Guys, I'm using TensorFlow and doing semantic segmentation and can't find a way to segment small and tiny parts as well. I'm thinking about modifying the loss function or adding some refinement block on top of my model. What suggestions do you have?
Here is what I mean.
Adding refinement block could be a solution, but if you want to modify the loss, maybe you need to find those areas (only small parts that need to be segmented) and put them in the loss function with a weighted number.
Or you can normalize your loss (e.g. cross-entropy) with the ground truth area and increase the weight of the loss, which contains only small numbers because if you have small numbers after normalization, it means you have small errors and in general your small errors contain the regions you are looking for.