OKGR: Occluded Keypoint Generation and Refinement for 3D Object Detection

Code

• github.com/mingqj/okgr
  OfficialIn paperpytorch★ 3

Abstract

Lidar-based 3D object detectors utilize point clouds to detect objects in autonomous driving. However, the point clouds are sparse and incomplete, which affects the detectors’ learning of shape knowledge and limits the 3D detection performance. Previous works improve performance through completing object shape at the point level or representation level, such as voxel. The former increases computational burden, while the latter has poor generalization ability to point-based detectors. In this paper, we present an approach, namely Occluded Keypoint Generation and Refinement (OKGR), which is effective to improve 3D detection performance by completing object features at the keypoint level. Specifically, Occluded Keypoint Generation (OKG) generates occluded keypoints to densify raw keypoints and learns the offsets between the generated keypoints and prototypes, while retaining the raw keypoints unchanged. Occluded Keypoint Refinement (OKR) assigns weights to the generated keypoints and conducts these weights to features to obtain high-quality complete features for detection. We apply our approach to two representative detectors, PV-RCNN++ and PDV, and evaluate the detectors on KITTI and Waymo Open Dataset. The experiments show significant performance improvement. Particularly, our OKGR applied on PV-RCNN++ achieves improvements of Pedestrian and Cyclist of +3.19%, +2.53% AP on average difficulty levels on KITTI, and +2.18%, +2.29% mAPH on Waymo Open Dataset. For more information, the supplementary material and code are available at https://github.com/Mingqj/OKGR.

Tasks

Reproductions