PowerPoint Presentation
Learning to PerceiveCoherent ObjectsNimrod Dorfman, Daniel Harari, Shimon UllmanWeizmann Institute of Science
WEIZMANNINSTITUTEOF SCIENCE
COGSCI 2013
Object segregation is learned25 monthsEven basic Gestalt cues are initially missing[Schmidt et al. 1986]Object segregation is learned
3Adults34How do we learnto segregate objects?5We propose a computational model:Explain the first steps of learningBased on psychophysical findingsComputationally tested on videosIt all begins with motion6MotionIt all begins with motionGrouping by common motionprecedes figural goodness[Spelke 1990 - review]
Motion discontinuities provide an early cuefor occlusion boundaries[Granrud et al. 1984]7
Our model8Static segregationLocal occlusion boundariesObject formMotion discontinuitiesCommon motionBoundary
GeneralAccurateNoisyIncompleteGlobal
Object-specificCompleteInaccurateMotion-based segregationIntensity edges?9
BoundaryOcclusion cues10
Extremal edgesConvexityT-junctions[Ghose & Palmer 2010]
Boundary10Familiar object11
Global12How does it actually work?
Moving object13MotionMoving object14FigureGroundUnknown
Motion15
MotionBoundaryGlobal
16Need many examples for good results (1000+)BoundaryGood examplesPrediction17
FigureorGround?FigureorGround?Novel object, novel background78% successUsing 100,000 training examplesBoundary17Entire image
18BoundaryFigureBackground18Learning an object19
Standard object recognition algorithmLearns local features and their relative locationsGlobal
Detection
20GlobalCombining information sources
21CombinedBoundaryAccurateNoisy & IncompleteGlobalComplete Inaccurate21More complex algorithms
Default GrabCutWith segregation cue[Rother et al. 2004]2222SummaryStatic segregation is learned from motionTwo simple mechanisms:BoundaryMotion discontinuities Occlusion boundaries(Need a rich library, including extremal edges)GlobalCommon motion Object formThese mechanisms work in synergyThis is enough to get started,adult segregation is much more complex2324ThankYou!
