Thank you. Managed to do this with the help of Claude Sonnet 4.6 which showed me how to retrain the data using something which has the same landmarks as in your Mediapipe pose tracking examples.
This seems to work very well, thank you.
Not sure how to share the process?
I took maybe 25 photos of each of 14 poses (yes, different handstand shapes, of course!)
Then organised them into this folder structure
training_data/
arms_crossed/
photo1.jpg
photo2.jpg
...
arms_raised/
photo1.jpg
...
lunge_left/
...
Here's my Colab notebook:
https://colab.research.google....
After I'd successfully trained that and got the files out the other end...
It then gave me the python mediapipe code.
https://www.dropbox.com/scl/fi...
And that outputs the OSC into Isadora. Two streams, one telling me which pose it is, and the other giving me the confidence about it.
Hope this is useful to someone!!!
Cheers
Mark (not that Mark, who I'm sure doesn't use Claude to do his coding!)