SPS-RCNN: Semantic-Guided Proposal Sampling for 3D Object Detection from LiDAR Point Clouds

Three-dimensional object detection using LiDAR has attracted significant attention due to its resilience to lighting conditions and ability to capture detailed geometric information. However, existing methods still face challenges, such as a high proportion of background points in the sampled point set and limited accuracy in detecting distant objects. To address these issues, we propose semantic-guided proposal sampling-RCNN (SPS-RCNN), a multi-stage detection framework based on point-voxel fusion. The framework comprises three components: a voxel-based region proposal network (RPN), a keypoint sampling stream (KSS), and a progressive refinement network (PRN). In the KSS, we propose a novel semantic-guided proposal sampling (SPS) method, which increases the proportion of foreground points and enhances sensitivity to outliers through multilevel sampling that integrates proposal-based local sampling and semantic-guided global sampling. In the PRN, a cascade attention module (CAM) is employed to aggregate features from multiple subnets, progressively refining region proposals to improve detection accuracy for medium- and long-range objects. Comprehensive experiments on the widely used KITTI dataset demonstrate that SPS-RCNN improves detection accuracy and exhibits enhanced robustness across categories compared to the baseline.