vtuber | 从一开始的 Vtuber | Machine Learning library

after spending some time learning basic computer vision concepts and techniques I started to notice how unreliable simple scripts can get when the luminosity or scale changes and how resource consuming is to use more advanced solutions like creating a well-made HAAR cascade or HOG-feature based svm. Furthermore, some even more advanced methods involving machine learning usually take a lot of time and GPUhours when a high quality model is created.

Recently while looking through YouTube I've found a lot of so called VTubers who use various software to control virtual avatars with somewhat precise motion tracking and what seems to be no errors whatsoever. While not something unimaginable, the amount of people using the software and the amount of software itself seems to be rather large.

Planning to investigate even further I looked into different ways similar technology works, but so far I only found a complex solutions involving either AI driven models or assistance from some sort of positional sensors attached to the body of the user. Still its hard to believe all of those people go through such measures, so I realised that perhaps this is accomplishable with some cv solution which is relatively easy on resource consumption. So far I looked into different ways to "map" model joints to human ones. On my own I tried basic counter matching, and greenscreen filtering to avoid errors. while I successfully managed to remove almost all errors, there still were moments when mapping snapped arm for example to elbow and etc.

How exactly is object recognition and motion tracking of such quality is achieved using only computer vision?