ROBOTICS INSTITUTE
CARNEGIE MELLON UNIVERSITY

16-721: Advanced Perception

Project Guidelines

One element of this class will be a semester-long research project.  The project should be at the frontier of current research, but need not necessarily move the frontier forward.  Replicating the results of an innovative recent paper would be a good project, for example.  Projects must be closely inspired by one or two specific recent high quality papers.  They should preferably have an experimental component, but "crass empiricism" is to be avoided.  As much as possible, you should use a high-level programming environment such as Matlab. The schedule for the projects is as follows.  After each deadline, the instructor will be available to discuss the projects in person.  You should take the initiative to set up an appointment, preferably on a Tuesday or Thursday.

(1) On Tuesday October 18 at 6pm or earlier, you should hand in a project proposal.  This should explain explicitly and clearly what you will do.  In particular, the proposal should include:

  • a list of four to six milestones, and deadlines to achieve the milestones
  • a list of questions to be answered during the project and discussed in the report
  • if the project is experimental, which existing software you will use directly or build upon
  • if the project is experimental, precisely which datasets you will use and where you can obtain them quickly
  • a few recent and very closely related papers that you will build on, with full bibliographic data.

The proposal should be written in well-organized continuous English, as opposed to just an outline.  Most of its text should be reusable in your final report. 

(2) On Tuesday October 25 at 6pm or earlier, please hand in a revised project proposal.  This must follow all the instructions above, and take into account comments received from the instructor and other sources.  It should be two to four pages long when formatted following the NIPS instructions or one to two pages long using the CVPR instructions. Since you will have to use LaTeX eventually, you should start now with the revised proposal.  In the first phase of the project, some important tasks include:

  • formulating clear, sensible hypotheses to test,
  • finding useful training sets and test sets,
  • surveying closely related recent papers, and
  • selecting and obtaining existing software, and developing new software.

These are tasks that can and should be performed mostly in parallel, not sequentially.

(3) On Tuesday November 22 at 6pm or earlier, you should hand in a complete, polished draft of your project report.  The report should resemble as closely as possible a good submission for a computer vision or pattern recognition conference.  You should follow either the NIPS or the CVPR formatting instructions and the 250A report guidelines.

(4) On Tuesday Nov. 29 at 2pm or earlier, you should submit the final version of your project report in both hardcopy and pdf format (by e-mail). This will be graded following the 250A grading criteria.  Perfect academic honesty is required.  The final version of your report should be a PDF file in NIPS or CVPR format that is under one megabyte.  Avoid these common formatting mistakes.  To maximize the readability of your paper, and to minimize the size of the PDF file, make sure that you use only standard Postscript fonts.  Use the command \usepackage{pslatex} in LaTeX to achieve this.  Also provide an ascii abstract of your report that follows these guidelines and is 150 to 200 words long.

While doing the project, remember that winning at research is similar to winning in many other fields of endeavor.
 

  • Build on an idea that has been successful in previous work.
  • Make the description of your work understandable, attractive, and memorable; pick catchy names.
  • Keep the work simple.  Let the basic ideas shine through.
  • More papers = more ways to have impact and be noticed.

It Takes Soup And a Dream     (Newsweek, March 26, 2001, page 8.)

Susan Runkle isn't just M'm! M'm! Good! She's the M'm! M'm! Best! Her Polynesian pork chops have been named a new Campbell's classic, an honor worth $20,000. Peri asked Runkle to dish on how to win a cooking contest:

  • 1. Polynesian pork doesn't just happen. Work backward by building on a taste you enjoy. (Runkle started with the tangy pork-pineapple combo.)
  • 2. Pick a catchy name. "Use alliteration: red raspberry... something."
  • 3. Keep it simple. Skip recipes "six pages long with chipotle chiles."


4. More recipes = more ways to win. "I sent in a pasta dish, a beef casserole--I call it pepper-pot beef--a Tuscan chicken, a Mexican pizza and a chicken potpie.  Husband's not too keen on fish."

Most recently updated on Sept. 13, 2005 by Serge Belongie. Content adapted from Charles Elkan's CSE 254 Project Guidelines.