Instance segmentation performance depending on the number of classes

Question

I have a setting in which I have approximately 10 classes, one of which I would need to put an extra emphasis on. Whilst training a mask-rcnn instance segmentation network (detectron2) I found out that I can do a lot better if I separate this one class and train a completely different network for it and then afterwards combine the predictions. Is this true in general for these networks? It would seem reasonable that the network would perform better with fewer classes to predict, but I couldn't find any good writings about this.

Soumya · Answer

The process you have used it spot on. Your problem: Give more importance to one class than the rest will have to be hard-coded in. There will be no pre-trained networks for that class especially.
One thing to note could be that assuming you have to segment roads in a road scene problem. You can create a function which can bias the weights towards the road specifically. This is just a thought, I haven't implemented it myself!

Instance segmentation performance depending on the number of classes

One Answer

Add your own answers!

Ask a Question