Abstract: We propose a deep learning framework for various 3D vision tasks, which takes a point cloud as input. The convolution is a basic operator for feature extraction in deep learning. However, it ...