CSC 665 Projects
(under construction)

Link to syllabus

General information about the projects

Projects should contribute to research in some way. Possibilities include engaging in a potentially publishable research direction, working on research infrastructure, or building a tool for some research project. Projects may be done individually or in groups. The scope of the project should be a function of the experience of the team, but all teams should feel challenged. Project directions need to be approved by Kobus, and my incur modifications as a consequence of proposal review by ones peers and Kobus.

All project will require substantive background reading. Likely you will want to integrate what you are reading for your project with what you are presenting to others as class lectures, but this is not required.

In creating a project, you should consider ways in which you can simplify the problem, at least for intermediate goals. Working with real raw data is often difficult. Perhaps you can start with synthetic data? Perhaps you can use image data that is manually marked, instead of extracting features.

Project Proposal

Project proposal RFP.

Link to some slides about writing in general.

Note that these need to be cleaned up, and that "class paper" refers to papers from a previous class where the project had to be about addressing a research problem with a statistical model.

Examples

The full scale proposals below might help you think about how to sell what you are doing. However, these are 3 year proposals, and so the technical plans are much broader than what you need to have for 3 month project. In some sense you are being asked to combine the qualities of a request for funds, with the qualities of a Thesis or Dissertation propoal. Regardless, the rule is always the same --- match what you write to the venue or reason for writing.

Please treat these documents as confidential!

Link to summaries for two research research proposals: CDI and SLIC

Link to descriptions for two research research proposals: CDI and SLIC

Project ideas

The following projects are real research projects, and all of them are challenging. The goal is to make a real contribution to ongoing research, and/or open up opportunities for further research after the course ends. However, it is not necessary to choose a project in the domains below. Conversely If more project ideas are needed, they will be forthcoming. Finally, if you need advice regarding creating a project in a particular domain, please contact Kobus.

Most of these projects are "encumbered" in the sense that they are both course projects and collaborations, and thus are subject to associated constraints which do not necessarily apply to standard course projects. In particular, your contribution is implicitly shared with your collaborators, and you respect that your collaborators part ownership of the intellectual property and artifacts (code, data, and papers). Of course, you can assume the reciprocal relationship with your collaborators.


Blurring faces in images.

This is an excellent semester size project. It also has potential for continued, and possibly funded, work in the spring. This is in collaboration with Nirav Merchant.

For many applications, it is important to be able to automatically de-identify people in images, while still maintaining the overall gist of the scene. This is a need of Google street view, and Nirav Merchant has a real live application that also needs such a capability. There has been some published work in this area, and the first step is an exhaustive literature survey. However, a basic system should not be too hard to implement. One can imagine using the OpenCV to find faces relatively reliably, then modifying the image so that faces are hard to recognize, but humans do not strongly object to the resulting image. Even if a new method for doing this does not emerge, the project will contribute to the vision lab's collaboration with Nirav. In this project it is required that the implementation is available to the vision lab and Nirav, and that the implementation is modular and the modules themselves are integrated into the vision lab software system.


SLIC   related projects.

There are potentially many SLIC oriented vision projects. For some, the first "right of first refusal has been granted to existing SLIC project people. However, there is plenty to do for everyone! The key vision issue in SLIC is to reliably match video frames to slides, determining both the identity, and the geometric mapping (a homography) between the frame and slide coordinate systems. This is relatively solved in good conditions. In more challenging conditions we have proposed a new approach (asymmetric matching) which has yet to be implemented. In addition, one can consider identifying and ignoring (or erasing) the speaker when they are standing in front of the slide, extending and speeding up "bundle adjustment" which makes the homographies more accurate, extracting slides from video when slide images are not available, and doing OCR on such a result.


The Mind's Eye related projects.

We have recently being funded to develop a system which recognizes complex, multi-agent interactions from a single video stream. For example, two people might approach each other, the first might pass a package to the second, and the second might take the package into a building. This all needs to be recognized, represented in a way that can be used by various inference processes, and done with a relatively small power budge. There are many components to building such a thing. If you are interested in the Mind's Eye project in general, a component for it could be undertaken as a course project.

This project requires collaboration with several people and working within collaboratively developed code bases.


Recognizing objects using 3D models

An ongoing vision lab project is to recognize specific objects such as machine parts (e.g., screws, gears) based on graphics representations of them which come from CAD files. For some of what can be done here, the first "right of first refusal has been granted to people working on the project. However, there is plenty to do for everyone. If there is interest in this, details can be provided.

This project requires collaboration with several people and working within existing code bases.


Words and pictures.

Kobus has done a lot of work inferring image labels from large data sets of images that have associated text. See, for example: Kobus Barnard, Pinar Duygulu, Nando de Freitas, David Forsyth, David Blei, and Michael I. Jordan, "Matching Words and Pictures," Journal of Machine Learning Research, Vol 3, pp 1107-1135, 2003.   Recently, we have received some funding for a particular direction for this project, and thus an appropriate course project could potentially be continued in the spring, possibly with funding.

The current state of words and pictures suggests several directions. First, we have work underway to test a number of algorithms, particularly on the region labeling task (most researchers focus on the image labeling task). A possible project is to contribute to this task by implementing some recent algorithms and running experiments.

Another direction is to develop an algorithm along the lines of our CVPR 2008 paper. This will likely require some contribution to the preceding idea as it would be necessary to compare new algorithms with recent approaches, especially if the new algorithm shares some ideas with them.

A third direction is to contribute to the funded pipeline framework for image annotation. Areas where help are needed include image feature extraction, word feature extraction, and integrating object identification (e.g., face identification).

Other projects are possible.

This project requires collaboration with several people and working within the code base which is in C.

Link to syllabus