Computer's Sights

Previous issues

INITIATIVES
Academic Commandos

RESEARCH
Open Research Day
Dealing With Stress, the Natural Way
Setting the Computer's Sights
Democracy is Not the Key to Peace in the Middle East
Islam and Christianity
Punctuating Mozart
Reshaping Distance Education

TECHNOLOGY
E-reader
Hybrid Books

GRADUATES
Prof. Gidi Gross: Cancer Researcher
Israel Adam: From Slum Alcoholic to Honors Student and a Master's Degree
All Doors Opened to Elishai Ezra
The Open University as His Greatest Challenge in Life: Sharli Afanger

STUDENTS
Walking in Their Footsteps in Ashdod
University Student Since 8th Grade

MILESTONES
Prof. Hagit Messer-Yaron
Prof. Ron Weidberg
Dr. Yuval Eylon
Dr. Adia Mendelsohn-Maoz
Prof. Aviva Halamish
Dr. Ophir Munz-Manor
Dr. Baruch Ziv
Dr. Milly Perry
Dr. Yoav Yair
SuperBrand University

EVENTS
Dance with Your Fears
Heavens Above

RESEARCH
Setting the Computer's Sights ...continued from previous page. From Artificial Intelligence to Security The evolution of the multidisciplinary science of computer vision can be traced back to AI (Artificial Intelligence), when scientists were interested in extracting information from images. Computer vision is an extremely interdisciplinary field with close relations to many other fields in computer science and indeed science in general. According to Tal, "computer vision is one of the biggest fields of research in computer sciences both for academic and commercial reasons. Indeed, the amounts of money that are being invested in this field are staggering and whichever came first – whether the interest or the moneys invested in research – the field has experienced unbelievable growth." In actuality, we have all encountered computer vision applications in our daily lives, although we may not categorize them as such – MRI's, X-ray tomography, computer games, facebook, and digital cameras. The military is a major user of computer vision (missile guidance, detection of enemy soldiers, transportation of men and materiel, etc.), and with the many and diverse security needs computer vision researchers have been catapulted to the forefront of the scientific field. Matching Image Pairs Dr. Tal Hassner and his colleagues from Tel Aviv University jumped into the crowded fray to address the following problem: are these two images, both of whom the computer has not seen before, the same or not? Think about the border crossing scenario. Or the thief absconding with the goods. And before continuing, consider this interesting anecdote to give you a sense of the difficulty in answering this question. About two years ago, a large scale government-funded computer vision trial research project was conducted in England. Researchers wrote an algorithm to help determine whether the person standing at the border is really the person in the passport photo. The standard sensitivity sent about 70% of the people for verification by human. And, by slightly lowering the sensitivity level, the computer ended up identifying Wynona Rider and Osama Ben Laden as the same person. Tal and his colleagues designed a new set of algorithms that were providing some outstanding results when measured against the international benchmarking system developed by the University of Massachusetts. Commonly known as the Labeled Faces in the Wild (LFW) image sets, researchers are given ten sets of image pairs, with each set containing 600 image pairs, 300 are the same and 300 are not. You perform the test ten separate times and calculate the standard deviation. The state of the art result in 2008 was 74%. Then Tal’s team came along. "We couldn’t beat that benchmark in the beginning," Tal reports. "When we first published we had reached 78%, but were soon beaten." Touche Unwilling to throw in the towel, they came back. "We developed new techniques which improved performance even more – our most recent results were better than 86%. Right now, we are considered state-of-the-art." A small Israeli start-up, face.com is basing their successful facebook application on the fruits of the research made by Tal's team. These most recent results are approaching the capabilities of face recognition by humans. "Humans can't make it to 100%" Tal elaborates, "we fall somewhat short of that, but the gap between human vision and computer vision is being narrowed." Currently, they are working on another test where more information is available and results are so far approaching 90%. Achieving that goal will make computer systems far more efficient and reliable in recognizing people's faces in general images – better than currently, and hopefully better than humans. Page: 1 2 3

RESEARCH

Setting the Computer's Sights

...continued from previous page.

From Artificial Intelligence to Security

The evolution of the multidisciplinary science of computer vision can be traced back to AI (Artificial Intelligence), when scientists were interested in extracting information from images. Computer vision is an extremely interdisciplinary field with close relations to many other fields in computer science and indeed science in general.

According to Tal, "computer vision is one of the biggest fields of research in computer sciences both for academic and commercial reasons. Indeed, the amounts of money that are being invested in this field are staggering and whichever came first – whether the interest or the moneys invested in research – the field has experienced unbelievable growth."

In actuality, we have all encountered computer vision applications in our daily lives, although we may not categorize them as such – MRI's, X-ray tomography, computer games, facebook, and digital cameras.

The military is a major user of computer vision (missile guidance, detection of enemy soldiers, transportation of men and materiel, etc.), and with the many and diverse security needs computer vision researchers have been catapulted to the forefront of the scientific field.

Matching Image Pairs

Dr. Tal Hassner and his colleagues from Tel Aviv University jumped into the crowded fray to address the following problem: are these two images, both of whom the computer has not seen before, the same or not?

Think about the border crossing scenario. Or the thief absconding with the goods.

And before continuing, consider this interesting anecdote to give you a sense of the difficulty in answering this question. About two years ago, a large scale government-funded computer vision trial research project was conducted in England. Researchers wrote an algorithm to help determine whether the person standing at the border is really the person in the passport photo. The standard sensitivity sent about 70% of the people for verification by human. And, by slightly lowering the sensitivity level, the computer ended up identifying Wynona Rider and Osama Ben Laden as the same person.

Tal and his colleagues designed a new set of algorithms that were providing some outstanding results when measured against the international benchmarking system developed by the University of Massachusetts. Commonly known as the Labeled Faces in the Wild (LFW) image sets, researchers are given ten sets of image pairs, with each set containing 600 image pairs, 300 are the same and 300 are not. You perform the test ten separate times and calculate the standard deviation.

The state of the art result in 2008 was 74%. Then Tal’s team came along. "We couldn’t beat that benchmark in the beginning," Tal reports. "When we first published we had reached 78%, but were soon beaten."

Touche

Unwilling to throw in the towel, they came back. "We developed new techniques which improved performance even more – our most recent results were better than 86%. Right now, we are considered state-of-the-art."

A small Israeli start-up, face.com is basing their successful facebook application on the fruits of the research made by Tal's team.

These most recent results are approaching the capabilities of face recognition by humans. "Humans can't make it to 100%" Tal elaborates, "we fall somewhat short of that, but the gap between human vision and computer vision is being narrowed."

Currently, they are working on another test where more information is available and results are so far approaching 90%.

Achieving that goal will make computer systems far more efficient and reliable in recognizing people's faces in general images – better than currently, and hopefully better than humans.

Page: 1 2 3