CS-599: Computational Architectures in Biological Vision
Review-Article Projects
For the review-article projects, you will survey the neuroscience
and computer vision literature on a particular topic, assess the state
of the art in both fields, and make a few suggestions on how one field
could benefit from the other.
Your main source of information for this work will be:
- The library,
- Google at http://www.google.com to
find actual systems, home pages of researchers in the field,
descriptions of ongoing projects, etc,
- Pubmed at
http://www.ncbi.nlm.nih.gov/PubMed/ to find neuroscience papers,
- The Web Of Science at
http://wos.isiglobalnet.com/ to find various papers,
- ResearchIndex at
http://citeseer.nj.nec.com/cs to find computer science online
articles (mostly technical reports or pre-versions of real
articles),
- And the online journals accessible through the USC library home
page, at
http://www.usc.edu/isd/elecresources/subject/j_a.html, from which
you will be able to download PDF versions of research articles.
You will make a selection of articles that you will briefly
describe in your review paper. You will not paraphrase teh authors of
the articles, but rather explain in a few words what the authors did
and what part of their work is interesting and relevant to your point.
It will be important to try to organize your paper in a logical
manner, so that you make a few clear points rather then dumping a lot
of unrelated information. Clearly, your paper will present a very
biased view on current research; you will not describe every existing
project in detail; rather, you will select a few projects which are
well known (based on the number of times they are cited in other
papers) and representative of an approach.
Your goal is to provide a busy reader with a quick assessment of
the current situation in a given research area. Several sample review
articles will be distributed in class, so that you can familiarize
yourself with the format.
Subject topics include:
- The "gist" of a scene. Humans can recognize in a very short time
(200ms) the basic setting of a scene. What is known about how this is
done in biological brains? Are there any computer vision algorithms
that do similar tasks?
- Computational and biological algorithms for face detection.
- Linking the "where" to the "what" pathways: interaction
between attention and object recognition.
- Interaction between visual scene understanding and textual scene
description.
- Adaptive computation: biological mechanisms of plasticity and
their counterparts in computer science.
- Perception-based image and video compression: exploiting knowledge
of the visual system to build better compression algorithms.
- Parallel implementations of vision algorithms and their
relationship to biological algorithms.
- Grouping parts into an object: biological evidence and computer
implementations.
- Coordinate transforms or perceiving a scene while moving our eyes:
biological mechanisms and computer algorithms.
- The best feature detector: what theory exists behind feature
detection, and how close to biological neurons come?
- Linking bottom-up attention to top-down attentional modulation:
biological evidence and potential computer vision usefulness.
Programming Projects
These projects consist of implementing a computer vision algorithm
that is relevant to biology. Implementation can be in the language of
your choice and on the platform of your choice. However, an important
part of the evaluation was to demonstrate that your algorithm can be
applied to real images (not only simple stimuli).
Possible topics include:
- Edge detection with non-classical modulation: Implement
an edge detection algorithm in which the responses from a basic edge
detection filter (e.g., Laplacian of Gaussian) are further enhanced
through two types of interactions: broad non-classical surround
inhibition from feature detectors tuned to similar features as the one
of interest, and specific long-range excitation along contours.
- Statistical Center-Surround: We saw that center-surround
neurons are excited by stimuli presented in a small central region of
the visual space but inihibited by stimuli presented in a broader
concentric surround region. One simple computational algorithm to
achieve this is simply to subtract the mean activity in the surround
from that in the center. One more sophisticated approach, however,
would be to test how statistically different is the distribution of
pixel intensities in the surround, compared to the center.