Occluded Video Instance Segmentation: A Benchmark

Abstract Can our video understanding systems perceive objects when a heavy occlusion exists in scene? To answer this question, we collect large-scale dataset called OVIS for occluded instance segmentation, that is, to simultaneously detect, segment, and track instances scenes. consists of 296k high-quality masks from 25 semantic categories, where object occlusions usually occur. While human vision can understand those by contextual reasoning association, experiments suggest current cannot. On the dataset, highest AP achieved state-of-the-art algorithms is only 16.3, which reveals are still at nascent stage objects, instances, videos real-world scenario. We also present simple plug-and-play module performs temporal feature calibration complement missing cues caused occlusion. Built upon MaskTrack R-CNN SipMask, obtain remarkable improvement on dataset. The project code available http://songbai.site/ovis .