3SNet: Semi-Anchor-Free 3D Object Detector With Slice Attention and Symmetric Features Propagation

This paper investigates 3D object detection from point clouds scene recovered by lidar and proposes a novel semi-anchor-free 3D object detector. Compared with the point-based method, the voxel-based method has more advantages in feature extraction and avoids the complicated neighborhood search. However, voxel-based methods usually require to preset a large number of anchors related to the objects to be detected, which brings the computational cost to the model inference. This paper introduces a framework with the semi-anchor-free mechanism that utilizes the voxel-based method to aid the learning of point features and achieves anchor-free performance in inference. Actually, feature learning still needs to be strengthened for point cloud encoding, and this paper designs a Directional Slice Attention to augment the discriminability of features. Additionally, the region of interest characterization based on Symmetric Feature Propagation is proposed in this paper to alleviate the obstacle brought by the incomplete scanning of objects in autonomous driving scenarios. Extensive experiments on both the KITTI dataset and Waymo Open dataset demonstrate the superiority of 3SNet over state-of-the-art methods.