| |
|
|
09:15 - 09:45
|
Refreshments and tagging
|
| |
|
|
09:45 - 10:00
|
Greetings :
Ohad Ben-Shahar, Department of Computer Science
Prof. Rivka Carmi, BGU President
|
| |
|
|
10:00 - 10:50
|
Shimon Ullman,
Weizmann Institute
Recognizing objects and natural classes
|
|
|
I will describe a computational scheme which learns to recognize new
object categories from image examples. Starting with a collection of
images illustrating examples of a category such as 'face', 'car', or
'horse', the model constructs a category representation, and then uses
it to identify novel members of the category. The approach is based on
representing shapes within a category by a hierarchy of shared
sub-structures called fragments, selected by maximizing the information
delivered for categorization. I will present results of applying the
model to natural object categories, and discuss relationships between
the model and parts of the primate visual system involved in object
perception.
|
| |
|
|
10:50 - 11:40
|
Its'hak Dinstein,
Ben-Gurion University
Binarization, skew detection, character extraction, and writer identification
in historical Hebrew calligraphy documents
|
|
|
We present our work on the paleographic analysis and recognition system intended for
processing of historical Hebrew calligraphy documents. One of the goals is to analyze
documents of different writing styles in order to identify the locations, dates,
and writers of test documents. We discuss our approach to binarization, skew detection,
and writer classification. Results of automatic extraction of pre-specified letters using the erosion
transform are presented. We further propose and test topological features for handwriting
style classification based on a selected subset of the Hebrew alphabet.
Joint work with Klara Kedem, Itay Bar-Yoseph, Amir Egozi, and Isaac Beckman, Ben Gurion University.
|
| |
|
|
11:40 - 12:30
|
Freddy Bruckstein,
Technion
On variational methods for image analysis
|
|
|
This talk will focus on variational methods in signal and image
processing raising questions on the choice of functionals that are
optimized for various tasks. It is argued that the selection of
functionals is more of an art than a science. Several examples will be
discussed from denoising to the Estimation of optic flow in video
streams.
|
| |
|
|
12:30 - 13:30
|
Lunch
|
| |
|
|
13:30 - 13:40
|
Greetings :
Prof. Moti Hershkowitz, Vice-President and Dean for Research and Development
|
| |
|
|
13:40 - 14:30
|
Nir Sochen,
Tel Aviv University
Segmentation from Descartes to Kant
|
|
|
We will treat the segmentation problem as an epistemological question
and hopefully illustrate complicated philosophical arguments with "simple"
segmentation problems. In particular we will discuss
texture segmentation, prior-shape segmentation and dynamic labeling approaches
via the variational and level-set frameworks.
|
| |
|
|
14:30 - 15:20
|
Ohad Ben-Shahar,
Ben-Gurion University
Does it make any sense to move your head when you look at a mirror?
|
|
|
The image of a curved, specular (mirror-like) surface is a distorted reflection
of the environment. Although the recovery of such specular shape from its
image appears futile without some knowledge of the environment, the
goal of this work is to develop a theoretical and practical framework
for solving this shape inference problem when the environment is
completely unknown. We show that although this general problem is
severely ill-posed, allowing relative object-environment motion induces
a dense specular flow in the image plane which can be
related to surface shape through a pair of coupled non-linear partial
differential equations that are independent on the environment content.
We examine the qualitative and geometric properties of these equations and
present analytic and numerical methods for recovery of specular shape in several cases.
Joint work with Yair Adato, Ben Gurion University, and Yuriv Vasilyev and Todd Zickler, Harvard University.
|
| |
|
|
15:20 - 15:40
|
Coffee break
|
| |
|
|
15:40 - 16:30
|
Amnon Shashua,
Hebrew University
The Role of multi-linear constrained Factorization in Image Coding, Clustering, and Visual Recognition
|
|
|
I will present a bird's eye view of novel connections between the task of
factorizing measurements arranged in multi-way arrays (tensors in general)
under certain convex constraints to core problems in Learning and Visual processing.
A general low-rank factorization under simplex constraints corresponds to a latent
class model solution, and a super-symmetric factorization corresponds to the general problem of achieving a
soft clustering assignment over hypergraph representations. A mixture of the general
and super-symmetric arrays corresponds to a "latent clustering" model in which the value
of the hidden variable determines the pairwise affinities between data points.
The framework has been applied to areas of image coding, multi-body segmentation,
"bag of words" inference tasks, and multi-class visual recognition using shared fragments.
|
| |
|
|
16:30 - 17:20
|
Daniel Keren,
Haifa University
Image Detection without Negative Examples
|
|
|
In a typical image detection system, a classifier
(e.g. neural net or support vector machine) is supplied
with positive and negative examples, and it constructs
a classification rule using these examples.
In this talk, I will describe a simple method to
replace the negative examples with a probability
distribution. In some cases, notably when the number
of positive examples is small, it performs better
than the standard classifier.
Joint work with Rita Osadchy and Tali Buchnik, Haifa University.
|
| |
|
 |
|
|
