SCDA-Net: Structure Completion and Density Awareness Network for LiDAR-Based 3D Object Detection

Abstract

As a fundamental task in various application scenarios, including autonomous driving and mobile robotic systems, 3D object detection has received extensive attention from researchers in both academia and industry. However, due to the working principle of LiDAR and external factors such as occlusion, the collected point cloud of the object is usually sparse and incomplete, which affects the performance of 3D object detector. In this letter, a Structure Completion and Density Awareness Network (SCDA-Net) is proposed for 3D object detection from point clouds. Specifically, a structure completion module is designed to predict dense shapes of complete point clouds by leveraging sequence transduction ability of the transformer architecture. Furthermore, we propose a density-aware voxel RoI pooling strategy to introduce density features that reflect the state information of the original objects in refinement stage. By restoring the complete structure of the objects and considering the true distribution of the points in raw point cloud, the proposed method achieves more accurate feature extraction and scene perception. Extensive experimental results on the KITTI and Waymo datasets demonstrate the effectiveness of the proposed SCDA-Net.