At the moment we are using pixel-wise loss functions for our segmentation models. These loss functions are great for overall segmentation problems but they fail to capture topology inherent in our satellite imagery.
One example use-case is roads. The pixel-wise loss functions don't penalize for connectivity issues.
We should look into topological-aware loss functions to help with these problems and to directly optimize for example for connectivity instead of having to do a much more difficult task in post-processing.
At the moment we are using pixel-wise loss functions for our segmentation models. These loss functions are great for overall segmentation problems but they fail to capture topology inherent in our satellite imagery.
One example use-case is roads. The pixel-wise loss functions don't penalize for connectivity issues.
We should look into topological-aware loss functions to help with these problems and to directly optimize for example for connectivity instead of having to do a much more difficult task in post-processing.
Idea:
a * pixelwise + b * topology