Intelligent interactive segmentation labelling through deep learning models

[juan-park]

Is your feature request related to a problem? Please describe.
Semantic segmentation labelling takes a long time to individually paint all the pixels of all the different classes using a simple brush painter. If we could incorporate interactive intelligent labelling functionality, that would be awesome.

Describe the solution you'd like
Either a implemented interactive segmentation option, or the availability of a pipeline for users to be able to implement their own.

This would involve, for each class, an option to "paint" a portion of the scene as a positive or negative label wrt a single class.
Once selected, the backend would take the image and the user input, generate pixel predictions and return back the mask result.

[makseq]

The support for this feature is realized in LS 1.3 and the latest ML backend.