Training detection model with optional segmentation available?

[hl-philips]

Hi,

I have dataset that has 20 classes and only 3 of them has segmentation mask.
In this case, can I train single model that learns detection of all 20 classes and instance segmentation of 3 classes only?
In other words, model that learns object detection and optional segmentation while training. For inference, segmentation network only enables for those 3 classes.

Thank you and looking forward to reply.