Object Detection Models on H7

I read in a post in May '21 that you were working on object detection for the H7. How is that coming along? Will it be possible to store large models in flash? Alternatively, can the person detector model in flash be replaced with a custom object detection model? Thanks.

1 Like

object detection for the H7. How is that coming along?

It’s done and it will be available when EdgeImpulse enable this feature.

can the person detector model in flash be replaced with a custom object detection model? Thanks

Yes, see openmv/README.md at master · openmv/openmv · GitHub

And it will be possible to train a model with EdgeImpulse and generate a firmware image with that model builtin.

1 Like

Ahh, that’s great. Where can I find the amount free flash available for models?

When you build the firmware it prints the text size used, the free is total flash size - text size. That’s something that changes between boards and releases.

With the latest firmware on the H7, and removing ‘person_detection’, about how much flash is available for a tensorflow lite model? This will guide the target size for a custom object detection model. Thanks.

Also, what is the size of the ‘person_detection’ model in flash, to give a sense of the speed to expect with a custom object detection model of similar size?

The person detection model is about 300KBs, but there’s more free flash, I can’t remember the exact number, I’d say about 400KBs.