CSI → pose (skeleton) overlaid on your laptop camera
camera: off
Camera is a visual reference only — it is NOT fed to the model. Overlay alignment is approximate (model trained in a different camera's frame).
live
CSI source—
nodes—
presence—
motion—
pose fps—
This skeleton is inferred from WiFi CSI only — no camera in the loop here. A model was
trained on paired (camera-pose, CSI) data in this room (ADR-079/180).
Honest accuracy: ~59.5% PCK@0.10 on held-out data (vs a 50% mean-pose baseline → +9.4 pp real signal). It captures coarse pose; fine detail is weak (PCK@0.05 ≈ 24%). Same person / room / session — not validated cross-day or through-wall.
Honest accuracy: ~59.5% PCK@0.10 on held-out data (vs a 50% mean-pose baseline → +9.4 pp real signal). It captures coarse pose; fine detail is weak (PCK@0.05 ≈ 24%). Same person / room / session — not validated cross-day or through-wall.