We've been looking at it. Have a model to generate training data.
There are some applications we could see it doing well in. I'm not convinced that sagittal plane analysis is one of them as there are so many variables and accurate data collection would still require user action, which is a huge source of error.

Say it quietly- the owner of this web site may realise he's going to be replaced by a robot 🤔

Are you trying to fit people to the equipment they have? Or giving people the fit coordinates to buy new equipment? From a machine learning perspective these are very different scenarios with very different solutions.

I can see the value of a quick algorithm that looks at knee angles and hip angles and tells you about saddle height, setback and crank length. Its pretty niche though and you would need to collect thousands of good and bad examples to teach a computer.

Getting a good parameterized training set is the biggest obstacle. It's not just static measurements (ex. femur length); but also issues like flexibility, pronation, etc. Even with the 1000s of folks who have been professionally fit using Retul, FIST, etc.; I doubt there is a comprehensive database of all the needed parameters. A lot of high quality 360 3D imaging of fits might be the best way to create scaleable sets.

A ML system needs to be really advanced to outperform even a cursory "eyeball" fitting, or just having someone use the "Lemond" method, KOPS, or other rules of thumb.

---

ECMGN Therapy Silicon Valley:
Depression, Neurocognitive problems, Dementias (Testing and Evaluation), Trauma and PTSD, Traumatic Brain Injury (TBI)

Where is the ML in what you're talking about?

Are you talking about using ML to estimate morphological measurements given conventional 2D imagery? And then just using those measurements to feed into an estimated starting position for the key fit parameters?

If that's the case, people are already doing that.

Or are you talking about taking a set of morphological measurements, and then instead of calculating the starting fit position from straightforward geometry calculations, and instead having ML act as a black box where it spits out estimated key fit parameters generated from a training database of "known good fits." (or something)?

Or both?

I could be wrong. But I tend to think good old algebra/geometry is pretty darn good at the second problem. Have a set of equations that describe the major skeletal lengths and angles. And another set of equations describing the bike contact points. Constrain those parameters by the type of fit. Fill in knowns. Solve for the unknowns.

Edit: But I'd love to be proven wrong, and an ML fit system to somehow work better.

I think a neat ML system would be to have real-time aero and power analysis on a fully motorized fit bike. Use cameras to estimate CdA in real time. Use a power meter and physiological (O2, HR) measurements in real time. Then have the fit bike automatically adjust the fit bike positions performing a type of gradient descent to arrive at the fit that solves the joint optimization problem of power efficiency vs. CdA.