convert a detect model to int8, the performance drops a lot

I use the yolov3 training a model to detect the traffic light, as we can see ,the detection target is very small relative to the whole images and there are few object from 1 to 3 in every images.
when i convert the model to float32, it performs well, however, the performance drops a lot with int8.
I also use another model to test and find that when with small and few target in images, the int8 model performs not well but performs well with big and not a few target.

anyone face the same problem? and what’s the reason, how to resolve it


To help us debug, can you provide more details about your dataset? How many classes and what percentage of images you use for calibration?

Thank you!