The Avatar Talk capstone: body, face, fingers and wrists driving one VRM from a single webcam, all in the browser. Why an arm solve can’t roll the wrist, the thumb bending about the wrong axis, caching frames and counting drops, the neutral face where 0.3 is the new zero, the mirror rule — and an honest ~9 fps, not the 30 I’d claimed.
The Avatar Talk finale: pose, face and hands driving one VRM in a single browser render loop. The One Euro filter that beats the jitter-vs-lag trade, filtering landmarks at the source, MediaPipe’s 52 ARKit blendshapes mapped to VRM expressions, and finger bend-not-splay hand tracking.
Same recipe — detect a pose, track it, classify it — but very different evidence. One 2026 paper grades a physical-therapy exercise against expert ground truth (RTMPose + GRU, ~96%); the other profiles a basketball move with no ground-truth labels, and honestly says so. What validates a pose pipeline isn’t the detector — it’s the ground truth.
You captured the motion — you’re not done. Two July 2026 papers trace the path from validated kinematics to a modeled injury number: an OpenCap markerless scoping review (51 studies), and a badminton-lunge ACL-load model (OpenSim + XGBoost, R² ≈ 0.91). Each inferential layer adds its own uncertainty.
Part 3 of Avatar Talk: retarget MediaPipe pose landmarks onto a VRM avatar’s bone rotations, in the browser. Positions vs rotations, the MediaPipe→VRM coordinate flip and mirror, rest→target quaternions, world→local chains, frame-rate-independent easing, and the single-camera depth fix.
Part 2 of Avatar Talk: load a real 3D VRM avatar in the browser with three.js and @pixiv/three-vrm. Why VRM’s standardized bone map matters, the loading cleanup that counts on a 15 MB model, two-line breathing, and the expression system for lip sync.
Real-time motion capture in the browser — no install, no server, about 200 lines of TypeScript. Part 1 of Avatar Talk: a webcam, a transparent canvas, and MediaPipe’s PoseLandmarker drawing a live upper-body skeleton, with the two gotchas that make it work.
Two 2026 papers take opposite bets: SoccerNet pools 427 teams and 1,129 submissions onto one shared benchmark, while a skateboarding study checks one commercial IMU against lab references — with mixed results.
Two reviews and one study show the distance between what AI-plus-wearables promises for injury prediction and what exists: a good classifier of expert judgment (0.86 AUC) — not a validated predictor of a future injury.
Cameras got cheap. Three 2026 studies — a 3D reconstruction pipeline, a systematic review, and one smartphone — all point the same way: global, gross measures are within reach; fine timing and joint angles are not, at any price point.