Consider AI hardware sticks

[roguedarkjedi]

So while this code is fairly optimized for what it is, there's the very clear and obvious problem that this code base can only process at 1 frame every 2 seconds (essentially running at .5fps). This is with the the full processing power and OpenCV optimizations which don't necessarily target the dnn module (which we use almost exclusively).

This leads into the problem of while we can be as fast as we can be, we'll always be processing dog images very slowly.

I believe I've hit the upperbounds of what the Pi can do. I don't think I can process this any better. So after some investigations, it might be worth considering investing in an AI processor expansion.

[roguedarkjedi]

Coral USB Board

https://coral.ai/products/accelerator/
https://coral.ai/docs/accelerator/get-started/#requirements

Pros:

• Extremely fast. In theory we'll be getting massive speed increases
• Will get us to about 20+fps
• Works out of the box with tensorflow
• Little additional setup needed
• Only $60

Cons:

• Cannot use OpenCV, will need to convert any image formats
• Must use a different AI code flow model as existing code is OpenCV.
• Must recompile the tensorflow models to the lite format (however the one for our current model can be found here)
• Not necessarily built for Raspberry Pi (though this might not be too big of an issue)
• Runs hot, risk of burning
• May need to consider cooling methods later.

Notes:

https://www.tensorflow.org/lite/guide/inference#load_and_run_a_model_in_python
https://medium.com/@aallan/hands-on-with-the-coral-usb-accelerator-a37fcb323553
https://www.pyimagesearch.com/2019/04/22/getting-started-with-google-corals-tpu-usb-accelerator/
https://www.pyimagesearch.com/2019/05/13/object-detection-and-image-classification-with-google-coral-usb-accelerator/

[roguedarkjedi]

Intel Compute Stick

https://movidius.github.io/ncsdk/ncs.html

Pros:

• Can use existing code already
• Slight speed upgrades (will get us to ~8fps)
• No major code changes
• Works with OpenCV

Cons:

• Setup is rather complicated.
• Workflow documentation has changed significantly due to an Intel refactor changing to an "OpenVIVO" format
• Rather bulky size

Notes:

https://towardsdatascience.com/how-to-stream-video-with-real-time-object-detection-on-raspberry-pi-f6503c46c7f9

https://www.pyimagesearch.com/2019/04/08/openvino-opencv-and-movidius-ncs-on-the-raspberry-pi/

https://www.pyimagesearch.com/2018/02/12/getting-started-with-the-intel-movidius-neural-compute-stick/

https://medium.com/hacksters-blog/getting-started-with-the-intel-neural-compute-stick-2-and-the-raspberry-pi-6904ccfe963

[roguedarkjedi]

This is implemented and works great. Closing this issue, but will pin it because the notes are good.