VGPCNet: viewport group point clouds network for 3D shape recognition

3D point cloud recognition is fundamental and popular in vision perceptual systems such as autonomous driving, robotics, and virtual reality. Due to the sparse distribution and irregularity of point clouds, previous 3D point networks perform convolution on nearby points, ignoring the long-range dependence on the global structure. To solve this problem, we propose a Viewport Group Point Cloud Network for 3D Shape Recognition (VGPCNet) in which features are grouped according to viewports instead of local neighbor points to model the long-range global context. First, we propose to use viewport as proxy to capture both local and global features from an outside view of the object. The related points are grouped by visibility attribute effectively and efficiently which can not only capture the inside local geometry details but also obtain the global structure from the outside viewport. Second, we use a graph-based feature consolidation module to enhance the viewport features by modeling interactions between different viewports. Finally, to aggregate a global representation from multiple viewport features, we propose a novel attention-based feature aggregation module. We evaluate our VGPCNet on three widely used benchmarks including ModelNet40/10, ScanObjectNN, and ShapeCore55 for shape classification and retrieval tasks. Extensive experiments have demonstrated the effectiveness and superior performance (94.1% on ModelNet40) of our method over state-of-the-art methods.