hkchengrex
commented
11 months ago eval.py supports using multiple input masks containing different objects (ref. YouTubeVOS evaluation does this). If the different camera angles contain substantially different content, you might be better off creating another evaluation session whenever the scene changes (e.g., treat them as separate videos). Otherwise the memory from a previous scene might creep in.
As a side note, our new project Cutie might work better.
Hello, thanks for your contribution, regarding provinding a mask without an interface and when the video has different camera angles, like consider a clip from a show where scene might change with different characters(considering person is the class of interest and multiple people appearing at the same time which I should track simaltaneously), how can I do that to provide the first frame mask of each scene and camera angle change?