CreaseLab
Record a batting shot or bowling action. It classifies the movement, measures the biomechanics and replays your own clip with coaching grounded in the measurements.
01 · Try it
Classification · experimental
–
Dashed line: displayed confidence is capped at 60%.
Measured from the pose
- front knee flexion
- –
- shoulder rotation
- –
- hand speed
- –
- head movement
- –
A pull shot keyframed in 3D for this page, shown from a raised broadcast angle, with scores modelled on a real CreaseLab result. The metrics are computed from the 3D joints as it plays. Try the live app with your own clip.
02 · The problem
Coaching feedback in cricket usually means someone watching you and describing what they saw. CreaseLab measures it instead: shoulder rotation, knee flexion, head movement, release height, all tied to frames of your own clip.
03 · How it works
Scroll the steps. The scene follows.
pixels
framing, background, lighting, camera
pose
joints only · the same for every camera
Athlete visible
Single athlete in frame
Shoulders, hips and knees visible
Full body inside the frame
Clear movement detected
✗ feet cut off · move back and re-record
Nothing stored. Nothing shared. Never used for training.
Illustrated from the project’s README and code. Values shown are examples.
- 01
Pose, not pixels
MediaPipe extracts a pose from each frame. Classifying on pose features normalises away framing, background and camera, which is exactly what broke the broadcast-trained video model it replaced.
- 02
A replay you can trust
Your clip is the centrepiece: pose overlay, 0.25× slow motion, frame stepping and phase markers. Every metric card seeks the replay to the moment it measured.
- 03
Coaching that cites its evidence
Coaching cues are deterministic by default. Each claim must point to a measurement. A capture-quality gate refuses to label poor footage and explains how to re-record.
- 04
Private by design
Clips are processed in a temporary directory and deleted. Nothing is stored, shared or used for training.
04 · Architecture
The whole system, running.
05 · Stated plainly
The models are evaluated leave-one-recording-session-out, because clips from one session are near-duplicates and a random split leaks. That honest split gives 37.5% for batting and 67.4% for bowling, and 54.5% and 86.1% on classes present in training.
The published broadcast video model scored 29.6% on these clips. I traced it to an upstream preprocessing bug and a genuine domain mismatch, and replaced it rather than ship the higher-looking number.
Built with