Computer Science Department

The University of Waterloo

CS 898 --- Topics in Computer Vision

Syllabus: Winter 2024

General Description

This course is a graduate seminar on computer vision covering standard pixel labeling problems (e.g. segmentation, stereo, reconstruction, restoration, etc). Most of the recent success in vision relate to fully-supervised settings, e.g. in image classification. However, accurate pixel-level supervision is prohibitively expensive and practically infeasible for most dense image-labeling probelms. This seminar focuses on much more difficult open-ended problems in unsupervised, meta-, weakly-, self-supervised, and other similar few-labels settings, which are far from being solved. Besides "shortage of labels", we may also look into "inacurate labels". While we will consider many general ideas, particular attention will be given to the design of the loss functions.

Since computer vision concerns objects and image data embedded in Euclidean spaces (3D world and its 2D projections), there are many important modelling constraints. Their representation is critical for weakly-supervised and unsupervised methods. We will start with several lectures covering standard structural regularization models (geometric, graphical, MRF/CRF, information- or physics-based, deformable) focussing on the corresponding loss functions and their standard optimization methods (combinatorial, relaxation-based, message passing, submodular approximations, spectral methods, mean-field, surogate optimization). After this general background is covered, we will discuss (recent) computer vision papers on "weakly-supervised" NN training. While this is a very broad area, we will explore general ideas addressing the shortage of labels, e.g. using structural and other modelling or regularization constraints. A tentative open list of proposed papers is posted. Each student will present 3 papers selected from the list on the first-come-first-serve basis (email me). Students can also suggest some papers to be added to the list.

Lectures: Wednesdays 12:00am - 2:50pm, DC 2568


Instructor Information

Instructor: Prof. Yuri Boykov
Office: DC3142
Office Hours : after class or by appointment
E-Mail: yboykov 'at' uwaterloo.ca


Prerequisite

Note: Students are responsible for ensuring that they have either the prerequisites for this course, or written special permission from the instructor. If a student does not have the course prerequisites, and has not been granted a special permission, it is in his/her best interest to drop the course well before the end of the add/drop period. If a student is not eligible for a course, he/she may be removed from it at any time, and will receive no adjustment to his/her fees. These decisions can not be appealed. Lack of prerequisites may not be used as the basis of appeal.

Course Website

The website for the course is https://cs.uwaterloo.ca/~yboykov/Courses/cs898/. Lecture notes, discussion papers, and other materials will be posted on this web site.

Textbooks, Lecture Notes, Research Papers

There is no required textbook for this course. Lecture notes will be available on the course web site. Substantial part of the course will be discussion of research papers in computer vision (to be posted). Some topics will have recommended readings from selected text-books on computer vision and optimization, e.g.

Course Content

The course will first focus on unsupervised low-level image segmentation and relared pixel labeling problems (depth, motion, restoration, inpainiting, etc) based on regularization and optimization. Then, we will discuss weakly-supervised or self-supervised methods covering the range between unsupervised methods and fully supervised learning-based techniques. Other related topics can be discussed. Suggestions from students are encouraged.

Course work and Evaluation

Half of the mark is based on the final project to be due at the end of the exam period (April 23). The projects should be related to the papers discussed in this course. There will be (tentatively two) homework assignments (or quizzes) in the first half of the course. Those will be administered via Learn. Besides (tentatively three) papers, students will also have to present a proposal of their project. Presentations of the papers and project proosal will be evaluated based on clarity.
Participation 5%
Assignmemnts/quizzes 15%
Paper Presentation(s) 30%
Final project (proposal+report) 50%

Academic Integrity

In order to maintain a culture of academic integrity, members of the University of Waterloo community are expected to promote honesty, trust, fairness, respect and responsibility. [Check www.uwaterloo.ca/academicintegrity/ for more information.]

Grievance

A student who believes that a decision affecting some aspect of his/her university life has been unfair or unreasonable may have grounds for initiating a grievance. Read Policy 70, Student Petitions and Grievances, Section 4, www.adm.uwaterloo.ca/infosec/Policies/policy70.htm. When in doubt please be certain to contact the department's administrative assistant who will provide further assistance.

Appeals:

A decision made or penalty imposed under Policy 70 (Student Petitions and Grievances) (other than a petition) or Policy 71 (Student Discipline) may be appealed if there is a ground. A student who believes he/she has a ground for an appeal should refer to Policy 72 (Student Appeals) www.adm.uwaterloo.ca/infosec/Policies/policy72.htm.

Note for Students with Disabilities:

The Office for persons with Disabilities (OPD), located in Needles Hall, Room 1132, collaborates with all academic departments to arrange appropriate accommodations for students with disabilities without compromising the academic integrity of the curriculum. If you require academic accommodations to lessen the impact of your disability, please register with the OPD at the beginning of each academic term.