LiDAR-based 3D object detection with localization reliability enhancement based on CenterPoint
Abstract
3D object detection from LiDAR point clouds is a fundamental perception task in autonomous driving, aiming to accurately localize and classify surrounding objects in sparse and irregular point-cloud data. CenterPoint is a widely adopted anchor-free baseline for LiDAR-based 3D object detection, but it often suffers from the generation and mis-selection of low-quality bounding boxes. To address this limitation, we present RefPoint, a simple yet effective variant that resolves these bottlenecks from three dimensions. First, to address feature representation deficiency, we propose a hybrid Voxel-Mamba (HyVoxM) backbone that leverages Mamba operators for efficient long-range context modeling and improved representation of large-scale objects. Secondly, we propose an IoU-aware head with geometric-mean rescoring to explicitly estimate box quality, thereby suppressing low-quality predictions and retaining accurate boxes during non-maximum suppression. Lastly, we introduce a four-channel BEV corner heatmap to provide explicit geometric supervision for object corners, thereby improving detection stability and localization accuracy. On nuScenes val, RefPoint improves CenterPoint from 58.0% mAP / 65.5% NDS to 64.3% mAP / 69.2% NDS (+6.3 mAP, +3.7 NDS). Furthermore, with a lightweight pre-tracker score gate, it also raises AMOTA from 0.6380 to 0.6862 (+0.0482).