Why is semantic segmentation not recommended by Roboflow?

I’m considering using semantic segmentation for a workflow but when I want to create it as a project type, it says ‘not recommended’.

Why is this the case? Does it generally perform worse than regular segmentation models? Does it only work well in specific cases, and if so, which ones?

It’s generally less useful than instance segmentation and you can flatten an instance segmentation prediction into a single mask but can’t convert a flattened mask into objects to get counts or differentiate between instances of the same class.

We’ve found that usually people are selecting semantic segmentation because they don’t know the difference or are confused about what to pick so de-emphasized it in the UI.

1 Like