Building compact subset of Vision model for MLKit

Question

Vision model in MLKit's Cloud API has all the labels I need. In fact I need only 100 of its 10000 labels. Can we retrain a compact version of that model to detect only 100 labels and deploy it on the android device so the app can run without internet connection?

Answer 1

That's precisely what Firebase ML Kit's image labeler does: it detects a subset of the labels that Cloud Vision detects by running a smaller ML model on-device.

If you want to control which labels it detects on the device, you'll have to train/use a custom model.