Deformable Shape-aware Point Generation for 3D Object Detection

Kai Wang, Xiaowei Zhang; Proceedings of the Asian Conference on Computer Vision (ACCV), 2024, pp. 2699-2715

As a fundamental task of the perception system for autonomous driving, LiDAR-based 3D object detection often suffers from the absence of incomplete object shapes under long distances and occlusion. 3D object detectors, based on point cloud completion methods, significantly improve detection performance by generating pseudo-point clouds. However, due to the scarcity of guidance provided by the geometric shape and orientation information of objects, it is more challenging to recover the accurate surface shape of objects. Motivated by this, we propose Deformable Shape-aware Point Generation(DSaPG) for 3D object detection. Specifically, we design a Geometry-guided Region Proposal Network (GgRPN) including a heatmap-guided shape branch and an orientation alignment supervision, which contributes to the generation of high-quality proposals. Moreover, we design a Density-aware Deformable Point Generation(DDPG) module, which generates points by encoding probability density at each grid ball query and deformation learning to accurately recover the shape of objects. The deformable-augmented generated points can more effectively represent the shape features of objects for 3D bounding box predictions. Experiments demonstrate that the proposed DSaPG achieves competitive performance on the KITTI dataset and the Waymo Open Dataset. The code will be available at https://github.com/Wkk121/DSaPG.