Abstract: Dynamic human-object interaction segmentation from videos focuses on the dense labeling of sequences involving multiple human-object interaction classes. Previous approaches often combined ...