is there an open source or even closed source model that does animation extraction from video in suboptimal conditions really well? video related, but imagine them rolling around and limbs intersecting, so even worse conditions for the biped skeleton detection algorithm. The best I found so far is viggle. I guess they use metas segment anything for human detection, then mask the detected human and apply the skeleton detection algorithm. It works pretty good as long the person can be segmented, which wouldnt be possible for the pinned girl in the video. So I would much prefer something that doesnt mask at all (or only subtract mask interfering elements) and then goes straight to joint detection and tracking, even if the joint rotation is all fucked up. I'm sure there's another algorithm to fix that.