Image processing time of FOMO in Edge Impulse

Image processing time is slow for me.
When I use FOMO last Octover, I could get 2 datas per 1 second, I probably.
But now I can get 1 data per 1.5 to 2 seconds. And when I show Serial monitor in ArduinoIDE, I can see 1 data per 10 seconds.
I want to use fomo by mobile robot so it is too slow for me.

Could you please share the way to get data more faster.

Hello @FT4444,

Which hardware runs your model?
To improve your inference speed, you can:

  • Use greyscale images if it is acceptable for your image
  • Lower your input size (96x96 images is usually pretty fast even on cortex-m7 hardware)
  • Use a smaller backbone (currently the FOMO MobileNetv2 0.1 is the smallest backbone we provide).

Keep in mind that you might lose some accuracy, consider training longer and adding more training data.



Thank you for replying fast. That would be very helpful.

I’m using ESP32S3-wroom(FREENOVE) and arduinoIDE.
Setting is greyscale and 96x96 images.
This is my project : Logo_kaiten - Dashboard - Edge Impulse

I use same machine learning result by FOMO but it has more time than before.
(classification time : Before:about 150ms, Now:about 11000ms)