Visualizations of Reconstruction Results. For each action sample, the ground truth is on the left, and the reconstruction result at 75% masking ratio is on the right.

The code is tested with Python 3.7.12, PyTorch 1.7.1, GCC 9.4.0, and CUDA 10.2.

Compile the CUDA layers for PointNet++ and Chamfer_Distance_Loss:

cd modules
python setup.py install

cd ./extensions/chamfer_dist
python setup.py install

We thank the authors of related repositories:

1. PSTNet: https://github.com/hehefan/Point-Spatio-Temporal-Convolution
2. P4Transformer: https://github.com/hehefan/P4Transformer
3. MAE: https://github.com/facebookresearch/mae