Semantic-Guided Sampling and Multi-Head Shape Encoding for Enhanced 3D Object Detection

Currently, voxel-based 3D single-stage detectors have been widely studied, but point-based single-stage methods still require further exploration. This study proposes a novel single-stage 3D point cloud detector that fully utilizes point cloud shape information. In 3D point cloud object detection tasks, existing methods have the following shortcomings: failing to fully utilize semantic and shape information of points during the segmentation of foreground and background points; not effectively utilizing this information when generating high-quality candidate boxes; and neglecting the correlation between channels, leading to limited feature representation capability. These factors collectively result in reduced detection accuracy. To enhance shape information in candidate points, this study adopts the following strategies: introducing a semantic-guided farthest point sampling method during the downsampling stage to increase the sampling rate of foreground points and suppress the generation of background points; applying a multi-head attention mechanism during the shapeEncode stage to effectively encode object shape information; and using a channel attention mechanism to fuse candidate features and shape features, thereby generating higher-quality candidate points. Extensive experiments were conducted on authoritative datasets such as ScanNet V2 and SUN RGB-D. Experimental results show that the proposed method outperforms existing methods on multiple benchmark datasets, fully validating the effectiveness of this framework in 3D point cloud object detection tasks.