Today, Alex Berg gave a VASC Seminar at CMU. At the end of the talk, Takeo Kanade made the interesting point that Alex's work on image parsing has a similar feel to the vision research in the late 70s. According to Takeo, in the good old days many PhD students would undertake the seemingly insurmountable task of building large vision systems and try to make it all work. More recently, PhD students are discouraged from such dreamer-style system-building and focus on more scientific study of narrower problems related to computer vision. In the past, students were actually discouraged from working on small 'easy' problems in vision and everybody tried to go for the gold.
An interesting question is: can we still build systems and be scientific about it (not simply program heuristics)? Alex speculated that vision will inevitably go back (is currently) to the system-building mentality due to the successess of using large amount of training data. I wonder how many things the new generation of PhD will re-invent.