How to handle with imbalance dataset

[hllj]

First, thank you for your work on this model. I have some problems that need your advice.

I have already trained your model with many types of data, I have the same result when some classes have less bounding boxes than others.

Do you have any solution to handle this kind of data (not really imbalance dataset, but some classes have less bounding boxes).

Another problem that I figure out when training, with small dataset (maybe 100-200 images), my experiment on resnet34 is much better than reset50 or resnet101, can you spot this problem ? Do you have any advice on how much data samples that we need to train models ?

Thank you.