I think you're right that the barrier is a lot lower now that there is a somewhat successful blackbox abstraction. In the past you need to know that to recogni

I think you're right that the barrier is a lot lower now that there is a somewhat successful blackbox abstraction. In the past you need to know that to recognize lines: hough transform, recognize polygons: line simplification, recognize face: cascades, etc etc. Now? You can almost just feed it arbitrary labeled training data and do well without any sort of feature engineering. Just another api to glue.