3D object detection combining semantic and geometric features from point clouds

Background: 3D object detection based on point clouds in road scenes has attracted much attention recently. The voxel-based methods voxelize the scene to regular grids, which can be processed with advanced feature learning frameworks convolutional layers for semantic learning. point-based extract geometric of due coordinate reservations. combination two is effective detection. However, current use a head anchors classification and localization. Although preset cover entire scene, it not suitable tasks larger multiple categories objects, limitation voxel size. Additionally, misalignment between predicted confidence proposals Regions Interest (ROI) selection bring obstacles detection. Methods: We investigate voxel-to-point module that captures features proposed paper. conducive small-size objects avoids presets inference stage. Moreover, adjustment center-boundary-aware solve regions interest selection. Results: method achieved state-of-the-art results Karlsruhe Institute Technology Toyota Technological (KITTI) dataset. Actually, as September 19, 2021, our ranked 1st Bird Eyes View (BEV) cyclists tagged difficulty level ‘easy’, 2nd ‘moderate’. Conclusions: propose an end-to-end two-stage detector module.