 |
Selected Topics in Computer Vision
Dr Vincent Lepetit -
Dr Fethallah Benmansour -
Dr Kevin Smith -
Dr Yunpeng Li -
Dr Mario Christoudias
Every Friday of Fall semester. 10:15-12:00, ROOM BC02.
Schedule for the paper presentations:
| Dec. 2 |
Anil Yuce |
BRAMBLE: A Bayesian Multi-Blob Tracker. Isard and MacCormick, ICCV 2001. |
| Dec. 2 |
Sofien Bouaziz |
Non-Local Sparse Models for Image Restoration. J. Mairal, F. Bach, J. Ponce, G. Sapiro, and A. Zisserman, ICCV 2009. |
| Dec. 2 |
Laurent El Shafey |
Fields of experts: A framework for learning image priors. S. Roth and M.J. Black, CVPR 2005. |
| Dec. 9 |
Ab Al-Hadi |
Higher-Order Clique Reduction in Binary Graph Cut. Hiroshi Ishikawa, CVPR 2009. |
| Dec. 9 |
Jean-Charles Bricola |
Snakes, Shapes, and Gradient Vector Flow. C. Xu and J.L. Prince, TIP 1998. |
| Dec. 9 |
G. Subrahmanyam |
On regularized Laplacian zero crossings and other optimal edge integrators. R. Kimmel and A.M. Bruckstein, IJCV 2003. |
| Dec. 9 |
Mirko Raca |
Efficient Belief Propagation for Early Vision. Pedro F. Feltzenszwalb and Daniel P. Huttenlocher, IJCV 2006. |
| Dec. 16 |
Eleni Krupi |
Mean Shift: A Robust Approach Toward Feature Space Analysis. D. Comaniciu and P. Meer, PAMI 2002. |
| Dec. 16 |
N. Thanikachalam |
Robust Tracking-by-Detection using a Detector Confidence Particle
Filter. Breitenstein, Reichlin, Leibe, Koller-Meier, and Van
Gool, ICCV 2009. |
| Dec. 16 |
Stefan Lienhard |
Removing camera shake from a single photograph. R. Fergus, B. Singh, A. Hertzmann, S. T. Roweis, W. T. Freeman, Siggraph 2006.
|
| Dec. 16 |
Zhou Xue |
A High-Quality Video Denoising Algorithm based on Reliable Motion Estimation. C. Liu and W. T. Freeman, ECCV 2010. |
The paper pdfs are still accessible here. Remember:
- Your presentation should last 20 minutes. Allow 2 or 3 minutes for questions.
- You should have about 1 slide / minute.
- A good way to start your presentation is by showing results: This
way, your audience will understand quickly what is the problem the
paper tackles, and what kind of results they can expect.
- Use formulas only when necessary, fewer is better.
Objectives:
The goal of Computer vision is to process images acquired with cameras
in order to produce a representation of objects in the world. Although
there already exists a number of working systems that perform parts of
this task in specialized domains, the generic "Vision Problem" is far
from being solved. No existing system can come close to emulating the
capabilities of a human.
The aim of this course is to teach the students how to reason in
rigorous manner on Computer Vision related problems. We will focus on
a few subsets and study them in depth. We will try to understand what
makes Computer Vision so hard, what is the state-of-the-art, and what
are promising directions for future research.
Content:
The content will vary from year to year. This
year (2011/12), we will focus on Computer Vision for biomedical images
analysis. However most the techniques we will discuss also apply to
all kinds of images:
- Image denoising: Linear and non-linear filtering. Wiener
filtering. Anisotropic diffusion. Non-local image
denoising. Deconvolution. Variational methods.
- Image segmentation: Global methods, region-based methods,
contour-based methods.
- 2D tracking: Recursive Bayesian filtering, Batch probabilistic
methods, non-probabilistic methods.
- Inference on graphical models: Belief Propagation and Graph-cuts.
- Data clustering: Expectation-Maximization and Mean-shift.
Lectures:
Assignments:
- Assignment #1 (please send a short report and the source code,
in the language you want
to Vincent Lepetit):
Implement the Wiener filtering method: Build a training set of
512x512 images by downloading images from the Web. Estimate the
variances for each pair of frequencies from the Fourier transforms
of these images. As shown during the lecture, this can be used to
compute the Fourier transform of the Wiener filter. Compute the
spatial version of this filter by applying the inverse Fourier
transform. Test it by convolving it with noisy images.
- Assignment #2 (please send a short report and the source code, in
the language you want
to Vincent Lepetit):
The goal of this assignment is to implement the algorithm given
Slide #86 of
the lecture
notes for the second session. Some matlab code is
provided here. You have to
complete the main file learn_sparse_features.m with the code for
the gradient descent on the D matrix, and the compute_t.m file.
The compute_t function must optimize the objective function over
t, given a training image patch o and the current D matrix. For
that, you can use the simple algorithm given Slide 81. Check
that the columns of Matrix D converge to patches that look like
gradient detectors (the convergence may take some time).
More details can be found
in Sparse
Coding with an Overcomplete Basis Set: A Strategy Employed by
V1? by Olshausen and Field for example.
- Assignment
#3 (unzip the file, open the html file, and follow the
instructions).
-
Assignment #4.
-
Assignment #5. Skeleton code and videos sequences [1, 2, 3].
-
Assignment #6. Supplemental code and images. Send your report to Mario Christoudias. Deadline: postponed to December 23th.
Required prior knowledge:
The students are expected to have already followed an introductory
Computer Vision class such as the "Introduction to Computer Vision"
Master's course and to be familiar with basic Computer Vision
concepts. A solid background in both programming and mathematics is
also a requirement for this class.
|
