Thanks for the comment. I did not use data augmentation technique because there was no need. I had enough data even for minority classes. The problem was that you cannot just add minority classes because other objects were there too. You want to add more 'bicycle' objects but 'cars' and 'person' classes are there always. Arghh! It helped a little bit for sure, but not as much as I hoped. But this is expected for all realistic datasets and your model should perform well even when the incoming data is imbalanced. Without using synthetic data, the best bet is to make a large "re-balanced" dataset for training. A "re-balanced" dataset is where you under sample majority classes and oversample minority classes if you take the data-centric approach (or re-weight if you take the model-centric approach). I'd happy to learn other ways. Thanks,