a modified version of SDF-SRN (MIT License) - Copyright (c) 2020 Chen-Hsuan Lin. Please also cite their great work if you use this codebase.

@inproceedings{lin2020sdfsrn,
  title={SDF-SRN: Learning Signed Distance 3D Object Reconstruction from Static Images},
  author={Lin, Chen-Hsuan and Wang, Chaoyang and Lucey, Simon},
  booktitle={Advances in Neural Information Processing Systems ({NeurIPS})},
  year={2020}
}