DualGroup for 3D instance and panoptic segmentation

Lin Zhao, Sijia Chen, Xu Tang, Wenbing Tao

Abstract: Highlights•We introduce an encoded center-shifted vector learning to improve the learning of smaller instances.•We propose a dual hierarchical grouping algorithm to generate instance proposals.•A new bottom-up segmentation framework, DualGroup, is proposed to tackle the 3D instance segmentation.•The extensive experimental results on ScanNet v2 and S3DIS datasets demonstrate the effectiveness of DualGroup.