I'm working on an image segmentation algorithm with FCN (Long et al., 2015) as the backbone network.
One idea I have is to use the argmax binary mask obtained from the final score layer (250x250x1) to generate some data (e.g. number of blobs in the mask) to modify the ground truth (e.g. set some pixels in the gt mask to 'ignore' labels) or in some way (partly) extract from the features (similar to RPN layer in FasterRCNN).
Does this violate any deep learning or machine learning rules?