Video Object of Interest Segmentation

In this work, we present a new computer vision task named video object of interest segmentation (VOIS). Given and target image interest, our objective is to simultaneously segment track all objects in the that are relevant image. This problem combines traditional with an additional indicating content users concerned with. Since no existing dataset perfectly suitable for task, specifically construct large-scale called LiveVideos, which contains 2418 pairs images live videos instance-level annotations. addition, propose transformer-based method task. We revisit Swin Transformer design dual-path structure fuse features. Then, transformer decoder employed generate proposals tracking from fused Extensive experiments on LiveVideos show superiority proposed method.