Google Coral TPU and teachable machine for no-code weed detection

[geezacoleman]

The Google Coral is a low cost TPU that plugs into the USB on the Pi. It would be a simple way to add green-on-green detection without the need to swap the Pi for the Jetson Nano and the complexities that brings.

Combined with Teachable Machine it would be a no-code option to provide people with in-crop image classification. There is a Raspberry Pi guide on how to implement it too.

Thanks Sébastien for the idea!

[geezacoleman]

Alternative to the Google Coral is using something like the OAK (Open AI Kit). The camera includes on-board processing for lightweight object detection algorithms. Process to implement would be:

1. Collect image data
2. Train a model - guide exists to convert YOLOv5 to .blob format
3. Transfer model onto hardware/camera
4. Receive detection locations from the camera a. would need to set up code here for it to work and integrate with existing code
5. activate regions based on detection locations

Very doable - with everything done on camera, would potentially be less code involved than the alternative. Biggest hurdle is training new algorithms and converting to .blob format. The guide above is very long.

[geezacoleman]

Google Coral support is now implemented in #69 .

Will continue to investigate no-code training options with Teachable Machine.

[geezacoleman]

Teachable Machine is quite limited for any object detection use cases. I think Google Coral and other object detection methods are likely the way forward. I will close this now, but feel free to add more comments and reopen if needed.