Diffusion Unit: Interpretable Edge Enhancement and Suppression Learning for 3D Point Cloud Segmentation

3D point clouds are discrete samples of continuous surfaces which can be usedfor various applications. However, the lack of true connectivity information,i.e., edge information, makes point cloud recognition challenging. Recentedge-aware methods incorporate edge modeling into network designs to betterdescribe local structures. Although these methods show that incorporating edgeinformation is beneficial, how edge information helps remains unclear, makingit difficult for users to analyze its usefulness. To shed light on this issue,in this study, we propose a new algorithm called Diffusion Unit (DU) thathandles edge information in a principled and interpretable manner whileproviding decent improvement. First, we theoretically show that DU learns toperform task-beneficial edge enhancement and suppression. Second, weexperimentally observe and verify the edge enhancement and suppressionbehavior. Third, we empirically demonstrate that this behavior contributes toperformance improvement. Extensive experiments and analyses performed onchallenging benchmarks verify the effectiveness of DU. Specifically, our methodachieves state-of-the-art performance in object part segmentation usingShapeNet part and scene segmentation using S3DIS. Our source code is availableat https://github.com/martianxiu/DiffusionUnit.