YOLOX convergence issue

[juneodie]

I'm trying to upgrade my model from Faster-R-CNN to YOLOX ( with custom dataset A ) because it performs better in my custom dataset B

The problem is that YOLOX model doesn't converge while Faster-R-CNN did

With Faster-R-CNN, mAP goes up to 0.78 but with YOLOX, mAP maintain still on 0.4~0.5

The config I use for YOLOX train is yolox_l_8xb8-300e_coco.py

I only changed the dataset part of the config and use default config for other parts

What approachs should I try to make it converge??

[juneodie]

Oh I think I found the reason. I guess this issue comes from the size of an object. My object is really small ( as small as size of a word ) What configs shall I change to make it fit well in small object detection??