Enhancing point cloud feature representation via historical node state increments in graph neural networks

With the increasing adoption of 3D sensors, extracting robust and discriminative features from point clouds has become a critical research area, particularly in autonomous driving and robotics. However, the unordered and unstructured nature of point clouds poses significant challenges for feature representation. Current Graph Neural Networks (GNNs) primarily focus on local relationships but overlook the temporal evolution of node states, leading to suboptimal performance in dynamic scenarios. In this paper, we propose an Attention-enhanced Point Graph Convolutional Module (APGCM), which incorporates the lightweight state-space modeling algorithm Mamba to efficiently update features by leveraging historical node state increments. To further enhance feature representation, we introduce a Multi-Decoder Residual Attention (MDRA) module, which reconstructs low-rank features and employs attention mechanisms to capture both local and global relationships. Our approach achieves state-of-the-art results on benchmark datasets such as ModelNet40, ShapeNetPart, and PCN, demonstrating its superiority in point cloud classification, segmentation and shape completion tasks.