Course Info

This course discusses advanced topics and current research in computer vision. Topics will be selected from various subareas such as physics based vision, geometry, motion and tracking, reconstruction, grouping and segmentation, recognition, activity and scene understanding, statistical methods and learning, systems and applications.

This course is geared toward graduate students who are interested in gaining research experience in computer vision. The course will consist of lectures, paper reading, in-class presentation/discussion, and a final project on a research topic. 

The goals of this course are:
  • To get exposed to a variety of research topics in computer vision and become familiar with state of the art techniques. 
  • To develop skills in reading research papers and evaluating other people's work. 
  • To gain experience in conducting original computer vision research.
Prerequisite: EECS442 or permission of the instructor. 


Time: TTH 3:00pm-4:30pm
Location: 1008 FXB
Instructor: Jia Deng (jiadeng).  Office hours: Thursdays 2-2:50pm or by appointment, 3640 Beyster. 
GSI: Yu-Wei (Johnny) Chao (ywchao)

Textbook

There are no required textbooks. You may find the following books useful as references. 

Course Organization

The main components of the course include:
  • Presentation (20% of grade)
  • Paper reading and reviewing (20% of grade)
  • In-class discussion (10% of grade)
  • Final project (50% of grade)
Presentation

You are expected to make at least one in-class presentation (~30min) on a topic based on a list of assigned papers, possibly with another student. Below is a suggested structure of your presentation:
  • Problem statement and motivation. Introduce the necessary background.
  • The technical approach
  • Discuss how the approach relates to other work. Discuss the commonalities and differences. 
  • Points for discussion (one slide): e.g.  strengths and weaknesses, possible extensions, alternative approaches etc. 
In your presentation, you should talk about the main technical content of the papers. You should try to strike a good balance between high-level intuitions and low-level technical details. In particular, you should not dump equations or pseudo code into your slides. You should always explain intuitions if you show math and illustrate with examples if you can. 

Equally importantly, you are also expected to discuss how the technique relates to prior work. This means that you need to think carefully about the commonalities and differences. 

You may use materials and slides from others, as long as you attribute the sources. For certain papers, slides and recorded talks are available online. In this case, you may use the slides but you may not recite the script or just play the recorded talk. You must always present in your own words. 

Important: you must email the instructor your draft slides by 5pm one week prior to the class you are signed up for. This allows time for iterations. 

Paper Reading and Reviewing

For each class there will be a list of papers to read. Some of them are required and some of them optional. Before each class, you must read all the required papers and write a summary for each of them. Your summary should have two paragraphs:
  • First paragraph (2-5 sentences): What is the problem the paper is solving and why is the problem important and/or interesting?
  • Second paragraph (2-5 sentences): What is the main approach? How is it evaluated and what are the main results?
In addition, you need to write reviews for assigned papers during the semester. A review is a summary plus the following two components:
  • (at least one paragraph) What are the main contributions and strengths of the paper? 
  • (at least one paragraph) What are the main limitations and weaknesses of the paper? How would you improve it or how would you do it differently? 
If you are writing a review for a paper, there is no need to submit a separate summary. We will assign reviews taking into account your preferences. 

You should write the reviews and summaries as if you are communicating the gist of a paper to a fellow graduate student who has not read it but has the same background knowledge as you do. You should be concise. This is not the place to elaborate on technical details unless you have major criticism on some technical points. 

You must write your reviews and summaries in your own words. You may use quotations occasionally, but you may not copy or closely paraphrase any parts of the paper. Discussing papers with fellow students is allowed and encouraged, but you must write your own reviews/summaries and may not read other people's reviews/summaries before submitting yours. 

Your reviews and summaries are due at the beginning of each class. An online submission system will be set up and posted soon. Late submissions will receive zero credit. Your lowest scored review and two lowest scored summaries will not count toward your grade. Reviews take up 10% of total grade and summaries anther 10%. 

In-Class Discussion

You should attend the classes and actively participate in discussions during class. 

Final Project

In your final project, you are expected to explore a research topic in depth. You can do a project on your own or form teams with 1-2 other students (the maximum team size is 3). The expectation will be proportionally higher for bigger teams. All team members will receive the same grade. 

Possible types of projects include but are not limited to:
Your project must include something original, which can be a new dataset, a new analysis, a new theory, a new algorithm, or a new programming framework, etc. Your project cannot be purely a literature survey. 

You are encouraged to be creative and "crazy" in your project proposal. To get a good grade, you are not required to produce positive, publishable results by the end of the semester. If you get negative results, you should do careful analysis to find out why. Negative results with insightful analysis will be graded generously. 

Your project topic must have the instructor's approval. You are strongly encouraged to schedule appointments or come to the office hours to discuss your proposal with the instructor. 

You are encouraged to reuse code developed by others as much as possible, as long as you have permission from the author(s), e.g. publicly available code with BSD licenses. This way you can focus your time on the original part of your project. In your final report, you should acknowledge all external code you have used. You are discouraged from re-inventing the wheel. For example, you should not implement SIFT from scratch unless there is something significantly original in your implementation. Keep in mind that your project will be judged the same way as research papers: what matters is your new ideas and results, not how you reproduced existing results. 

Initial proposal (hardcopy due in class on 9/23): The initial proposal should be 1~2 page in the CVPR format (including references). It should include the names of your teammates (if any), a problem statement, a brief review of related work, an outline of your approach and why you believe it would work, and the expected outcome. 

Final proposal (hardcopy due in class on 10/7): By the time you turn in your final proposal, you should have iterated with the instructor and reached an agreement on the topic and the scope of the project. The final proposal should be a revised version of the initial proposal, including a concrete technical approach as well as weekly milestones starting the week of 10/7. The final proposal should be at most 4 pages in CVPR format. 

Midterm report (hardcopy due in class on 11/4): The midterm report should give a detailed account of the current progress and include any preliminary results. This is an opportunity to get feedback on your project. 

Project presentation (12/4 or 12/9): You will present your project to the class at the end of the semester. All team members should participate in the presentation. 

Final report (due on 12/10): The final report should be written as a research paper in the CVPR format. It should be no more than 8 pages (references included). As is typical in a conference submission, you can include supplemental materials with no length limit but your reader/reviewer should be able to evaluate your contributions by reading only your main report. Along with your report, you need to submit all your source code and documentation on how to reproduce your results.