Day 59: Pixel measurements from pose: the math
Assembling a measurement vector
Combine yesterday's distances into the set of measurements a size recommendation needs — shoulder width, chest/torso width, arm length, torso height, hip width — each in pixels for now, each carrying a confidence derived from the visibility of the keypoints it used. This measurement vector is the pipeline's core numeric output, still in pixels until calibration (Days 77–79) converts it to centimeters.
def measurement(a_lm, b_lm, w, h):
a, b = point(a_lm, w, h), point(b_lm, w, h)
conf = min(a_lm.visibility, b_lm.visibility) # weakest link
return {"px": dist(a, b), "confidence": round(conf, 3)}
measurements = {
"shoulder_width": measurement(L.LEFT_SHOULDER, L.RIGHT_SHOULDER, w, h),
"torso_height": measurement(L.LEFT_SHOULDER, L.LEFT_HIP, w, h),
"arm_length": {"px": arm_length_px, "confidence": round(min(vis_ls, vis_le, vis_lw), 3)},
"hip_width": measurement(L.LEFT_HIP, L.RIGHT_HIP, w, h),
}Confidence propagates through the pipeline
A measurement is only as trustworthy as its least-visible keypoint, so confidence takes the minimum. These per-measurement confidences will combine into the single '91%' the engine reports (Day 81). Threading uncertainty through every step — rather than bolting on a fake confidence at the end — is what makes the final number honest.
Key terms
- Measurement vector
- The set of body measurements (shoulder, torso, arm, hip...) the size recommendation consumes.
- Confidence propagation
- Carrying and combining per-step uncertainty through the pipeline so the final output reflects true reliability.
Why does a derived measurement take the minimum visibility of its contributing keypoints as its confidence, rather than the average?
