Deformable Pyramid R-CNN for 3D object detection (ChinaMM2022)

3D object detector performance has gradually improved, but it still faces huge challenges for accurate detection from point clouds due to incomplete shapes and sparse point clouds. To solve this problem, recent two-stage approaches usually divide the region of interest (RoI) into regular grids and aggregate the internal features of the RoI within a certain radius for refinement proposals. However, for objects with sparse points, only the features inside the RoI cannot satisfy the need for accurate detection. In view of this objective fact, we present a new two-stage 3D object detection framework from point clouds to 3D voxels inside and outside RoI, named Deformable Pyramid R-CNN. Specially, we firstly propose the Voxel Feature Pyramid, which dynamically selects multi-layers of 3D features based on the sparsity of non-empty voxels within the RoI for proposal refinement. Secondly, a new Deformable Voxel RoI Pooling is introduced to abstract rich context information from the voxels of interest outside RoI, to better align semantic information by deformation grid points set with multiple receptive fields for accurate detection. Our method achieves progressive performance in both KITTI Dataset and Waymo Open Dataset. Especially on KITTI Dataset, our method outperforms PV-RCNN by 0.47%, 1.63%, 1.34% on val set easy, moderate and hard levels for car detection, respectively. Code will be available at https://github.com/june2678/DP-RCNN.