First Class: Tuesday September 9, 2003
Recent advances in recording, storing and processing large amounts of visual data have focused attention in a number of interesting problems in the junction of Machine Learning and Computer Vision. Traditionally, computer vision research has been aiming to find appropriate models to describe the physics of natural scenes. In this course we will focus on how machine learning and inference can be used to derive or augment such models by using previously recorded visual data, and examine recent results in the field.
We will cover computer vision topics in i) object detection and segmentation, ii) object tracking, iii) object recognition, iv) texture analysis and synthesis v) scene analysis and inference.
Some of the machine learning methods we will be looking at include: Bayesian networks, graphical models, the EM algorithm, Hidden Markov Models, Markov Random Fields, markov chain monte carlo methods, particle filtering, PCA, ICA, kernel based methods, Support Vector Machines, boosting and bagging, graph cuts. We will mostly look at this methods in the context of the computer vision problems mentioned above, with a brief theoretical introduction as necessary. However, many of the ideas and techniques used here are also used in other areas of AI (e.g. robotics, natural language understanding, learning).
The intended audience for this course is graduate students with a background in computer vision, image processing, AI, and computer graphics (esp. of the image based type!). This is NOT an introductory course although due to an error in SOLAR the course is listed as "Introduction to Computer Vision". Students from any department are welcome as long as they have the necessary programming skills for the project, as well as a mathematical background that includes linear algebra, calculus, and basic probability.
The course will be organized as a combination of lectures by the instructor and paper presentations by the students. Each student will have to do two paper presentations, as well as a final programming project. The project should be at the frontier of current research (although not necessarily move the frontier forward). There will be no final exam, however in the beginning of some of the presentations students will be asked to answer a short easy question on the paper that will be presented (to ensure that everyone has read it).
Grading will be: i) 65% Project: Project proposals will be due
on October 21. Mid term demo will be on November 20. Final demo will be
on December 18-19. ii) 20% Presentations. iii)15% Class participation
With permission from the instructor students who do not wish to do the project can take the course for 0/1 credit by registering for the seminar in Computer Vision CSE 656.
References will be mostly research papers as well as the following
Forsyth and Ponce, Computer Vision -- A Modern Approach
Trucco and Verri, Introductory Techniques for 3-D Computer Vision
Duda and Hart and Stork, Pattern Classification
Tom Mitchell, Machine Learning
D. Samaras, Tel. 631-632-8464
Office Hours: Wed 1:30pm to 3pm, Thu 4 to 5:30pm or by appointment
Computer Science room 2429