HyperG-PS: Voxel correlation modeling via hypergraph for LiDAR panoptic segmentation

Light-detection-and-ranging (LiDAR) point cloud panoptic segmentation is a fundamental task in autonomous driving since it integrates the tasks of static environmental understanding and dynamic object identification, which have recently gained significant research interest. In this paper, we propose a bottom-up panoptic segmentation framework based on hypergraph learning, named HyperG-PS, which addresses the core problem of LiDAR panoptic segmentation by improving the cluster performance of instance segmentation. Specifically, our proposed method takes the raw LiDAR point cloud as input and utilizes a multi-view feature-extraction framework to fuse the 3D point cloud and 2D BEV features at the voxel level. Afterward, we model the correlation among voxels using a hypergraph to bridge the gap between voxel features and instance labels. We enhance the representation of voxels, thus improving the cluster algorithm performance while directly avoiding predicting point cloud offsets using our hypergraph-learning module. Extensive experiments on the SemanticKITTI and nuScenes datasets demonstrated the superior performance of HyperG-PS compared with state-of-the-art methods.