How to address class imbalance in computer vision datasets?