Reducing the inferencing time

Question/Issue:
Hi, I trained FOMO MobileNETv2.0.1 and uploaded it to ESP-32. I used RGB 48x48 pixel images and my object detection takes approximately 1 second to get some results. I want to lower this to say 250ms to be able to get the results close to “real” time.

I have two questions:

  1. Can I run the object detection on my computer instead of ESP-32 itself? ESP-32 will just take pictures, send images through WIFI to my PC, which will then process the data a lot faster than ESP-32 itself.
  2. How can I tweak the model to make if work faster on ESP-32 itself?

Project ID:
531463

Environment:

  • Platform: ESP 32
  • Build Environment Details: Arduino IDE
  • OS Version: Windows 11