a bayesian model of visual attention - mit9.520/spring09/classes/06may2009... · 2010-01-22 ·...

A Bayesian Model of Visual  Attention

Sharat Chikkerur, Thomas Serre, Cheston Tan & Tomaso Poggio

Center for Biological and Computational Learning, MIT

Outline

• IntroductionLimitations of feed‐forward processing

Role of attention

• A computational model of attention

• ApplicationsModeling human eye‐movements 

Object recognition under clutter

Outline

• IntroductionLimitations of feed‐forward processing

Role of attention

• A computational model of attention

• ApplicationsModeling human eye‐movements 

Object recognition under clutter

Riesenhuber & Poggio 1999, 2000; Serre Kouh Cadieu Knoblich Kreiman & Poggio 2005; Serre Oliva Poggio 2007

*Modified from (Gross, 1998)

Feed‐forward processing

see also Broadbent 1952 1954; Treisman 1960; Treisman & Gelade 1see also Broadbent 1952 1954; Treisman 1960; Treisman & Gelade 1980; Duncan & Desimone 1995; Wolfe, 1997; Tsotsos and  many othe980; Duncan & Desimone 1995; Wolfe, 1997; Tsotsos and  many othersrs

Zoccolan Kouh Poggio DiCarlo 2007 Reynolds Chelazzi & 

Desimone 1999Serre Oliva Poggio 2007

Role of attention

Parallel processing  (No attention) Serial processing (With attention)

Biology of attention

Feed-fowardFeature-attention

Rao 2005; Lee & Mumford 2003Rao 2005; Lee & Mumford 2003

Attention as Bayesian inference

OO

FiFi

FliFli

II

LL

NN

LIP/FEFLIP/FEF

V2V2

V4V4

ITIT

PFCPFC

Spatial attention

•We use a Bayesian framework to model 

the interaction between the ventral 

stream and LIP/FEF

•Feed forward connections

within the 

ventral stream are modeled as bottom‐up

evidence.  Feedback

connections from 

higher areas are modeled as top‐down

priors.

•The posterior probability of location 

generates a task‐based saliency map.

Model description

Fi

Fil

L

N

“What”“Where”

Feature-maps

Image

Feature-maps

Model properties: invariance

Fi

Fil

L

N

“What”“Where”

Model properties: crowding

Fi

Fil

L

N

“What”“Where”

Model: spatial attention

Fi

Fil

L

N

What is at location X?

X

* * *

Model: feature‐based attention

Fi

Fil

L

N

Where is object X?

X

* * *

Outline

• IntroductionLimitations of feed‐forward processing

Role of attention

• A computational model of attention

• ApplicationsModeling human eye‐movements 

Object recognition under clutter

Matching human eye‐movements

Car search Pedestrian search

Dataset100 CBCL street-scenes images having cars & pedestrians20 images with neither objects

Experiment8 subjects where shown these 120 images in random order. Each image in the stimuli-set was presented twiceThe subjects were asked to count the number of cars/pedestrians For each of these block trials, the subject’s eye movements were recorded using an infra-red eye tracker.

(Psychophysics done by Cheston T )

Model instantiation

OOO

FiFFii

FliFFllii

III

LLL

NN

LIP/FEFLIP/FEF

V2V2

V4V4

ITIT

PFCPFC

Examples

Examples

Quantitative evaluation: ROC

0

0.25

0.5

0.75

1

car pedestrian

Humans Bottom-upTop-down (feature-based) Feaure-based + contextual cues

RO

Car

ea

Integrating (local) featureIntegrating (local) feature--based + (global) contextbased + (global) context--based based cues accounts for cues accounts for 92%92% of interof inter--subject agreement!subject agreement!

Chikkerur ,Tan Serre & Poggio (SFN Chikkerur ,Tan Serre & Poggio (SFN ‘‘09,VSS 09,VSS ‘‘09)09)

Quantitative evaluation: ROC

Outline

• IntroductionLimitations of feed‐forward processing

Role of attention

• A computational model of attention

• ApplicationsModeling human eye‐movements

Object recognition under clutter

Effect of clutter on detection

recognition without attention

recognition under attention

Scale and location prediction

Performance improves under  attention

0

1

2

3

Model Humans

perfo

rman

ce (d

perfo

rman

ce (d

’’ ))

no attentionno attention one shift of one shift of attentionattention

Tan, Tan, ChikkerurChikkerur , , SerreSerre & & PoggioPoggio (VSS (VSS ‘‘09)09)

Thank you!

Questions?