Special Topics in Computer Vision, 2P2019, UTB

Time: Wed 2:00-5:00pm Place: A2-202 Homepage: opilab.utb.edu.co/topics-computer-vision/ Instructor: Andrés Marrugo, PhD Email: agmarrugo@utb.edu.co

Course overview

Course description:

This course discusses advanced topics and current research in computer vision. Students are expected to read papers selected from various subareas such as 3D reconstruction, machine vision and inspection, segmentation and grouping, and pattern recognition. Approaches for learning from image data will be covered and include topics from convolutional neural networks, sparse and redundant representations, and others. The course will be a mix of lecture, student presentation and discussion.

Course Goals:

• Gaining research experience in computer vision.
• Get exposed to a variety of topics and become familiar with state-of-the-art techniques.
• Learn to read and evaluate other people’s work.
• Get hands dirty on a research project.

Requirements:

Students will be responsible for reading and reviewing a subset of papers discussed each week (submitting their reviews via a link on the website), participating in class discussions, giving at least one presentation to the class, and completing a research project. In preparing a presentation, students will also be expected to experiment with the methods they describe, either by trying out existing code online or implementing a simplified version of the method. Students will have the liberty to decide what methods and papers interest them in presenting a topic, and should feel free to suggest other topics that interest them.

Presentations will be due on Monday before they are to be presented, meaning that you will need to read the papers, do any necessary experiments, and prepare your presentation several days before your presentation slot. This will allow time for the instructor to give you feedback on your presentation.

###List of suggested topics

1. Feedforward neural networks
2. Image representations (sparse and redundant representations)
3. Feature detection
4. 3D geometry and calibration
5. Segmentation and Grouping
6. Denoising and phase retrieval
7. 3D Reconstruction
8. Object Recognition
9. Semantic Segmentation

Requirements

Students are expected to do the assigned readings, participate in class discussions, write one to two research paper reviews per week, and complete a final research project. In addition, each student will prepare a presentation on a current topic in computer vision, drawing on a set of suggested research papers. The presentation will summarize the topic, as well as present the results of at least one of the student’s own experiments with this topic, using existing code or the student’s own implementation. Note that presentations are due several days before you are scheduled to present, in order to leave time for feedback from the instructor.

Paper reviews:

You will submit two paper reviews per week for the assigned papers. Since we’ll usually be reading more than two papers per week, you can choose which papers to review. Reviews are due the night before the paper will be covered in class, and can be submitted via a link on the website. Each review should address the following:

• Give a brief summary of the paper in your own words, including the main contribution
• What are the strengths and weaknesses of the paper?
• How convincing are the experiments? What, if anything, is lacking?
• Describe one specific way in which the work could be extended.
• Additional comments, including points that were unclear

If you are presenting a paper in a given week, you need not submit reviews that week.

Paper presentations:

Each student will give a presentation in class covering two papers on a topic selected from the list of suggested topics, or another topic of interest. Each presentation should overview the papers and explain key technical details, and synthesize any underlying commonalities or highlight interesting distinctions. The talk should be well-organized and polished, sticking to about 45 minutes. Please run through it beforehand and check the time. Include these components in the presentation:

• Clear statement of the problem
• Why the problem is interesting, important, difficult
• Key technical ideas, how they work, main contributions, strengths and weaknesses
• Evaluation, summary of key experiments and data
• How the technical approaches agree/differ
• Open issues raised in the papers, likely extensions

Try to use applications to motivate the work when possible, and look for visual elements (images, videos) to put in the presentation. Check out the links on the class webpage, and also look at authors’ webpages for supplementary materials. It’s ok to grab a few slides from conference talks etc. when available, but be sure to clearly cite the source on each slide that is not your own.

In addition, in your presentation you will present the results of some experimental evaluation of the main idea in a presented paper. The goal is to implement a distilled version of an essential technical idea in the paper, and show us some toy example of how this works in practice. For many papers, you will be able to find code or binaries provided by the authors online (see the journal Image Processing Online - IPOL). Your experiment should help us gain a more complete intuition about the work we are studying. You might:

• Experiment with different types of datasets
• Examine the method’s sensitivity to relevant parameters
• Show a simplified example highlighting an strength or weakness of the method

The goal here is not to recreate published results or to build systems as described in the paper. Instead, you are looking to make a small illustrative demo that will let us more deeply understand the papers we are reading.

Spend some time playing with your implementation, and put thought into what would be an instructive toy example to show the class. The demo should allow us to learn something about the method, not just see it. If you had to implement something yourself, explain how you did it, and especially point out any details or choices that weren’t straightforward, in case others in the class can build on your experience later when working on the final project. Be sure to explain the rationale for the outcomes, and conclude with a summary of the messages your example illustrates.

In addition to the presentation, make a simple webpage to outline the demo and include links to any existing code, data, etc. you have used. We’ll point to that page for the rest of the class to reference.

Projects:

Students will do a final research project (individually, or possibly in pairs with instructor approval). The final project should be related to a topic we covered in class, but should involve new research, either by extending a topic discussed, doing a thorough experimental evaluation of two or more techniques in a given area, or posing a new approach and performing appropriate experiments. The final projects will involve a proposal, a midterm report (possibly involving a short presentation), and a final written paper (possibly to be submitted to a conference).

Grading

Grading for this class will roughly follow these guidelines:

• Reading assignments (~6, 40%).
• Oral presentation (30%).
• Final project with written paper (30%).

This template is based on the syllabus by Noah Snavely.