[4eyes] Fwd: ICPR12 contest on Kinect-based gesture recognition

[Steffen Gauglitz]

FYI

-------- Original Message --------
Subject: ICPR12 contest on Kinect-based gesture recognition
Date: Thu, 03 May 2012 22:18:24 +0100
From: Bob Fisher <rbf at inf.ed.ac.uk>
To: sgauglitz at cs.ucsb.edu

The ICPR kinect-based gesture recognition challenge opens on May 7
(cash prizes & more). See URL and below:

   http://gesture.chalearn.org/dissemination/icpr2012

================

ChaLearn takes gesture recognition to the crowd with Microsoft Kinect(TM)

A competition to help improve the accuracy of gesture recognition
using Microsoft Kinect(TM) motion sensor technology promises to take
man-machine interfaces to a whole new level. From controlling the
lights or thermostat in your home to flicking channels on the TV, all
it will take is a simple wave of the hand. And the same technology may
even make it possible to automatically detect more complex human
behaviors, to allow surveillance systems to sound an alarm when
someone is acting suspiciously, for example, or to send help whenever
a bedridden patient shows signs of distress.

Through its low cost 3D depth-sensing cameras, Microsoft Kinect(TM)
has already kick-started this revolution by bringing gesture
recognition into the home.  Humans can recognize new gestures after
seeing just one example (one-shot-learning). With computers though,
recognizing even well-defined gestures, such as sign language, is much
more challenging and has traditionally required thousands of training
examples to teach the software.

To see what the machines are capable of, ChaLearn launched a
competition hosted by Kaggle with prizes donated by Microsoft, in the
hope they can give the state of the art a rapid boost. The ChaLearn
team has been organizing competitions since 2003, featuring hard
problems such as discovering cause-effect relationships in data. It
has selected the young and dynamic startup Kaggle to host the gesture
challenge because Kaggle has very rapidly established a track record
for using crowdsourcing to find solutions that outperform state-of-the-
art algorithms and predictive models in a wide variety of domains
(from helping NASA build algorithms to map dark matter to helping
insurance companies improves claims prediction). And now the first
round of the gesture challenge helped narrow down the gap between
machine and human performance. Over a period of four months starting
in December 2011, 153 contestants making 573 entries have built
software systems that are capable of learning from a single training
example of a hand gesture (so-called one-shot-learning). They lowered
the error rate, starting from a baseline method making more than 50%
error to less than 10% error.

The winner of the challenge, Alfonso Nieto Castanon, used a method he
invented, which is inspired by the human vision system. He and the
second and third place winners will be awarded $5000, $3000 and $2000
respectively and get an opportunity to present their results in front
of an audience of experts at the CVPR 2012 conference in Rhode Island,
USA, in June. A demonstration competition of gesture recognition
systems using Kinect(TM) will also be held in conjunction with this
event, with similar prizes donated by Microsoft.

Now, from May 7 and until September 10, new competitors can enter
round 2 of the challenge and get a chance to close the gap with human
performance, which is under 2% error! The entrants are given a set of
examples with which to apply and test their algorithms, so that they
may improve them. Compared to round 1, they will benefit from a wealth
of resources including the fact sheets and published papers of the
participants of round 1, data annotations, and data transformations
having had success in round 1. During a four month period they will be
able to compare their system with those of other contestants, by using
it to predict gestures from a feedback sample. Throughout the
competition the evaluations of these are posted on a live leaderboard,
so participants can monitor their performance in real time. The
contestants will then have the opportunity to put their best
algorithms to the final test in an evaluation phase. Here they will be
given a few days to train their system on an entirely new set of
gestures, after which the one with the best recognition score will be
rewarded with $5000. Those coming second and third place will receive
$3000 and $2000 respectively. Similarly as in round 1, the results
will be discussed at a scientific conference (ICPR 2012, Tsukuba,
Japan, November 2012) where a demonstration competition will be held
also crowned with prizes in the same amount. Microsoft will be
evaluating successful participants in all challenge rounds for two
potential IP agreements of $100,000 each. See official challenge rules
for more details at http://gesture.chalearn.org.

The winner of the first round believes that it is possible to reach
and even beat human performance. Others will also join in the race.
According to Kaggle, that is the power of the crowd: bringing together
expert talent, sometimes from previously untapped quarters. And with
Microsoft interested in buying the intellectual property, the hope is
that the new algorithms that emerge from the contest will not only
boost accuracy but also open the doors to a whole new range of
applications. From using communicating with Kinect(TM) through sign
language or even speaking, with the algorithms interpreting what you
say by reading your lips to smart homes or using gestures to control
surgical robots.

The challenge was initiated by the US Defense Advanced Research
Projects Agency (DARPA) Deep Learning Program and is supported by the
US National Science Foundation, the European Pascal2 network of
excellence, Microsoft and Texas Instruments. Any opinions, findings,
and conclusions or recommendations expressed in this material are
those of the authors and do not necessarily reflect the views of the
sponsors and funding agencies.

Isabelle Guyon, Vassilis Athitsos


-- 
The University of Edinburgh is a charitable body, registered in
Scotland, with registration number SC005336.