PHN: PointHeatNet for Point Clouds Classification and Segmentation

In recent years, point cloud learning has gained significant attention in fields such as computer vision, autono-mous driving, and robotics. The unordered and sparse nature of point cloud data makes traditional 2D image processing methods difficult to apply directly. The emergence of the PointNet model enabled end-to-end learning of point cloud data by proposing a new method to process unordered point clouds directly. Subse-quently, PointNet++ improved local feature learning through a hierarchical abstraction approach, but it still faces challenges such as high computational complexity and limited generalization capability. To further enhance the performance of point cloud feature extraction, this paper proposes the PointHeat network, inspired by the vHeat model based on the heat conduction equation used in 2D image processing. PointHeat leverages the Heat Conduction Operator (HCO) to process point cloud features in the frequency domain, achieving superior local and global feature extraction capabilities. Experimental results show that PointHeat improves accuracy by 1.7% in classification tasks and 1.2% in segmentation tasks, effectively enhancing the analytical performance of point cloud data.