Work

PillarFocusNet for 3D object detection with perceptual diffusion and key feature understanding

Year: 2025

Type: article

Abstract: This study presents PillarFocusNet, a novel network about 3D point cloud object detection that optimizes the PointPillars framework to improve detection performance. First, we propose the Pillar Clustering Sampling Method to address the sparse and uneven distribution of 3D point cloud data. Second, we introduce the Mixed Pooling Dilated Convolution Layer (MPDC layer) to enhance feature extraction. Finally, we integrate the Space-Channel Synergistic Enhancement Module (SCS-EM) to improve feature representation across both spatial and channel dimensions. Results from KITTI dataset experiments demonstrate that PillarFocusNet enhances bbox, bev, and 3d detection performance by 1.3%, 2.9%, and 3.4%, respectively, when compared to PointPillars. The code and models are publicly available at https://github.com/Gaoooyhan/PillarFocusNet .

Source: Scientific Reports

DOI: 10.1038/s41598-025-92338-5