wanted to record youtube videos without a face: cut the background out of a webcam feed and maybe drive a character with my movement. research made it clear these are two different problems that get lumped together. matting is per-pixel is-this-a-person and gives you a clean cutout, while motion capture extracts joints and drives a rig. they share a camera and nothing else.
for the cutout, the honest toolchain is boring and effective. the obs background removal plugin runs mediapipe selfie segmentation live and gives a transparent person without a green screen. robust video matting is the quality pick because it uses temporal information, so the mask flickers less around hair and hands. modnet is the real-time portrait option. the gotcha is the output path: obs virtual camera doesn't reliably carry alpha to every app, and a normal mp4 has no transparency at all, so record the composite or record the mask separately and let the editor use it as a matte.
for the character version, the pipeline is webcam to landmarks to bone rotations to rig. mediapipe pose gives thirty three body points at real time, and posecap ties a full smpl-x estimate into blender to drive a rigged character and record keyframes. freemocap is the multi-camera option if i ever want proper 3d reconstruction. rust was my original fantasy for all this and the verdict is blunt: rust is fine for the capture bridge and useless for making the model faster, and obs, blender, and mediapipe already solved the integration work. start with the obs plugin for the cutout and posecap for the character, and only reach for custom code when something specific is missing.