Cylindrical and Asymmetrical 3D Convolution Networks for LiDAR Segmentation

  • 2020-11-19 18:53:11
  • Xinge Zhu, Hui Zhou, Tai Wang, Fangzhou Hong, Yuexin Ma, Wei Li, Hongsheng Li, Dahua Lin
  • 2

Abstract

State-of-the-art methods for large-scale driving-scene LiDAR segmentationoften project the point clouds to 2D space and then process them via 2Dconvolution. Although this corporation shows the competitiveness in the pointcloud, it inevitably alters and abandons the 3D topology and geometricrelations. A natural remedy is to utilize the3D voxelization and 3D convolutionnetwork. However, we found that in the outdoor point cloud, the improvementobtained in this way is quite limited. An important reason is the property ofthe outdoor point cloud, namely sparsity and varying density. Motivated by thisinvestigation, we propose a new framework for the outdoor LiDAR segmentation,where cylindrical partition and asymmetrical 3D convolution networks aredesigned to explore the 3D geometric pat-tern while maintaining these inherentproperties. Moreover, a point-wise refinement module is introduced to alleviatethe interference of lossy voxel-based label encoding. We evaluate the proposedmodel on two large-scale datasets, i.e., SemanticKITTI and nuScenes. Our methodachieves the 1st place in the leaderboard of SemanticKITTI and outperformsexisting methods on nuScenes with a noticeable margin, about 4%. Furthermore,the proposed 3D framework also generalizes well to LiDAR panoptic segmentationand LiDAR 3D detection.

 

Quick Read (beta)

loading the full paper ...