GACM-Net: A Geometry-Aware Contextual Memory Network for Efficient 3D Point Cloud Understanding

Three-dimensional point cloud understanding plays an important role in autonomous perception, robotic navigation, and LiDAR-based remote sensing. However, the irregular and unordered nature of point clouds makes it challenging to model long-range contextual dependencies while preserving geometric awareness, particularly under noise, occlusion, and non-uniform sampling. To address these challenges, we propose a Geometry-Aware Contextual Memory Network (GACM-Net) for 3D point cloud analysis. Specifically, an Adaptive Geometric Prior Encoding (AdGPE) module is introduced to dynamically coordinate multiple geometric priors, including absolute coordinates, center-relative coordinates, and distance-based cues, thereby enhancing geometry-aware contextual interaction during long-range propagation. Furthermore, a Structure-Aware Point Memory Unit (SAPM) is designed to achieve stable contextual memory learning through normalized gate interaction, peephole memory regulation, gated candidate filtering, and residual feature propagation. Based on SAPM, a Bidirectional Structure-Aware Point Memory module (BiSAPM) further captures complementary geometric dependencies from opposite propagation directions, improving contextual completeness and structural consistency for irregular point cloud representations. Extensive experiments on ModelNet40, ScanObjectNN, and ShapeNetPart demonstrate that GACM-Net achieves competitive classification and part segmentation performance with a compact model size and favorable computational efficiency. The results further show that the proposed framework provides a good balance among accuracy, efficiency, and robustness for 3D point cloud understanding.