Narayanan Sundaram. Conference stats ~300 papers 40 orals Rest were posters ~700 people ~50% from Europe ~33% from U.S & North America Rest from Asia.

Update on ECCVSeptember 5-11, Greece

Narayanan Sundaram

Conference stats~300 papers

40 oralsRest were posters

~700 people~50% from Europe~33% from U.S & North AmericaRest from Asia & Oceania

4 days main conference1 day tutorials2 day workshops

Some interesting trendsConvergence of vision and graphics (image

composition, restoration of old video tapes etc.)Structure from motion for elastic surfaces (mostly

faces)Facial expression/beauty analysis/synthesisUnderstanding volumes (buildings/indoor scenes etc)Object recognition/segmentation (lot of papers use

Markov random fields in some way or the other)Recognition with humans in the loopConvex optimization for vision problems (avoiding the

initialization problem)

Session on “Vision & Industry”Talks from

IBMGESiemensLockheed MartinAdobeDxO Labs (France)

Where has CV worked and where has it failed?

What are the opportunities and threats?

Visual Analytics for Shrinking Checkout Shrink (Sharathchandra Pankanti - IBM)CV from security cameras for checking retail fraud

~$10/lane/day (~10’s of billions of $ lost in US & Europe every year)

50% due to dishonest employees/cashiersFraud in self-checkouts, various types of cashier fraudUsual action sequence is Pickup->Scan->Drop

Huge variations (coupons, returns mid-way, barcode errors etc)Constrained viterbi algorithm for optimal alignment recognitionVisual Verification Service (VVS) sold a product

ChallengesOnline learning (list of products, barcodes keeps changing)20-30,000 unique barcodes/storeNeed days worth of training dataDifferentiating very similar looking products (store brand vs name

brand)

GE's Computer Vision: from Prisons to Healthcare (Peter Tu)Projects

Shopping: mostly tracking shoppersDHS : tracking people at stadium (>2000

people/hour)Mock prison riots: people tracking + group dynamicsAction recognition : crouching, throwing etc.Face recognition: Under unreliable environmentsSocial network analysis: Identify leaders in groups“People as sensors”“FBI : “Face from skull”Pathology: Cell counting

GE (contd)Main question : “Is there enough Return on

investment for customers from computer vision?”Not much at the momentMostly added benefits, not primary benefits

E.g. video analytics industry makes as much as the septic tank industry

2 death sentencesHigh false alarm rateNeed a PhD to install the product and calibrate the

setupProving algorithms in real sites under real

environments in real time will make a difference

Visual Cognition, Semantic & Quantitative Imaging (Ramesh Visvanathan - Siemens)Mostly in medical and industrial imaging

Cardiac segmentation (4D)User guided segmentation of other organsGas turbines – defect modeling + analysisPeople – queue management – How long does it take

for a person to go through airport security?Challenges

How do we represent and update knowledge systematically?

What guarantees can we give?Do not want to do what A.I did in the 80’s

(overpromise and under deliver)

Siemens (contd)Video analytics

Low end is already commoditized (badly)Algorithms are not robust at all – no guaranteesPrices down to ~$100/licenseOverpromising customers may lead to a

backlash against all commodity CV productsHigh end video analytics

Require very robust algorithms<0.1% false alarms>95% true detection at 120 meters

Applications of Computer Vision in Airborne Surveillance (Michael E. Bazakos – Lockheed Martin)“Need capabilities, not technologies”

Need timely and actionable information from data

Unless false alarm rate goes down, pilots will just switch off the CV

Specify the conditions under which the algorithm will work Daylight only (or) x pixels on objects (or) upto

n degrees of gaze etc.

From PostScript to face detectors: How computer vision is transforming Adobe (Lubomir Bourdev)Adobe doing more and more vision and graphics researchAdobe has transitioned from

Low level -> High level analysisPixel level -> Cross-image content

ExamplesRed-eye removalPanoramic compositionTagging peopleSelecting objects

ChallengesAlgorithms are not robustUser Interface issues (how to get feedback from users)

Information capacity: a measure of potential image quality of a digital camera (Frederic Guichard – DxO labs)IP licensing of IC design, optics design &

software“Image processing at capture level”Image processing on cameras is huge 2008-

2010Processing in easy (Moore’s law)Lens+sensor – costly and slow

Camera phones already at photonic/diffraction/thermal noise limitsSLRs are getting there

Computational imaging is the future

DxO labs (contd)Future:

Improve digital calibration Takes 20 sec – 1 min for calibrating a single iPhone

camera Have to use on-the-fly correction Get rid of calibration

New functions for photographers (several shots /several cameras) Extended DOF (autofocus is costly and bulky) DxO’s solution based on different focus depths for

RGB channels and correct from there

Greece was amazing!