CS279 Fall 2006

TR 4:00-5:50, 5252 Boelter Hall
COMPUTER SCIENCE 279: VISUAL RECOGNITION
SYLLABUS
 
DATE
TOPIC
QUESTIONS
REFERENCES
9/28/06
Introduction; organization; administration


10/3
Background material

Any book on statistical decision theory; e.g. Duda-Hart
10/5
Local features and wide-baseline matching: SIFT, MSER, Gabor Jets
How can we tie the choice of feature to the task? Currently features are designed for wide-baseline matching, but used for category recognition. Can we devise better features for the latter task?
Lowe, Matas, Koenderink, Kadir, Sirovich-Kirby
10/10*
MURI Kickoff Meeting


10/12
Viewpoint invariance: Affine invariance, scale-space, deformations
Assuming statistics on  likely deformations are available, can we relax the notion of invariance to "insensitivity" "to most likely deformations? When should invariance/insensitivity be achieved by design, and when should it be achieved as part of the matching process? (recognition via features vs. recognition via reconstruction)
Schmid, Vedaldi, Lindeberg
10/17
(Meltzer, Voroninski)
Illumination: Simple phenomenological models (Phong,  Torrance-Sparrow), Lambertian reflection, bas-relief ambiguity. Global approximate models: Contrast functions, gradient direction.
General illumination invariants do not exist: Can we design features that are insensitive to the most likely illuminations? Can color help?
Any book on computer graphics; Morel, Belhumeur
10/19
(Nabil)
Feature selection: Link to task (object vs. category recognition)
See 10/5, in the broader context of feature selection in the AI literature.
Koller
10/24
(Raptis, Fontanelli)
Dimensionality reduction: PCA, LDA, TCA, manifold learning
Mostly background material (tutorial)

10/26
(Vedaldi)
Information-theoretic approaches: Basics of info-theory, Rate-Distortion theory; Rate-Recognition theory.
How can we use information theory to guide the selection of features? We want maximally insensitive to viewpoint and illumination, but maximally informative with respect to the recognition task. How do we characterize that? What are the tradeoffs?

10/31
(Fulkerson)
Computational Approaches: Complexity, Boosting, Bagging
Some background (Adaboost); how do these approach relate to those based on a model, or based on complexity?
Geman, Viola, Freund,
11/2
(Jurgens, Zhu)
Grammars and Semantics
Why is it necessary to "parse" a scene in order to "understand" it?

11/7*
Illumination (Paul Debevec, USC)
Guest lecture

11/9
(Ling)
Parts and Compositions
What is a "part"? How can we model complex scenes as the juxtaposition of simpler parts?

11/14*
Actions (Trevor Darrell, MIT)
Guest lecture

11/16
(Wnuk)
Context
What is context? Local vs. Global?
Torralba
11/21
(Guidi, Sasai)
Natural Statistics; occlusions, textures, perceptual transitions
How can we collect statistics on the scene (rather than the image) from images? How do we exploit such statistics? Shape statistics? Refleectance statistics? Deformation statistics?
Mumford
11/23
Thanksgiving Holiday
Happy Turkey

11/28*
SLAM (Frank Dellaert, GeorgiaTech)
Guest lecture

11/30
(Hong, Yi)
Deformable Templates: Sparse vs. Dense
Grenander's tenet: Variability in the domain of the image (group deformation) vs. range (intensity)
Grenander
12/5
(Bissacco)
Actions and events


12/7
Modeling vs. Learning, Recognition vs. Segmentation
How should we decide what to "learn away" and what to model explicitly? What are fundamental complexity bounds?

* indicates a different location, to be announced in class.


COURSE DESCRIPTION

This is an advanced course on computer vision, focusing on visual recognition, that is the problem of recognizing objects or scenes from images. The course will explore the state of the art in the field. Each student will be required to participate actively with presentations and discussions. For each  topic, a small group will be formed that is responsible to present the state of the art in that topic, and to lead a discussion.

PREREQUISITES

There are no formal prerequisites. However, a solid background in linear algebra and basic probability and stochastic processes is highly recommended.
GRADING POLICY

This course is listed as "letter grade only." Grade will be based on participation and on the quality of the review work and role in the discussions.

OFFICE HOURS

Tuesdays 12:00-1:45. Additional hours by appointment. Please send email if you plan to attend office hour.