EDGCNet: Joint dynamic hyperbolic graph convolution and dual squeeze-and-attention for 3D point cloud segmentation

Abstract

This paper proposes a novel 3D point cloud segmentation network called EDGCNet. Structurally, the network combines the encoder–decoder structure and graph convolution to improve the processing efficiency and the exploration of point-to-point correlation. In addition, it contains two sub-modules: the dynamic hyperbaric graph convolution module and the dual squeeze-and-attention module. To mitigate the issue of incorrect inference in existing graph convolution networks for multi-category hybrid regions, the dynamic hyperbolic graph convolution module maps the features captured in the Euclidean space to the hyperbolic space to aggregate the neighborhoods and dynamically update the edge weights. Compared to Euclidean geometry, hyperbolic embedding with learnable curvature can deeply fit the geometric topology of point cloud by transforming the manifold shape to enrich feature representation. The dual squeeze-and-attention module recalibrates the original features by modeling the correlation between channels from both global and local perspectives to improve the feature shift caused by the deepening of the convolution. Empirical experiments on three public datasets verify that EDGCNet has an excellent segmentation effect on both artificial shapes and real scenes. Furthermore, ablation studies and confirmatory experiments show that the modules in EDGCNet are contributing, low-complexity and robust.