Overfitting Under Class Imbalance

I am working with semantic segmentation on dataset with 5 classes including a null class including unlabeled pixels due to uncertainty, I am ignoring the null class in loss function computation, one of the classes consider is under represented , I have tried weighted loss, over and under sampling , augmentation but the model is overfitting the minority class with IoU of 93% and on test it is 16%, I tried a lot of loss functions, and a paper considered asymmetric techniques with mixup and adversarial training but nothing seems to work. Do you have any suggestion to deal with this problem. I am using mIoU for comparing results as the main metric