Hausdorff point convolution with geometric priors

Developing point convolution for irregular clouds to extract deep features remains challenging. Current methods evaluate the response by computing set distances which account only spatial alignment between two sets, but not quite their underlying shapes. Without a shape-aware response, it is hard characterize 3D geometry of cloud efficiently with compact kernels. In this paper, we advocate use modified Hausdorff distance as measure calculating convolutional responses. The technique present, coined (HPC), shape-aware. We show that HPC constitutes powerful feature learning rather four types geometric priors further develop an HPC-based neural network (HPC-DNN). Task-specific can be achieved tuning weights combining shortest input and kernel sets. also realize hierarchical designing multi-kernel multi-scale encoding. Extensive experiments demonstrate HPC-DNN outperforms strong baselines (e.g., KPConv), achieving 2.8% mIoU performance boost on S3DIS 1.5% SemanticKITTI semantic segmentation task.