It would be useful have Video AI analyze a video and then perform automatic segmentation of the different parts of the video contents. This segmentation could be rendered out as a cryptomatte pass for use in post-production softwares. This would prove to be very useful for automatic rotoscoping.
This is no doubt a big endeavor but would be quite useful for many people. Maybe this is included in Video AI as a new model type or maybe it’s a new software altogether.