Computer Vision News - July 2020

ScanNet Indoor Scene Understanding 39 This year we are focusing on the challenging task of 3D instance segmentation, which is more difficult than semantic segmentation as two chairs or two cabinets that are next to each other need to be separated. Speakers from the top three teams on the benchmark presented their methods (all are papers at the main conference) and told us what are the special techniques they used to tackle this challenging problem. One thing that all these methods have in common is that they use the aforementioned sparse convolutional backbones for feature extraction. Lei Han , from the overall winner team for both instance segmentation and semantic segmentation presented on OccuSeg: Occupancy-aware 3D Instance Segmentation . Other winning teams at the instance segmentation were represented by Li Jiang who presented PointGroup: Dual-Set Point Grouping for 3D Instance Segmentation (see figures for example results), and Francis Engelmann who presented 3D-MPA: Multi Proposal Aggregation for 3D Semantic Instance Segmentation . Figure from Li Jiang’s PointGroup showing an input point cloud and the output with semantic classes and object instances. Best of CVPR 2020