[dahart]

One of the most striking things about intro computer vision explanations & courses today is how completely things have changed in the last 20 years. My “machine vision” textbook from college has near zero overlap with the subjects listed in this blog post, except for face and object recognition as goals, but the techniques taught for object and face recognition today are different from what was taught not so long ago. Seems like CNNs really flipped everything and that nobody starts off with Sobol operators or Medial Axes or Hough Transforms anymore. Most of computer graphics and computer science in general is still building on foundations from thirty, fifty, and seventy years ago, but it seems like vision has changed more dramatically than most other sub-field of CS.

[mistrial9]

at Berkeley I studied multiple papers by students of The Mighty Malik(tm); as said, they may not appear to be aging well, but the topic is far from exhausted.

https://people.eecs.berkeley.edu/~malik/

[uniqueuid]

Absolutely, except - when you get a bunch of embeddings out of your CNN, there will still be a lot of vector math to do anything sensible with it.

[ramacastro]

Yes definitely, I had the same experience in college a couple of years ago, the AI course was about "expert systems" and that kind of old symbolic AI. No neural nets at all. But, interestingly enough I think neural nets will likely be combined with something like symbolic methods when we finally build AGI, mainly for reasoning tasks.

[amelius]

Perhaps they should rename the field to "classical computer vision".