LiDAR-based 3D Object Detection with Cylindrical Representation

Abstract

Cylindrical representation has shown its superiority for the LiDAR point cloud object detection task[1] for preserving sensor’s streaming property. However, little effort has been devoted to solve shape distortion and scale inconsistency problems of objects in cylindrical view, which degrades the performance. To han- dle the above limitation, we propose a novel 3D detection framework CylinDet which can benefit from the advantages of both anchor-free architecture and cylindrical data representation. First, specifically designed asymmetrical RPN and dilated asymmetric sparse backbone are employed to handle the scale inconsistency by structure-aware context embedding.  Then, a novel cylindrical anchor-free head is proposed to encode the 3D boxes in cylindrical coordinates with range-aware supervision labels.  Finally, our framework has been verified on the public KITTI benchmark and extensive experimental results show that our frame- work outperforms other one-stage baselines with a remarkable margin, especially on small objects (e.g., pedestrian), the performance improvement reaches nearly 8 points on average AP.1