Looks like the "learn it all end-to-end" deep learning hype is starting to fade as some people realize putting some actual prior knowledge in computer vision systems is helpfull.

http://alexgkendall.com/computer_vision/have_we_forgotten_about_geometry_in_computer_vision/