(SP22-CS 598) Efficient & Predictive Vision

Grading breakdown: 10% Participation + 30% Presentation + 60% Project

Participation

For each regular class, you are required to submit the pre-class questionnaire. Typically, we have two papers within a topic. You are free to comment either of the pappers (or all of them!). In addition, you are encouraged to engage in discussions on Campuswire. The presenters are encouraged to incorporate some comments from the posts in their presentations.

Here are some things to think about

What might be the limitations of the approach and its assumptions: where might it fail, even if all of its components work as intended?
What might this paper have done that was previously impossible?
What are broader take-aways from the paper that can be applied elsewhere?
What's the most important part of the paper or the method?
Do the experiments justify the paper's argument?

Remember also that while criticizing (especially in hindsight) is often easy, it's equally important to defend a paper.

Presentation

Students will form groups of up to four (1-4) and jointly develop a presentation on an assigned topic. Each group member must deliver a portion of the presentation. Each student should present twice during the course.

Signup: An email has gone out to all the registered students with the link to a signup sheet (the link has also been posted on Campuswire). Topic assignment is first come, first served, though in case nobody signs up for some topics, we may ask some students to switch to ensure coverage. Each student should sign up two topics.
After the deadline for registered students, any unregistered students interested in taking the course are free to sign up for the remaining spots. The presentation schedule will be finalized by the end of January 28.
Practice presentation (5% of grade): You should reach out to Liangyan and/or Shengcao/Yunze to schedule a time for your practice presentation approximately a week before your presentation date. The goal is to enable feedback to ensure the highest possible quality of the in-class presentation. All group members must attend. The practice presentation is not expected to be polished or 100% complete, but the grading will be based primarily on evidence that the group is taking the preparation seriously.
Slides (10% of grade): By the day before the scheduled presentation, the group must post the link of the slides on Campuswire, to be made available to all the students. (The slides can be further polished before the class). After the presentation, the group must submit the slides in PowerPoint format to Liangyan, Shengcao, and Yunze, and we will make them available on the course webpage. Late submission of the links or slides will forfeit this portion of the grade.
In-class presentation (15% of grade): The presentations will be graded based on clarity, technical depth, successful synthesis of content from multiple papers, ability to involve the students, and responsiveness to feedback from the practice presentation. Presentations will be recorded and made available to other students registered for the class.

Here are two important guidelines

Within each topic, two papers are typically listed. The key objective is to identify a coherent story over these papers. Your presentation typically should not be organized as a sequence of separate single-paper summaries. You are also not expected to cover every part of every paper.
Use of external sources: It is not allowed to use an entire slide deck from another source "as is" as the basis for your presentation. It is fine to "borrow" some slides, graphics, or demos. However, please explicitly give credit whenever you use material from other sources.

Timeline:

Jan 27, 10:00PM: Presentation signup (for registered students) due.
Jan 28, 10:00PM: Presentation signup (for unregistered students) due.
One week before the presentation, specific time to be scheduled: Practice presentation due.
The day before the presentation, 10:00PM: Links of presentation slides due.
The presentation day, 10:00PM: Presentation slides due.

Course Project

This is an opportunity to try out the ideas discussed in the class. Computer vision is a rapidly developing field. We are looking for projects that get you to think about things differently: a well-motivated and well-executed crazy idea that experiments show doesn't work as well as existing methods will receive a higher grade than adding predictable features to an existing method and getting a performance gain. Students will form groups of up to three (1-3) for projects. You are free to come up your own projects, and here a list of project topics for reference:

Implement prediction architectures (e.g. RNN, Transformers, GCN). Try on different datasets and/or on different domains.
Analyze and discuss the existing evaluation metrics of prediction tasks.Propose some new evaluation metrics that you think are more reasonable.
Implement and test several network compression methods (e.g. weight pruning, quantization), and analyze their accuracy-efficiency trade-off on a benchmark. In addition to compare individual compression techniques, try to find a proper combination of them to achieve a better performance.
Use efficient architectures as building blocks to create a novel model and perform an interesting task, e.g. real-time 3D detector running on a cell phone
Propose a method to accelerate / robustify existing neural architecture search algorithms, and evaluate it on widely-used benchmarks.
Transfer a technique originally developed for efficient vision models to other areas, such as natural language processing or robotics.

An ideal project should

Be different. Take this as an opportunity to try something new and exciting that might not work.
Be the result of a considerable amount of effort in terms of thinking and working. You should be able to explain why you're doing what you're doing.
Be described by a polished report and presentation. Remember: we see the report and presentation and not your code and the hours you put in.

Timeline:

Feb 21, 10:00PM (10% of grade): Two page project proposal due. >
May 11, 7-10PM (20% of grade): Final project presentations.
May 13 (30% of grade), 10:00PM: Four page final report due. Use CVPR/ICCV format (Latex+Word Templates here).