2016 Israel Computer Vision Day
Sunday, December 25, 2016
Vision Day Schedule
Time 
Speaker
and Collaborators 
Affiliation 
Title 
08:5009:20 
Gathering 

09:2009:40 
Gerard Medioni Tal Hassner 
USC 
Faces, deep learning and the pursuit of
training data 
9:4510:05 
Elad
Richardson Roy
OrEl Ron
Kimmel 
Technion 

10:1010:30 
Margarita Osadchy Stuart Gibson Orr Dunkelman Daniel PerezCabo 
Haifa U of Kent U of Vigo 

10:3510:55 
Nadav Cohen Amnon Shashua 

Inductive Bias of Deep Convolutional Networks through Pooling
Geometry 
11:0011:30 
Coffee Break 

11:3011:50 
Dan Feldman Ibrahim Jubran 
Haifa 
Lowcost and Faster Tracking Systems Using
Coresets for PoseEstimation 
11:5512:15 
Ronen Basri Soumyadip Sengupta Tal Amir Meirav Galun Tom Goldstein David Jacobs Amit Singer 
Weizmann Princeton 

12:2012:40 
Mike Werman Shmuel Peleg Ben Arzi Yoni Kesten Tavi Halperin 
HUJI 

12:4513:05 
Ehud Barnea Ohad
BenShahar 
BGU 
HighOrder Contextual
Object Detection with a Few Relevant Neighbors 
13:0514:10 
Lunch


14:1014:20 
"Intermezzo" 

14:2014:40 
Yaniv Taigman 
Facebook
AI Research (FAIR), TLV 

14:4515:05 
Micha
Lindenbaum Avi Kaplan Tamar Avraham 
Technion 
Interpreting the Ratio Criterion for Matching SIFT
Descriptors 
15:1015:30 
Or Litany Alex Bornstein Michael Bronstein 
TAU 

15:3515:55 
Itamar
Talmi Roey
Mechrez Lihi
ZelnikManor 
Technion 

16:0016:20 
Coffee
Break 

16:2016:40 
Netalee Efrat Piotr Didyk Mike
Foshey Wojciech Matusik Anat Levin 
Weizmann MIT 

16:4517:05 
David Avidar Meir Barzohar 
Technion 

17:1017:30 
Yael Moses Shachaf Melman Gerard
Medioni Yinghao Cai 
IDC 
The Israel Computer Vision Day is sponsored
this year by:
Abstracts 

Gerard Medioni and Tal Hassner – USC,
OPENU 

Elad Richardson, Matan Sela, Roy OrEl and Ron Kimmel  Technion 

No Bot Expects the DeepCAPTCHA! Introducing
Immutable Adversarial Examples, with Applications to CAPTCHA Generation 
Margarita Osadchy, Julio HernandezCastro, Stuart Gibson, Orr Dunkelman Daniel PerezCabo – Haifa, U of Kent, U of Vigo 
In this work we introduce DeepCAPTCHA, a new and secure
CAPTCHA scheme based on adversarial examples, an inherit limitation of the
current Deep Learning networks. These adversarial examples are constructed visual inputs,
either synthesized from scratch or computed by adding a small and specific
perturbation called adversarial noise to correctly classified items, causing
the targeted DL network to misclassify them. We show that plain adversarial
noise is insufficient to achieve secure visual CAPTCHA schemes, which leads
us to introduce immutable adversarial noise  an adversarial noise
resistant to removal attempts. We implement
a proof of concept system, and its analysis shows that the scheme
offers high security and good usability compared to the best previously
existing CAPTCHAs. 
Inductive Bias of Deep Convolutional Networks through Pooling Geometry 
Nadav Cohen
and Amnon Shashua  HUJI 

Lowcost and Faster Tracking Systems Using Coresets for
PoseEstimation 
Dan
Feldman, Soliman Nasser and Ibrahim Jubran  Haifa 
We
prove that every set has a weighted subset (coreset) of constant size
(independent of n), such that computing the optimal orientation of the small
coreset would yield exactly the same result as using the full set of n
markers. A deterministic algorithm for computing this coreset in O(n) time
is provided, using the Caratheodory Theorem from computational geometry. We then
developed a $50 tracking system based on this algorithm that turns a toy
drone into an autonomous drone. The experimental results are almost identical
to those obtained via a commercial $10,000 tracking system (OptiTrack). 
A New Rank Constraint on Multiview Fundamental
Matrices and its Application to Camera Location Recovery 
Ronen Basri , Soumyadip Sengupta, Tal Amir,
Meirav Galun, Tom Goldstein, David Jacobs, Amit Singer – Weizmann,
UMD, Princeton 

