We provide our PyTorch implementation of our paper 'Domain Adaptation on Point Clouds via Geometry-Aware Implicits' (IEEE CVPR 2022). By using geometry-awrae implicits representation, our method can align point clouds from different domains in feature space well.

Here we show point clouds from different domains in an image.

Class-wise MMD for the task: ModelNet to ScanNet in PointDA-10 dataset. Diagonal shows source-target distances of the same class. Upper and lower triangular matrices indicate distances between different classes in the source and target domain, respectively. Our method maintains class-wise distances well.

For generating point clouds from GraspNet, we need to render depth maps firstly. Refer to my repo ObjsDepthRender for more information.

From Google Drive Link.

• Python > 3.7
• CUDA > 10.0

We suggest installing torch manually, depending on the python and CUDA versions.

pip install -r requirements.txt

python train.py --name $EXP_NAME --datapath_graspnet $PATH_TO_GRASPNETPC

Part of this implementations is based on DGCNN. We also thank Synchronized-BatchNorm-PyTorch for synchronized batchnorm implementation.

So far, this repo only includes the self-supervised pre-training part. As for domain adaptation, my suggestion is to use GAST which is a sufficient codebase for benchmark comparisons.

If you find this useful for your research, please cite the following paper.

@InProceedings{Shen_2022_CVPR,
    author    = {Shen, Yuefan and Yang, Yanchao and Yan, Mi and Wang, He and Zheng, Youyi and Guibas, Leonidas J.},
    title     = {Domain Adaptation on Point Clouds via Geometry-Aware Implicits},
    booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
    month     = {June},
    year      = {2022},
    pages     = {7223-7232}
}