train-pose

Build a CSI→pose model without overstating it. The project has a retracted 92.9%/100% history — the discipline below exists so it never recurs.

The non-negotiable: mean-pose baseline first

A pose model that always predicts the dataset's mean pose already scores ~50% PCK. Quote PCK only as a delta over that baseline, on a held-out split with no subject or temporal leakage. Example honest result (ADR-181):

Held-out PCK@20 59.5% vs a 50% mean-pose baseline = +9.4 pp real signal — MEASURED.

Paths