Open 51N84D opened 4 years ago
I think you have interesting results regarding this issue @51N84D ; can you write up a sentence or 2 to close this issue? (+pictures)
Actually I haven't tried the patch idea yet. What did produce interesting results was using perceptual loss on the reconstructed region, but I haven't yet found a counterpart for the flooded region. Still an open problem so maybe we can leave this one open a while longer
Perceptual loss works out of the box with images that share a common mask (the edge artifacts cancel out) but not with images that have differing masks.
One idea: Sample a square crop from the respective masked regions and calculate a perceptual loss between these two. Not sure if this is a differentiable operation though