I have a heavily imbalanced dataset, GFL learns it really well, but I'm now in need of segmentation output and cascade mask rcnn seems to be struggling.
I'm a little limited in my choice because I want to be able to export to ONNX, so a change in loss would be ideal, but several other issues seemed to have had little success with this approach
This is a more general question wrt. what are my options to improve performance on an imbalanced dataset for a mask_rcnn type model. Not specific to loss functions which I've seen discussed elsewhere.
I have a heavily imbalanced dataset, GFL learns it really well, but I'm now in need of segmentation output and cascade mask rcnn seems to be struggling.
I'm a little limited in my choice because I want to be able to export to ONNX, so a change in loss would be ideal, but several other issues seemed to have had little success with this approach