wifi-densepose/harness/ruview/skills/train-pose.md

34 lines
1.5 KiB
Markdown
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

---
name: train-pose
description: Train/evaluate WiFi pose models honestly — camera-supervised (MediaPipe + CSI) and camera-free (WiFlow), always checked against the mean-pose baseline before any PCK is quoted.
---
# 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
- **camera-supervised** (ADR-079) — MediaPipe Pose labels the camera frame; paired CSI
trains the net. Train/infer in one camera frame so the skeleton aligns.
- **camera-free** (WiFlow, ADR-152) — no camera at inference; geometry-conditioned.
- **in-browser** (ADR-181) — WebGPU/WASM trainer; the active backend is shown as a badge
(honest about what's executing).
## Before you publish a number
1. Run the mean-pose baseline on the same split.
2. Report `(model baseline)` in pp, with the split definition (chronological /
blocked-gap / grouped-bucket; no leakage).
3. `ruview_claim_check` the writeup — it flags any untagged or 100%/perfect claim.
4. If it's a benchmark vs SOTA, tag MEASURED-EQUIVALENT only with the reproducer.