Without a public testable version, it's hard to say how it would handle weird edge cases or drawings that don't closely match objects in its training data (which we don't know the extent of).