Mike Werman, Shmuel Peleg, Ben Arzi, Yoni Kasten and Tavi Halperin  HUJI 
The fundamental matrix is the basic building block of multiple view
geometry and its computation is the first step in many vision tasks. Its
computation is usually based on pairs of corresponding points. It is known
that the fundamental matrix can also be computed from three matching epipolar
lines. This was rarely used as there were no good methods to find these
correspondences. Here we present 3 practical methods of finding such epipolar
line correspondences resulting in superior fundamental matrices. 
HighOrder Contextual Object Detection with a Few Relevant
Neighbors 
Ehud Barnea
and Ohad BenShahar  BGU 
A natural way to improve the detection of objects is to consider
contextual constraints imposed by the detections of other objects in the
scene. In this work we exploit the spatial relations between objects to
improve a given set of detections and analyze the different properties of the
problem in an exact probabilistic setting. In contrast to previous methods
that are based on various complicated assumptions but typically focus on
pairwise interactions only, here we employ a single realistic assumption that
the existence of an object at any given location is influenced by just few
other relevant locations in space in order to facilitate a more exact
calculation of object probability while using higher order interactions as
well. We suggest a method for identifying these relevant locations and integrate
them into an exact calculation of probability based on the raw detector
responses. Among other insights, we argue that while it is generally
difficult but possible to learn when an object reduces the probability of
another, in many cases it is practically impossible to do for the task of
improving the results of an object detector. We show that this also applies
to some cases where an object greatly increases the probability of another,
but that generally this occurs less than the former case. Finally, we
demonstrate that the suggested approach improves detection results more than
previous approaches over the challenging KITTI dataset. 
Yaniv
Taigman, Adam Polyak and Lior Wolf – Facebook AI Research (FAIR), TLV 

Interpreting the Ratio Criterion for Matching
SIFT Descriptors 
Micha Lindenbaum, Avi Kaplan, Tamar
Avraham  Technion 
We provide two alternative
interpretations of this criterion, which show that it is not only an
effective heuristic but can also be formally justified. The first
interpretation shows that Lowe's ratio corresponds to a conditional
probability that the match is incorrect. The second shows that the ratio
corresponds to the Markov bound on this probability. The interpretations make
it possible to slightly increase the effectiveness of the ratio criterion,
and to obtain matching performance that exceeds all previous (nonlearning
based) results. 
Or Litany and Alex Bornstein, Emanuele
Rodolà, Michael Bronstein – 

Itamar Talmi, Roey
Mechrez, Lihi ZelnikManor  Technion 

Netalee Efrat, Piotr Didyk, Mike Foshey, Wojciech Matusik,
Anat Levin – Weizmann 

David Avidar, David Malah and Meir Barzohar  Technion 
The use of 3D point clouds is currently of much interest. One of the
cornerstones of 3D point cloud research and applications is point cloud
registration. Given two point clouds, the goal of registration is aligning
them in a common coordinate system. In particular, we seek in this work to
align a sparse and noisy local point cloud, created from a single stereo pair
of images, to a dense and largescale global point cloud, representing an
urban outdoors environment. The common approach of keypointbased
registration, tends to fail due to the sparsity and low quality of the stereo
local cloud. We propose here a new approach. It consists of the creation of a
dictionary of much smaller clouds using a grid of synthetic viewpoints over
the dense global cloud. We then perform registration via an efficient
dictionary search. Our approach shows promising results on data acquired in
an urban environment.

Yael Moses, Shachaf Melman, Gerard Medioni and
Yinghao Cai – IDC, TAU, USC 
We present a joint Deep Convolutional Neural Network and Support
Vector Regression approach for estimating a person’s age from a face. We
start by leaning a robust face representation using deep
network, followed by kernelbased support vector regression. We then
show the age estimation accuracy can be further improved that by learning an
agerelated dimensionality reduction metric. The proposed schemes were
successfully applied to the MORPHII and FGNet datasets outperforming
contemporary stateoftheart approaches. 