Choose a topic based on option 1 or 2 below. When choosing a topic, remember that as a yardstick, the project should require roughly twice the amount of effort of one of the problem sets. This includes effort in writing the report and presentation. Choose something that is not trivial, but at the same time, practical to complete by the end of the term. (Avoid anything that requires training large neural networks.)
Once you have selected a topic, submit a 2-3 paragraph proposal by 11:59pm Sunday Oct 21st. Describe what you intend to do for the final project, including details of any relevant papers and online resources, plans of what you will likely implement, analyze, etc.
The final project report will be due December 9th at 11:59pm. The report should be 4 pages (not including references) in two-column CVPR format. You can use this latex template to prepare your report.
Option 1: Read and analyze one or more computer vision papers.
- Recent or classic papers from ICCV, ECCV, CVPR. Vision papers from other venues also appropriate.
- Either implement---all or a subset---or if implementation available, modify / analyze.
- If purely theoretical paper, do simulations, find examples, etc.
- Key is to demonstrate you understood method
- What problem did it solve ?
- Why was it needed over and above what existed before ?
- What are its limitations ? Is the problem solved ?
See suggestions / pointers below.
Option 2: Apply what you've learned in class to a problem you care about.
- Read up on most relevant related work.
- Implement adapted method for your problem.
- Analyze results. Did it work ? If so, how well. If not, why not.
Your grade will depend as much on presentation and analysis as on technical contribution.
- Abstract: 3 pts (One paragraph succinct summary)
- Introduction/Motivation: 5 pts
Why is this problem important, what is the vision task, prelude to rest of the report.
- Related work: 4 pts
How have other people solved it ? What are other similar problems ? Read, describe.
- Description / Experiments / Technical Correctness: 10 pts
- Conclusion: 3 pts
Pointers for Option 1
Note that depending on the paper, you might want to implement part of it, all of it, or the paper as well as some prior/previous work. You are always free to select papers not in here.
- Barron & Malik: Shape Illumination and Reflectance from Shading - PAMI 2015 - Code at http://www.jonbarron.info
- Wu et al., Robust Photometric Stereo - CVPR 2010
- Johnson & Adelson: Shape Estimation - CVPR 2011
- Buades et al.: Non Local Means Denoising - CVPR 2005
- Krishnan and Fergus: Fast Image Deconvolution - NIPS 2009
- Zoran & Weiss: EPLL Algorithm for Image Restoration - ICCV 2011 - Code available at https://people.csail.mit.edu/danielzoran/
- Zontak and Irani: Image Statistics with Patch Recurrence - CVPR 2011
- Yamaguchi et al.: Motion & Stereo - CVPR 2013
- Brox et al.: Optical Flow - CVPR 2009
- Levin et al.: Interactive Colorization - SIGGRAPH 2004
- Rother et al.: Interactive Segmentation with GrabCut - SIGGRAPH 2004
- Efros & Freeman: Texture Synthesis - SIGGRAPH 2001