PointStack: A Point Cloud Processing Network with Enhanced Global Feature Aggregation

Abstract

Point cloud data, characterized by its rich geometric information, dense data representation, and flexible topological structure, finds extensive application in fields such as computer graphics, computer vision, and robotics perception. The rapid advancement and upgrade of three-dimensional point cloud acquisition devices have made the acquisition of point cloud data increasingly convenient. This, in turn, has spurred an urgent need for research on point cloud data processing. The irregularity and disorderliness of point clouds necessitate the design of network models that differ from those used for image processing. Presently, most network models for point cloud processing focus primarily on the design for local feature extraction, even though the effectiveness of these improvements has become somewhat marginal. In response to this, a stacked diffusion network model designed to capture global context has been proposed. This model excels at capturing global information and can effectively integrate it with local information. Extensive experiments conducted on standard benchmark datasets such as ModelNet40 and ShapeNetPart indicate that this model, in comparison to DGCNN, improves point cloud classification by 1.2%, achieving an Overall Accuracy (OA) of 94.1%. Furthermore, it enhances point cloud part segmentation by 1.3%, reaching an mIOU evaluation metric of 86.5%. When compared to other advanced methods, this model also exhibits noticeable improvements in point cloud classification and part segmentation.