Occlusion labeling semantic segmentation

I’m currently working on labelling data for semantic segmentation and I have a question about handling occlusions. How do you typically approach occlusion labelling in this context?

Do I label the visible parts of the objects multiple times, or do I label the object along with the hindering object and set the mask layer of the hindering object in front of the object that I want to label?

If I am supposed to label the object along the hindering object, how should I label it? Should I try avoiding it as much as possible or should I make a rough mask of how my object behind would be shaped like?

I’m asking this because roboflow is only giving me one mask per instance. If I create another mask it takes it as another instance.

In the following case. 1st picture shows how the player in the blue jersey is behind the red jersey player. So how should I label this? Should I label it like 2nd picture including the complete leg along with the red jersey player’s leg? Or should I label it like the 3rd picture with just a few overlaps?

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.