Point cloud analyzing and processing have attracted extensive attention due to their broad application in numerous sectors. Although many previous deep learning-based frameworks have had significant improvement, they often struggle with processing efficiency and neglect the spatial relationships between points. In this paper, we introduce CSP-Former, a novel framework for point cloud classification and segmentation. Inspired by the impressive strides of self-attention obtained in NLP and CV tasks, we designed a transformer-based network as a backbone for feature extraction; additionally, a fast sampling layer based on compressed sensing theory is proposed to enhance the sampling efficiency, which speeds up the sampling process through only once matrix multiplication. Subsequently, a hierarchical spatial self-attention module is also proposed to better capture the spatial relationships between points, improving segmentation performance. Extensive experiments on the ModelNet40 and ShapeNet part datasets demonstrate that our proposed framework achieves superior performance in point cloud classification and segmentation tasks.
Read full abstract