Hi! You might have seen the preview in @dansitu’s talk but we’ve now released our computer vision pipeline in Edge Impulse, with support for the OpenMV Cam H7 Plus out of the box, but can be compiled on other Cortex-M7 and equivalent targets. Thanks to transfer learning and data augmentation we get to models that work surprisingly well with very little data (around a 100 images for the 3-class classifier that we train in the tutorial) on very small targets (resulting tflite file of about 250k on grayscale model).
Getting started tutorial here: https://docs.edgeimpulse.com/docs/image-classification (also works with your phone as a sensor!) and hope this will finally let some community members train tiny computer vision on their own data (I know for me this was the first time)!