BarPath · velocity-based training

Measure first.
Then show it.

A side-on video of a lift goes in. Bar speed per rep comes out. Right now this page shows what we have measured honestly — not a chart we cannot yet earn.

"This demo is honesty-first: every figure below is either measured by our own pipeline or labelled as unverified."

That is the whole product promise. Everything else is interface.

The claim, in one line

Bar speed is the honest measure of fatigue —
it drops before your form does.

The pipeline, measured not asserted

88.1%

of video frames produce a tracked body

392 of 445 sampled frames

complete reps detected by our own tracker

the app's own counter read 10 — one rep was not visible

3.33s

mean rep period, ±0.16s

plate tracked in 884 of 1,335 frames

—

bar velocity, m/s

not shown until a trustworthy scale exists

The finding that changed the plan

In a side-on view one hand is always behind the body. So the obvious way to track the bar — average both wrists — fails almost completely.

confidence rule	near wrist	far wrist	either	both
0.1	100.0%	50.8%	100.0%	50.8%
0.3	100.0%	5.1%	100.0%	5.1%
0.5	95.9%	2.8%	95.9%	2.8%
0.7	11.2%	0.0%	11.2%	0.0%

At our default threshold of 0.3, both wrists are usable on 5.1% of frames — but either one alone is usable on 100.0%. So we select the wrist by confidence, never by name, and never average.

Note the last row: even the clearly visible wrist barely clears 0.7. Confidence has a ceiling near 0.79 — so "raise the threshold to be safe" would destroy coverage.

Wrist visibility over time: one wrist stays high, the other is near zero for almost the whole clip

What we are deliberately not copying

The source recording reports Mean V = 0.43 m/s.

We do not reproduce this number. It is shown only because it is the one figure legible in the source recording, and so you can see what we are deliberately not copying.

0.43 m/s sustained over a ~2.6 s concentric implies roughly 1.1 m of bar travel — implausible for the lift shown. Either the app corrects for perspective in 3-D, or the figure is modelled rather than measured. We cannot see which, so we report our own measurement instead of matching theirs.

What this page does not do yet

Upload a video — waits for the pipeline to produce real numbers
Per-rep velocity chart — waits for a trustworthy scale reference
Skeleton overlay on your own footage
Live camera

Each of those is deliberately absent rather than faked. A chart drawn from invented numbers would look better than this page and mean less.

Where the numbers came from

Pose rate and wrist table — occlusion_check.py run over the source recording on 4 Nov 2026
Rep count and period — plate tracked by disc detection plus a circle fit on the plate rim
Source clip — a 44.6 s screen recording of an existing app, sent by the project owner. Not our own footage, which is why the one number that matters most (m/s) is missing.