Jump phases. From pose signal to measurable skating events.

A pose-based computer-vision system that detects take-off, apex and landing, measures biomechanics and analyses what happens after contact. Phase detection, landing outcome and validation are deliberately separated.

v3 RAW-TRANSITION LANDING OUTCOME v1 COCO-17 VALIDATION NOT A JUDGING SCORE

External samples

318 Unseen single-jump skeleton sequences processed.

Pipeline completion

100% 318 / 318 completed without detector exceptions.

Raw structural sanity

83.0% 264 / 318 across the mixed public dataset.

Compatible cohort

94.8% 255 / 269 video-pose-compatible sequences.

Architecture

One jump. Several independent decisions.

The system does not reduce a jump to one opaque score. It produces measurable intermediate state and keeps heuristic interpretation separate.

01Pose MediaPipe for video. COCO-17 skeletons for the external benchmark.

02Features Knee angle, torso inclination, hip/ankle geometry and raw transitions.

03Phases Take-off → apex → landing using the v3 raw-transition detector.

04Outcome Post-contact window detects recovery instability or collapse signal.

05Report Measured facts, timing diagnostics, evidence strength and limitations.

Failure analysis

54 failures became three different problems.

Suspected domain shift / converted mocap

Fixed 150-frame sequences with effectively absent confidence channel. These samples are not used to retune detector weights.

Video-pose edge cases

Apex occurs near the clip boundary. Raw trajectories must be reviewed before declaring detector failure.

Validation-threshold case

A short clip violates the 60% flight-ratio sanity rule. This may be a validator issue rather than a detector error.

Current boundaries

Measured facts ≠ interpretation.

Measured Frames, timestamps, airtime, knee angle, torso angle, pose quality and post-landing signal.

Heuristic Phase confidence and landing evidence are signal-strength measures, not calibrated probabilities.

Not yet classified Axel, Lutz, Flip, Loop, Salchow and Toe Loop require a separately validated classifier.

2D limitation Camera movement, viewpoint, occlusion and motion blur can affect reconstructed biomechanics.

Next validation wave Manual ground truth on unseen videos → phase MAE → ±1 / ±2 / ±3 frame accuracy → landing-outcome confusion matrix → failure stratification.