RE-Det3D: RoI-enhanced 3D object detector

Abstract

This paper proposes a two-stage point-based detector referred to as RoI-enhanced 3D object detector (RE-Det3D), which enhances the RoI-feature extraction ability for two-stage point-based 3D object detectors. The proposed detector is characterized by an RoI shape-aware module (RSAM), RoI keypoints sampling module (RKSM), and RoI self-attention module (RAM). More precisely, the RSAM uses the proposal which possesses accurate boundary information as auxiliary supervision, to reinforce the framework to be more aware of the object shape. Simultaneously, RKSM uses the tilted sampling strategy to obtain more representative keypoints from the RoI. Afterwards, the plug-and-play module RAM cascades the set-abstraction and self-attention, exploring the interactions of keypoints and aggregating local features, to produce discriminative feature representations for 3D box refinement. Comprehensive experiments are conducted on the widely used KITTI dataset and the latest large-scale dataset (ONCE). The results demonstrate that the RE-Det3D can bolster the baseline by a significant margin and achieve comparable accuracy as several strong voxel-based detectors.