Recorded running, frame by frame

Colin running past a phone camera, March 2026. Camera pixels are recorded; the saved 26-point pose stream contains model estimates. Wearable timing is not linked in this example.

Saved pose estimates, paired by recorded frame ordinals. Point errors may be visible; this is not a joint-accuracy test. Anatomical sides are not independently validated. Wearable timing is not linked.

Gold dots: model estimates. Cyan ring: selected point. No pose interpolation or fitted timing offset is applied. Frame jumps seek inside each camera PTS interval. Open camera video.

Loading the bounded pose record. Native camera playback remains available.

Video clock
0.000000 s
Decoded camera frame PTS
—
Original pose timestamp
—
Selected estimate x / y
—
Raw model score
—

Vertical image position

Selected point vertical image position 238 saved pixel values against actual camera presentation timestamps. The cyan cursor follows the decoded video frame. Image y increases downward.
Dots are the 238 saved values; the cyan cursor follows the decoded frame. Image y is in pixels from the top edge and increases downward. Perspective and camera motion affect these positions; this is not calibrated biomechanics.

Pixel origin: top-left; x right, y down. Raw scores are uncalibrated and can exceed 1. They are not probabilities. Point indices retain model order, without claiming anatomical identity.

238 consecutive saved ordinals, no pose-record holes. That does not establish absence of original capture drops. No linked accelerometer, gyroscope or heart-rate stream is supplied.