I have a question is there a reason the documentation for Axon inferencing on a rPi calls for resolution of 300x300 or greater? Though we have some hardware on the way to speed thing up our frame rate is less than 2 fps at the suggested 640x480 I was thinking of going to 320x240 but that is less than the 300x300 Though when first trying it out we did see detections doing even smaller resolutions.
While I did not write the documentation, I believe this is due to how it ML Vision works.
The AI model relies on feature extraction. With less resolution, you have less features / more unclear features. Combine that with convolution, as is common in many of these cases, and it won’t be able to properly identify the features.
So, you could use a lower resolution, but it will give you very bad results.
If you are worried about fps, look into getting a TPU module for your pi.
Thanks for the answer. We did get a tpu and that took us to 14-15 fps at 630x480. We do have an upgrade from a pi 3 to 4 on the way as well.