Baseline reconstruction + 16 Aug 2026 progression update

Muscle-up
digital twin.

The original three-attempt baseline remains reconstructed frame by frame below. New footage now shows three full turnovers to stable support; the current bottleneck is repeatability across later attempts.

Open the new 7-attempt analysisAnalyze your own clip locally
60 Hz primary pose33 tracked landmarks10 Hz scene tracking±9.6 px bar uncertainty
NEW / 16 AUG 2026

Three first-attempt successes. Seven attempts, three clips, one repeatability pattern.

Inspect original and annotated video, synchronized telemetry, every reviewed attempt window, uncertainty, coaching priorities and the next-session capture protocol.

Open full practice update
Attempts detected32.28–24.68 s
Best attempt#1relative within video
Chest deficit11 cm9–14 cm sensitivity envelope
Elbow timing+0.32 safter peak COM, #1
Model agreement13.6 pxmedian common-joint gap
01 / Synchronized evidence

Playback lab

Scrub once; video, telemetry, 2D overlay, plots, and 3D twin stay synchronized.

Source video + live reconstruction
00:00.00
Between attempts
download annotated MP4
Interactive 3D twin
drag to rotate · wheel to zoom
Tracked barloading…independent per-frame 3D lineAthlete ↔ bar depthloading…camera-frame, estimated metresGrip ↔ barloading…3D finger distance during contact
COM / bar−1.22 m
Vertical speed+0.04 m/s
Elbow angle161°
Power proxy37 W
Frame-level telemetryClick a trace or drag to zoom. Dashed lines mark attempt windows.
02 / Performance diagnosis

What stopped the transition

Each conclusion is tied to a timestamped metric and bounded by what one oblique camera can support.

Four-panel comparison of chest height, velocity, elbow asymmetry and power proxy
Attempt-level comparison. Power is a COM-derived proxy.
Timing of upward velocity, knee tuck, COM peak and elbow flexion
All attempts complete elbow flexion after peak body height.
Center-of-mass trajectories for three attempts
Camera-stabilized, tracked-bar-relative COM paths; color runs start to finish.
Sensitivity ranges for key metrics
Sensitivity to smoothing, ±7% image scale, and tracked-bar centerline uncertainty.
Camera translation, rotation, zoom, bar height and bar slope diagnostics
Measured handheld-camera motion and active grip-bar trajectory across the full recording.
Chronological source frames with the tracked sloped grip bar overlaid
Visual audit: the tracked centerline follows the perspective-sloped bar actually gripped in every shown frame.
03 / Scene depth

Relative depth + segmentation

Depth Anything V2 estimates per-frame relative inverse depth. Color is normalized independently for each frame and is not metres.

Selected source frame
Source
Selected person segmentation
Person segmentation
Selected relative inverse depth
Relative inverse depth closer → warmer

Do not compare depth colors across timestamps. Each view is independently normalized; it supports scene layering and a visualization-only point cloud, not metric travel.

04 / Frame evidence

Key movement events

Knee tuck, maximum COM height, and minimum elbow angle for each trial.

Nine annotated frames showing three key events for each attempt
tracked active grip bar estimated COM pose model
05 / Scientific honesty

Measured, modeled, proxied

The twin is useful because its boundaries are visible—not hidden.

Observed

Direct video evidence

  • Frames and timestamps
  • Static-background features
  • Paired grip-bar edges
  • 2D landmark locations / outcome
Estimated

Calibrated + model-derived

  • Camera transform + bar interpolation
  • Image-plane metres from height
  • 33-joint pose + YOLO cross-check
  • Relative scene depth / masks
Proxy

Mechanics approximations

  • Segment-weighted COM
  • Velocity and acceleration
  • Vertical force / energy / power
  • Within-video ranking
Not identifiable

Requires more sensors

  • Ground-truth scene depth / joint torque
  • Individual muscle activation
  • Tendon load or injury diagnosis
  • Force-plate-grade power
Pipeline, calibration, and limitations expand
06 / Open evidence

Download everything

Original identifiable footage is public with the athlete’s explicit approval. Derived data preserve provenance and proxy labels.

Original source4K HEVC videodirect evidence · 471 MB Annotated playbackPose + telemetry MP430 fps H.264 · 33 MB Analysis packageCSV · JSON · Parquet · reportreproducible data bundle Readable reportMethods + coachingMarkdown · evidence-linked Correction auditCamera + grip-bar evidenceold vs corrected metrics · diagnostics