This was my first thought: I could do this in half a day. Then you look at their site and all the domains of application. My guess is they probably use a varie

This was my first thought: I could do this in half a day. Then you look at their site and all the domains of application. My guess is they probably use a variety of models to get 100%: edge detection, CNN-style object detection, all sorts. And then aggregate/choose between the resulting predictions. Then they will probably have some layers of geometrical estimators. The challenge here is 100% and on a wide variety of images. They'll need to maintain and collect data across a lot of domains, and find ways of coping with non-ideal ("in the field") input.