1.
Decide on a project (0 points)
a.
Fill in a spreadsheet with team, external
collaborators, faculty advisors (if any), and a one sentence summary (“tweet”) of
the research question.
b.
Deadline: September 16. Receive feedback by September
20.
2.
Submit a project proposal (20 points)
a.
A 2 page document
with sections Motivation, Problem Definition, Related Work and Proposed Work.
b.
Deadline: September 30. Receive feedback by
October 7.
3.
Submit an extended abstract (20 points)
a.
A 4 page document
with sections Motivation, Problem Definition, Related Work, Method and
Experimental results (if any).
b.
Aim is to have initial method with initial
results.
c.
Deadline: November 1.
4.
Review work by peers (15 points)
a.
Will be given 1 project to review
b.
Provide a 2 sentence
summary of the project, identifying the key contributions.
c.
Provide 3 concrete suggestions for improving
the project/paper.
d.
Deadline: November 10. Receive feedback by
November 20.
5.
Submit a 3 minute
recorded presentation for the project. Be available for 1 minute of Q&A on
December 7. (15 points)
a.
Deadline:
December 5.
6.
Submit a 6-8 page
final paper. (30 points)
a.
Deadline: Based on registrar.
Project team: Each project must be done in groups of 2-4 people. Single student projects are not allowed: find partners! Projects of more than 4 people are allowed, but you need a strong rationale and you have to run it by me (the instructor).
Project scope: We will aim here for something akin to
a workshop submission. However, if you are using this project for multiple
courses, or if you have external collaborators/ faculty advisors, you can aim
higher (e.g. conference submissions).
To get a sense of what this entails, consider: advanced PhD students in computer vision will typically require at least 3 months to go from concrete project proposal to full conference submission. Here we only have 2 months, and you will have other commitments as well, so aim appropriately. It is better to do a small project that achieves something concrete / answers some concrete questions rather than an ambitious project that will take the entire semester to even frame properly. If you do have an ambitious goal you want to pursue, try to identify the first step that is achievable in this time frame.
What should a project “achieve”?
Good question. The key criterion is that you should contribute something novel to the existing body of knowledge, either in computer vision, or in another field of your liking. Here are the different kinds of projects you can do:
1. You can take an established computer vision problem and try to come up with a better solution. To make this feasible in the time frame, you must identify some significant limitation in the current state-of-the-art, have a concrete idea for addressing this limitation and a justification for why your approach might work. Your project should then evaluate this idea on one benchmark (that may be simplified or reduced in size to allow tractability). E.g.,: you want a new object detection approach that works better for transparent objects.
2. You can define a new problem that hasn’t been tackled before. To do so, you will need to define a benchmark for this problem and show in your project that existing approaches for other problems are either not appropriate, or that they fail badly due to some inherent limitation. You don’t need to evaluate every single possible approach, just the most appropriate ones. E.g., you want to estimate the location of the light source in a scene.
3. You can take existing problems and existing solutions, but you want to use this to solve an important problem / answer an important question in a different field. The standard here is the standard of your field: how well do you do compared to prior work? You can also use this application as a way of commenting on how accurate existing computer vision solutions are and where they fail. E.g., you want to track people to navigate around them in a robotics problem.
4. You can take existing problems and existing solutions but evaluate them in a different way to expose limitations that haven’t been thought of before. The standard here is that you should discover some limitation that was not known before and suggest ways of removing that limitation. An example of this is audits for racial bias the way Joy Buolamwini and Timnit Gebru did in the gender shades project. Another example is evaluating segmentation algorithms for how well they predict the boundaries of each object.
5. Other kinds of projects are fine too, if they satisfy the key criterion above.
1. Clarity of writing [5 points]
2. Concreteness of plan [6 points]
3. Coverage of prior work [6 points]
4. Appropriate scope [3 points]
1. Clarity of writing [5 points]
2. Coverage of prior work [5 points]
3. Correctness of approach / experimental setup [6 points]
4. Sufficient progress towards goal [4 points]
1. Correct summary [5 points]
2. Technical quality of suggestions [6 points; 2 for each suggestion]
3. Politeness and respect for authors [4 points]
1. Clearly specified problem / research question [5 points]
2. Clearly identified limitations of prior work [3 points]
3. Clearly communicated approach and results [7 points]
1. Clarity of writing [5 points]
2. Coverage of prior work [5 points]
3. Correctness of approach / experimental setup [5 points]
4. Incorporation of meta reviews where possible [5 points]
5. Discussion and interpretation of results [5 points]
6. Significance of contribution [5 points]
Roughly speaking, we will keep the following absolute scale:
1. [90, 100] : A
2. [80, 90) : A-
3. [55, 80) : B+
4. [45, 55) : B
5. [35, 45) : B-
6. < 35 : C+
You are guaranteed a grade of C+ if you show up to class. You will likely get at least a B+ if you do some relevant project.
A+ grades will be reserved for students who (a) qualify for an A, and (b) either did exceptional projects, or were exceptional in class discussions.