Handling foreground-foreground class imbalance in object detection.

[chandan5362]

Hi all,

I am working on an object detection problem where I am using Faster-RCNN network for detection. I have images such that the count of one class label is always less than the other in a single image (say 3:7). It highly affects the model performance on the test data and only a few instances of underrepresented class are being detected. I was wondering that is there a way other than annotating more images that can mitigate this issue?.

I would highly appreciate any suggestions.