There are two main things for which postprocessing steps could be used in ground/flood masks obtained with Deeplab models:

• Blobs in the sky / Small noisy disconnected components:
  

Ideas for criteria

• Look at surrounding pixels of a connected component - e.g. reject if all the pixels around are segmented as sky
• Look at location of the blob if the image: reject if > x % of the blob if in the top y % part of the image
• Holes in the ground:
  • the close_holes.py script closes all holes that are smaller than a chosen threshold.
    What is a good threshold ? How can we adapt it depending on the input image?
    If it is too big, the postprocessing step might lead to even more erroneous masks.
    Example below: ground segmentation with Deeplabv3 and ground segmentation with the postprocessing step.