STC: Spatio-Temporal Contrastive Learning for Video Instance Segmentation

Video Instance Segmentation (VIS) is a task that simultaneously requires classification, segmentation, and instance association in video. Recent VIS approaches rely on sophisticated pipelines to achieve this goal, including RoI-related operations or 3D convolutions. In contrast, we present simple efficient single-stage framework based the segmentation method CondInst by adding an extra tracking head. To improve accuracy, novel bi-directional spatio-temporal contrastive learning strategy for embedding across frames proposed. Moreover, instance-wise temporal consistency scheme utilized produce temporally coherent results. Experiments conducted YouTube-VIS-2019, YouTube-VIS-2021, OVIS-2021 datasets validate effectiveness efficiency of proposed method. We hope can serve as strong baseline other instance-level video tasks.