MaiCoach Rate your throw →

MaiCoach

Ordinary video in. Measured biomechanics out.

Reality capture and coaching the athlete can act on.

What we do

MaiCoach turns video of an athlete into a training program built from a measured biomechanical assessment.

Computer vision extracts the movement, a biomechanical model constrains it, and deterministic scoring decides what the athlete works on.

How we do it

  1. Extract physiological landmarks and track them through space over time.

  2. Quantify split-second biomechanical events as kinematic series.

  3. Score measured performance against a skill-based mechanics rubric.

  4. Prescribe a program driven by competency in the individual skills.

Prioritised, actionable, and specific to the athlete who was measured.

Why you can trust it

The same problem as any reality-capture pipeline: raw sensor data is not yet an answer, and the gap between them is where the work lives.

A scaled 3D skeleton, reconstructed from ordinary 2D video.

Two calibrated cameras can triangulate depth where they both see a joint. One camera can’t see depth at all. No one can see a single moment from every angle.

A frame of the throw with the tracked 2D landmarks drawn over the athlete, arm raised with the ball still in hand.
What the camera saw.
The same frame reconstructed in three dimensions and viewed from behind and to one side, an angle the footage does not contain.
The same frame, from an angle no camera had.

Constrained by a biomechanical model, not smoothed into shape.

The model rules out configurations a human body cannot reach, rather than averaging away the evidence. Clean data and correct data are not the same thing.

Machine produced, human reviewed.

Ambiguous frames get adjudicated by a person, and every measurement carries a record of how it was produced.

The same throw, frame for frame, before and after review. Review moved about five joints in every frame — the largest by more than half a metre — and over seventy per cent of that movement was depth, the one thing the camera that shot it could not measure.

Measurement becomes language without the model inventing anything.

Skill priority and drill selection are decided in code. The language model writes the prose and is never handed a number it could get wrong.

Objective assessment of core skills

Hip-to-shoulder separation

In a well-sequenced throw the hips rotate first and the shoulders follow, and the separation between them is where velocity comes from. When they rotate together, the athlete is throwing with their arm alone — the most common power leak in the sport, and one of the most correctable.

How it's measured. Plot hip and shoulder rotational velocity on the same axes: the interval between the two peaks is the measurement, and the measurement is the verdict. A failed verdict selects four specific drills out of eighteen — prescribed by what was measured, not assembled from a template.

The measurement 59.94 FPS · hips lead by 3 frames

Hip and shoulder rotational velocity plotted against time through a throw. The hips peak first and the shoulders peak 50 milliseconds later; the interval between the two peaks is shaded and labelled hip-to-shoulder separation.
Hips lead, shoulders follow. We can assess skill competency by measuring the space between peak velocities.

Disclaimer

Every clip on this page is me.

A research project, built to learn how to do automated measurement honestly. The numbers are internally consistent and repeatable; they are not calibrated against laboratory motion capture, and that distinction is deliberate.

[email protected]