复旦大学数字信号处理与传输实验室的冯辉老师的合作论文HgCA: Hypergraph neural network with cross-attention for point cloud analysis被Neurocomputing期刊录用

论文摘要：Point cloud classification and segmentation are two fundamental tasks in 3D point cloud analysis, aiming to recognize object categories and partition shapes or scenes into semantically meaningful components, with wide-ranging applications in intelligent systems. Graph Convolutional Networks (GCNs) are commonly used for processing point clouds, as they excel at extracting topological information. However, GCNs focus on lower-order pairwise relationships, which limits their ability to model complex higher-order dependencies inherent in the point cloud. To overcome this limitation, we propose a Hypergraph Neural Network with Cross-Attention (HgCA) for point cloud analysis, which integrates enhanced lower-order geometric modeling and adaptive higher-order semantic learning using an attention mechanism. First, to enhance the extraction of lower-order neighborhood geometric information, we design a Geometric Edge-Conditioned Aggregation (GECA) module that modulates feature aggregation using local geometric cues. Second, a hierarchical clustering strategy is employed for adaptive hypergraph structure learning, thereby capturing higher-order semantic dependencies beyond pairwise relationships. Finally, we introduce a bipartite hypergraph with a cross-attention mechanism to bridge lower-order geometric features with higher-order semantic dependencies encoded in the hypergraph structure. Extensive experiments demonstrate the strong performance of HgCA on both segmentation and classification tasks, achieving 87.5% mIoU on ShapeNet-Part, 72.8% mIoU on S3DIS, and 94.5% OA on ModelNet40.

论文作者：Xinxin Hou,  Hui Feng, Zhengpin Li, Shubo Zhou, Jian Wang, Xue-Qin Jiang