Experiments: sprint auto-start and auto-splits

done2026-10-10 (overnight)Sprint direction

Question: can pose detection find the start and the split crossings from a single down-the-line phone view, so a person only has to confirm? Everything is App-Measured; tape owns the metre; no auto-DQ. All auto events are unconfirmed suggestions (humanReviewed=false). No app code was changed.

New (2026-10-11)

Luke's own video, simulated: annotated clips, keyframes and proxy tables (assumed scale, ±25-30%, illustrative only).
Iterations v0.2.2 to v0.2.6: change, evidence, verified vs compile-checked for each build.

Luke's video (annotated, proxy)

Run 1 (proxy metres, illustrative). Burned-in overlay: skeleton box, START, running time, proxy split labels, proxy distance from camera (assumed scale). Direct link.

Run 2 (proxy metres, illustrative). Burned-in overlay: skeleton box, START, running time, proxy split labels, proxy distance from camera (assumed scale). Direct link.

Keyframes, tables and caveats →

Verdict

Start (first movement): viable as a proposal. Found 7 of 7; median absolute error about 67 ms (median signed +40 ms); 6 of 7 within ±67 ms; worst −167 ms.
Auto split crossings from one down-the-line phone view: not viable as an unreviewed measurement. Found 17 of 18 gates, but timing noise is about 130 ms median and the metre scale cannot come from the video alone (±20% scale means about ±150-200 ms). True gate accuracy is unknown: no markers in the footage.
Recommended next: frame-1 calibration (tap cone points), then a marked-course trial; show detections only as unconfirmed suggestions. See the proposed next node.

Annotated videos

Run 1 (best): skeleton, START marker, running time, split flags, proxy-distance bar. Proxy metres are illustrative only. Direct link. Source footage: YouTube wKP0iWaAlZI, CC BY, uploader ლევანი რობაქიძე.

Run 0. Direct link. Same source and licence.

Images

Shoulder-width curves with auto vs manual start and gates for all six reps — Shoulder-width scale curve, auto vs manual start, and gates, all 6 reps.

Manual reference frames for run 0 start — Manual reference frames for the run 0 start (reference itself is ±1-2 frames).

Manual reference frames, block start clip — Block-start clip (YouTube 3hREO5HBdw0, CC BY, informedinsport.com) manual reference frames.

Gate bar check, run 1 — Human check of shoulder-width bar: only reliable to about 2-3 frames (±70-100 ms).

Start detection vs manual frames

run	auto frame	manual	error
0	356	354	+2 f (+67 ms)
1	1384	1383	+1 (+33 ms)
2	2844	2845	−1 (−33 ms)
3	4225	4230	−5 (−167 ms; slow crouch release)
4	4988	4986	+2 (+67 ms)
5	5760	5758	+2 (+67 ms)
3hRE (25 fps, side)	193	192	+1 (+40 ms)

Pose coverage vs runner size

runner bbox height	frames	full-frame pose	x3 crop pass
60-100 px (far)	585	0%	100%
100-150 px	197	0%	100%
150-250 px	802	14%	100%

Reference is the crop pass itself, so its 100% is optimistic.

Frame-rate sensitivity

rate	with interpolation (median / mean abs / p90 / max)	no interpolation (median / mean abs)
15 fps	−4 / 24 / 49 / 118 ms	+28 / 39 ms
10 fps	−4 / 43 / 107 / 248 ms	+64 / 83 ms

Take-away: 30 fps minimum with interpolation; 60 fps recommended.

Limits (honest)

Six reps, one athlete, one camera; manual references are mine to ±1-2 frames; scale assumptions unverified; CPU model, not a phone model; real-world will be worse.

Full report
Luke's video simulation
Iterations v0.2.2 to v0.2.6
Simulated app output tables (proxy metres, illustrative)
Footage sources and licences (links only)
results.json