Perceptual Errors in Object Recognition are Reduced by the...
Transcript of Perceptual Errors in Object Recognition are Reduced by the...
Perceptual Errors in ObjectPerceptual Errors in ObjectRecognition are ReducedRecognition are Reduced
by the Presence of Contextby the Presence of ContextObjectsObjects
M. AucklandM. Auckland11, K. Cave, K. Cave22,,N. DonnellyN. Donnelly11 and F. Gomes-Pinto and F. Gomes-Pinto11
11University of SouthamptonUniversity of Southampton22University of MassachusettsUniversity of Massachusetts
Explanations for Contextual EffectsExplanations for Contextual Effects
nn Description EnhancementDescription Enhancement•• Biederman Biederman (1981)(1981)•• Context activates schema, which facilitates theContext activates schema, which facilitates the
perceptual analysis of objectsperceptual analysis of objects
nn Criterion ModulationCriterion Modulation•• Palmer (1975), Friedman (1979), Palmer (1975), Friedman (1979), Kosslyn Kosslyn (1994)(1994)•• Context activates schema, which lowers theContext activates schema, which lowers the
activation threshold of related objectsactivation threshold of related objects
nn Functional IsolationFunctional Isolation•• Hollingworth Hollingworth & Henderson (1998, 1999)& Henderson (1998, 1999)•• Semantic and perceptual processes isolatedSemantic and perceptual processes isolated•• Effects explained by response biasEffects explained by response bias
Explanations for Contextual EffectsExplanations for Contextual Effects
nn Hollingworth Hollingworth & Henderson (1998, 1999)& Henderson (1998, 1999)provide strong evidence against Descriptionprovide strong evidence against DescriptionEnhancementEnhancement•• Over-estimation of sensitivity when calculatingOver-estimation of sensitivity when calculating d d’’ in in
Biederman Biederman (1981)(1981)•• Two alternative forced choice experiment toTwo alternative forced choice experiment to
remove response biasremove response bias
nn Nature of forced-choice experiments did notNature of forced-choice experiments did notprovide such strong evidence againstprovide such strong evidence againstCriterion ModulationCriterion Modulation
Aim of ExperimentAim of Experiment
nn This study attempted to use an alternativeThis study attempted to use an alternativemethod of removing response bias to assessmethod of removing response bias to assessCriterion ModulationCriterion Modulation
nn Also tested the effect of seeing the contextAlso tested the effect of seeing the contextslightly before the targetslightly before the target
New ParadigmNew Paradigm
nn 60 participants60 participantsnn Cue to control Cue to control attentional attentional orientationorientation
1000ms
Cue
New ParadigmNew Paradigm
nn Picture stimulusPicture stimulus•• Centrally placed targetCentrally placed target•• Four objects placed around target to form contextFour objects placed around target to form context
In-context Out of context
New ParadigmNew Paradigm
nn SOA of context to target was manipulated soSOA of context to target was manipulated socontext was seen:context was seen:•• Simultaneously with target (SOA = 0ms)Simultaneously with target (SOA = 0ms)•• Or before target (SOA = -52ms & -104ms)Or before target (SOA = -52ms & -104ms)
nn Display time was 52ms for all context andDisplay time was 52ms for all context andtarget imagestarget images
1000ms
Cue
52ms
Picture Stimulus (SOA = 0ms)
1000ms
Cue
52ms52ms
52ms
Picture Stimulus (SOA -104ms)
New ParadigmNew Paradigm
nn Target offset was masked (1000ms)Target offset was masked (1000ms)nn Six alternative forced choice response (errorSix alternative forced choice response (error
rates)rates)•• Allows differentiation between number of perceptualAllows differentiation between number of perceptual
errors and number of semantic errorserrors and number of semantic errors
1000ms
Cue
52ms
Picture Stimulus
1000ms
Mask
Self-Paced
6AFC
New ParadigmNew Paradigm
nn Six alternative forced choice responseSix alternative forced choice response1.1. TargetTarget e.g. Cardse.g. Cards2.2. Physically similar to targetPhysically similar to target e.g. Paper Fane.g. Paper Fan3.3. Semantically related to contextSemantically related to context e.g. Roulette wheel (in)e.g. Roulette wheel (in)
e.g. Banana (out)e.g. Banana (out)
•• Possible guessing strategyPossible guessing strategynn Participant can deduce contextParticipant can deduce contextnn Guess has 1:2 chance for in-context trials but 1:3Guess has 1:2 chance for in-context trials but 1:3
chance for out of context trialschance for out of context trials
New ParadigmNew Paradigm
nn Six alternative forced choice responseSix alternative forced choice response1.1. TargetTarget e.g. Cardse.g. Cards2.2. Physically similar to targetPhysically similar to target e.g. Paper Fane.g. Paper Fan3.3. Semantically related to contextSemantically related to context e.g. Roulette wheel (in)e.g. Roulette wheel (in)
e.g. Banana (out)e.g. Banana (out)
4.4. Physically similar to choice 3Physically similar to choice 3 e.g. Tyre (in)e.g. Tyre (in)e.g. Horn (out)e.g. Horn (out)
5.5. Semantically related to choice 2Semantically related to choice 2 e.g. Chopsticks (out)e.g. Chopsticks (out)No relationship No relationship e.g. Mobile phone (in)e.g. Mobile phone (in)
6.6. Physically similar to choice 5Physically similar to choice 5 e.g. Straws (out)e.g. Straws (out)e.g. Calculator (in)e.g. Calculator (in)
Guessing correction 1Guessing correction 1
nn Picking one of two contextually relatedPicking one of two contextually relatedchoices?choices?
