Monday, January 18th, Intro lecture on theories of visual perception and low-level representations.
Paper 1: K. Nakayama, A reductionst Explanation of Perception for the 21st Century.
This paper summarizes the theories of vision and their roles and corresponds to the first part of the slides.
Paper 2: P. Cavanagh, Vision is getting easier every day.
Provides arguments for "simple", 2D-based mechanisms for vision. The paper includes the detailed argumentation based on 2-tone and drawings that was summarized in the slides.
Please use this system for posting comments on the papers.
Investigating how the brain functions has inspired some nice practical tools as discussed in-class. However, is trying to truly model the brains behavior going to help practically? As an extreme example, suppose the brain is really just a neural net w/ millions of nodes & layers. With this known, how can we hope to train it? Of course there is some middle ground, and personally I believe it lies in better representation.
ReplyDeleteAgreed. Just because the Human Visual System doesn't seem to trace contours doesn't mean that we can't extract meaningful information from doing so in software. Maybe you can make a case that the only way to do as well as the HVS is to mimic it, but we clearly don't understand it well enough to mimic it, any attempt to mimic it would probably require a comparable level of complexity, and there are some things for which computer vision is already better than the HVS anyway, such as analyzing medical images.
ReplyDeleteWell, you can also consider the reductionist approach in a similar way as 'deep learning' systems. There are multiple specialized regions that can be trained independently to extract low level information from images, and there can be a multi-level hierarchy of algorithms that are based on the outputs of other algorithms to create more complex behaviors. Organizing visual information into specialized algorithms with multiple layers is something that can certainly be used in computer vision to create better-performing systems than the standard 'black box' approach of taking billions of pictures, defining an arbitrary distance metric and mixing everything together by optimizing a weight vector.
ReplyDeleteBefore I forget: Those of you who are interested in this topic should considered Aloysha's new class on human perception coming up as a new addition to the sequence of vision classes, if you're around.
ReplyDelete