A Sparse Point Cloud 3D Object Detection Method Based on Size-Aware Attention Mechanism and BEV Backbone Network

Abstract

Addressing the challenge of effectively detecting distant and small-scale objects in sparse point cloud scenarios, this paper proposes an improved 3D object detection method based on a size-aware attention mechanism and a BEV backbone network, achieving performance enhancements within the PV-RCNN framework. First, the Size-Aware Dual-Branch Attention mechanism is introduced during the keypoint feature fusion stage. By estimating the voxel density in keypoint neighborhoods, it dynamically determines target scales. Channel attention and spatial attention are applied differentially to enhance keypoint features, thereby improving the model's ability to represent multi-scale features. Second, the BaseBEV backbone network is designed and integrated into the BEV feature extraction stage, replacing the original SECOND and SECONDFPN structures to enhance multi-scale feature modeling capabilities in the BEV space. Experimental validation on the KITTI 3D object detection dataset demonstrates that the proposed method achieves consistent improvements in both BEV detection and 3D detection tasks. Ablation studies further validate the effectiveness of each improved module, confirming that our approach effectively enhances 3D object detection performance in sparse point cloud environments.