In-contextIn-context Out of contextOut of contextTarget:Target: Cards Cards –– G1 G1 CardsCardsPhyMatchPhyMatch:: Paper FanPaper Fan Paper Fan Paper Fan –– G1 G1ConMatchConMatch:: Roulette Wheel Roulette Wheel –– G1 G1 BananaBananaEType3:EType3: TyreTyre HornHornEType4:EType4: Mobile PhoneMobile Phone Chopsticks Chopsticks –– G1 G1EType5:EType5: CalculatorCalculator StrawsStraws
G1 = EType4[Out] G1 = EType4[Out] –– (EType3[Out] + EType5[Out])/2 (EType3[Out] + EType5[Out])/2
PerceptuallySimilar Pairs
SemanticallyRelated Pairs
Guessing correction 2Guessing correction 2
nn Guessing based on context alone? (ResponseGuessing based on context alone? (Responsebias correction)bias correction)
In-contextIn-context Out of contextOut of contextTarget:Target: Cards Cards –– G2/2 G2/2 CardsCardsPhyMatchPhyMatch:: Paper FanPaper Fan Paper FanPaper FanConMatchConMatch:: Roulette Wheel Roulette Wheel –– G2/2 G2/2 Banana Banana –– G2 G2EType3:EType3: TyreTyre HornHornEType4:EType4: Mobile PhoneMobile Phone ChopsticksChopsticksEType5:EType5: CalculatorCalculator StrawsStraws
G2 = G2 = ConMatchConMatch[Out] [Out] –– (EType3[Out] + EType5[Out])/2 (EType3[Out] + EType5[Out])/2
PerceptuallySimilar Pairs
SemanticallyRelated Pairs
Results Results –– Error Rates Error Rates
Repeated measures ANOVARepeated measures ANOVAnn Main effect of contextMain effect of context
•• Uncorrected:Uncorrected: F(1, 48) = 20.979, p < 0.01F(1, 48) = 20.979, p < 0.01•• Corrected:Corrected: F(1, 48) = 8.407, p < 0.01F(1, 48) = 8.407, p < 0.01
nn Main effect of SOAMain effect of SOA•• Uncorrected:Uncorrected: F(2, 96) = 60.874, p < 0.01F(2, 96) = 60.874, p < 0.01•• Corrected:Corrected: F(2, 96) = 3.980, p < 0.05F(2, 96) = 3.980, p < 0.05
Results Results –– Error Rates Error Rates
nn Context x SOA interactionContext x SOA interaction•• Uncorrected:Uncorrected: F(2, 96) = 5.888, p < 0.01F(2, 96) = 5.888, p < 0.01•• Corrected:Corrected: not significantnot significant
Uncorrected line graph data Corrected line graph data
Uncorrected Context X SOA
20
25
30
35
40
45
50
In OutContext
Erro
r Rat
e (%
)
0 ms
-52 ms
-104 ms
Corrected Context X SOA
20
25
30
35
40
45
50
In OutContext
Erro
r Rat
e (%
)0 ms
-52 ms
-104 ms
Results Results –– Broken down by Error TypeBroken down by Error Type
Repeated measures ANOVARepeated measures ANOVAnn Main effect of error typeMain effect of error type
•• Uncorrected:Uncorrected: F(4, 192) = 38.177, p < 0.01F(4, 192) = 38.177, p < 0.01•• Corrected:Corrected: F(4, 192) = 61.381, p < 0.01F(4, 192) = 61.381, p < 0.01
nn SOA x error type interactionSOA x error type interaction•• Uncorrected: Uncorrected: F(8, 384) = 5.926, p < 0.01F(8, 384) = 5.926, p < 0.01•• Corrected:Corrected: not significantnot significant
nn Planned comparisons: (corrected)Planned comparisons: (corrected)•• Perceptual errors: Perceptual errors: F(1, 48) = 4.994, p < 0.05F(1, 48) = 4.994, p < 0.05•• Semantic errors:Semantic errors: not significantnot significant
Results Results –– Broken down by Error TypeBroken down by Error Type
nn Effect of response bias correction onEffect of response bias correction onsemantic errorssemantic errors
Corrected SOA x Error Type
0
2
4
6
8
10
12
14
0 ms -52 ms -104 msSOA
PerceptualSemantic345
Error Type
Uncorrected SOA x Error Type
0
2
4
6
8
10
12
14
0 ms -52 ms -104 msSOA
PerceptualSemantic345
Error Type
Results Results –– Broken down by Error TypeBroken down by Error Type
nn Fewer perceptual errors in in-context conditionFewer perceptual errors in in-context condition
Corrected Context x Error Type
02468
101214
In OutContext
Erro
r Rat
e (%
)
PerceptualSemantic345
ConclusionsConclusions
nn Eliminating response bias does removeEliminating response bias does removesome of the context effectsome of the context effect
nn A context consistent to the target alsoA context consistent to the target alsoincreases sensitivity in recognition tasksincreases sensitivity in recognition tasks
nn Combination of perceptual analysis andCombination of perceptual analysis andcontextual information reduces perceptualcontextual information reduces perceptualerrorserrors
nn Consistent with Criterion Modulation notConsistent with Criterion Modulation notFunctional IsolationFunctional Isolation