About the project
Upload a video of someone moving, or press Record and do the move in front of the camera. cermin finds the person and follows their body, fingers and face frame by frame, on this Mac — MediaPipe for the person, then a second, much larger model (DWPose, on the graphics card) that places the legs and feet even under loose trousers, folded back into 3D by bone length. On a dance video that took the legs from found in 88% of frames to 99%.
While tracking it scans the floor beside the feet, so it knows when the camera moves, turns or zooms: the feet stay on the floor, and the body leaves it only for a real jump — one where the hips rise and the time in the air fits gravity. Planted feet are pinned so they do not slide, a take can be trimmed, and the motion goes onto a standard 65-bone rig with fingers and a face, out as .glb, .fbx, .bvh and a face .csv, or straight to gerak inside bengkel to be cleaned up by hand.
Optional characters: Wooden Mannequin by James Wright, Lowpoly Human Skeleton by Void, Casual Man Character by Bogdan Strielecki and Female Character stylized by Guilherme Alves, all CC BY 4.0. The picture is from a public-domain US Army drill video.