I am using yolox-s architecture for 11 class object detection model which is trained from scratch. My training data is well balanced and the input image size is 160x96. The training data is about 50,000 images. Average object size is 20x10px. I am able to get mAP of 0.88
Two things I have observed are:
- Two of the 11 classes under perform in mAP metrics.
- loss_cls, loss_reg and loss_iou reduces but loss_obj starts increasing after 40 epochs.
How to improve the performance of the two classes. Why is loss_obj increasing after 40 epochs.