Geometric Back-Projection Network for Point Cloud Classification

As the basic task of point cloud analysis, classification is fundamental but always challenging. To address some unsolved problems existing methods, we propose a network that captures geometric features clouds for better representations. achieve this, on one hand, enrich information points in low-level 3D space explicitly. On other apply CNN-based structures high-level feature spaces to learn local context implicitly. Specifically, leverage an idea error-correcting feedback structure capture comprehensively. Furthermore, attention module based channel affinity assists map avoid possible redundancy by emphasizing its distinct channels. The performance both synthetic and real-world datasets demonstrate superiority applicability our network. Comparing with state-of-the-art approach balances accuracy efficiency.