Self supervised Learning of Point Clouds via Orientation Estimation

Code for the paper "Self-supervised Learning of Point Clouds via Orientation Estimation", 3DV 2020.

Installation

We use tensorflow 1.14, python 3.6 and h5py.

Usage

To train the rotation prediction model with 18 random rotation angles run:

python train_rotation_prediction.py --num_angles 18 --dataset shapenet --log_dir log_rotation_shapenet --no_transformation_loss --no_input_transform --no_feature_transform;

Please see train_rotation_prediction.py for more arguments.

Here is a pretrained model of running the script above.

To train a linear SVM for object classification on top of the pretrained weights run:

python SVM.py --model pointnet_cls_rot_svm_scoped --svm_c 0.001 --dataset modelnet10 --model_path PATH/TO/CHECKPOINT

Datasets

The ModelNet data will be automatically downloaded to the data/ directory.

The ShapeNet data can be downloaded from here. Set SHAPENET_DIR in provider.py to the ShapeNet folder.

Citation

If you use the code in this repository in your paper, please consider citing:

@article{poursaeed2020self,
  title={Self-supervised Learning of Point Clouds via Orientation Estimation},
  author={Poursaeed, Omid and Jiang, Tianxing and Qiao, Quintessa and Xu, Nayun and Kim, Vladimir G.},
  journal={arXiv preprint arXiv:2008.00305},
  year={2020}
}