Computer Vision News - April 2022

30 Computer Vision Book Some will focus more on image processing and low-level computer vision, while others will focus more on machine learning and other aspects of computer vision. There will always be overlap, but maybe a little more intentional separation than a single computer vision track of education. Of all the novelties and innovations you have seen in the last few years, can you pick theoneyou thinkexperts andscholars should be most careful not to neglect? MT: It would be hard to choose just one, but something that strikes me as important is the general limitations of what we know about machine learning. Machine learning has made incredible progress, but much of it has been pragmatic in the sense that people have tried many things, and some have been successful, but there’s a lot we don’t know fundamentally or theoretically about limits. For example, how much data is needed for a given problem? How do we know whether there is enough variety in the data? There is helpful best practice and guidance, but still a lot of missing parts. The whole subfield of adversarial learning is an example where people have found situations that don’t work so well. People are exploring that field to get a pragmatic view of these limitations, but we need to know theoretically and fundamentally what the limitations and possibilities are and the appropriate guidance. RD: I think Matthew is absolutely right about the need to be able to theoretically ascertain the amount and variety of data needed in any application. Indeed, that is the1-million-dollarquestion that previously put paid to the old neural networks of the 1980s. to produce chapters that were even better than their initial drafts – and of coursemore didactic and beneficial for the eventual readers! What was it like working together? MT: Believe it or not, we still haven’t met face to face, but we’ve been able to meet online. Roy is a very organized person and easy to interact with. We ended up complementing each other well in terms of our styles and focus. I didn’t know his prior textbooks in detail before this, so it allowed me to get to know them. His most recent introductory computer vision textbook is an excellent book to have on your shelf and use for classes. RD: I found it a remarkable experience. As Matthewsays,wehadnotmet inadvanceof working together, but it soon became clear that our different past experience usefully led to different global and local slants on the writing. I should also remark that I was overwhelmed that this was the same Matthew Turk who had long ago invented the impressive eigenface approach to face recognition! With computer vision changing so much in the last few years, how do we balance the fundamentals of the field with the things that are most important now? The Editors: That’s a question people are still struggling to answer. Knowing the fundamentals of computer vision and image processing has been crucial in making all this tremendous progress, and many people are worried we’ll have a whole new generation coming along who don’t know the basics. Part of the answer is that the field will probably start to break up more than it has in the past. It’s hard for one person to be an expert in everything.