• a postprocessing trick for semantic segmentation
  • After your model makes a prediction. You can get a better DiceLoss by going through every predicted segment, and put each pixel into a disjoint set.
    • if there are sets with a small pixel count, it means that there are clusters of pixels that have fooled the model and aren’t really part of the target in the image.
    • so just remove those small clusters
  • Note: be careful when adjusting the “minimum size of pixel clusters to keep”. Be conservative. You don’t want to remove a bunch of important pixels