Sky segmentation is a prevalent task within the Lightroom ecosystem, for example as shown in the recent release of Quick Actions. This ecosystem includes mobile, desktop, as well as cloud deployments and we have previously utilized different models for these environments. To simplify this, I have recently developed a joint sky segmentation approach that not only fits into the efficiency constraints of all target environments but also significantly outperforms all of our previous models.

This is a byproduct of my multi-year effort on depth estimation. After all, when looking at a depth estimate, one should immediately be able to identify where the sky region is without even having to look at the corresponding image.