Deciphering the Faceclopinet.com/isabelle/Projects/CVPR2011/slides/aleix.pdfAleix M. Martinez...
38
Deciphering the Face Deciphering the Face Aleix M. Martinez Computational Biology Computational Biology and Cognitive Science Lab li@ d aleix@ece.osu.edu
Transcript of Deciphering the Faceclopinet.com/isabelle/Projects/CVPR2011/slides/aleix.pdfAleix M. Martinez...
Deciphering the FaceDeciphering the Face
Aleix M. MartinezComputational BiologyComputational Biology
and Cognitive Science Labl i @ [email protected]
Human-ComputerPoliticsInteraction
HumanHumanfacefaceArt
Sign Language
faceface
Language
CognitiveCognitiveScience Computer Vision
Models of Face Perception
• Features: Shape vs. texture. ……
• 2D vs. 3D
• Form of the computational space:p p
Continuous vs. Categorical
What we are going to show
• What is the form of the computational space in human face perception? Hybrid approach:in human face perception? Hybrid approach: Linear combination of continuous representations of categoriesrepresentations of categories.
+ c2c1 + … + cn
• What are the dimensions? Mostly configural.
21 n
• In computer vision we need precise detailed detection of faces and facial features.detect o o aces a d ac a eatu es.
Identity
Same or different?
Identity
Same or different?
Identity
Same or different?Identity, expression, gender, etc.
Dimensions of the Face Space
Same or different?Configural processing
Form of the Computational Face SpaceComputational Face Space
Exemplar-based modelExemplar based model
Exemplar cells …
Norm-based modelMid-level
cells
vision
Low-level vision
Facial Expressions of Emotion
Muscle Positions Model
Muscle Positions Model
• Global shape (bone structure)determines identity – configural.y g
• But ONLY muscles are responsiblefor expression interactionfor expression, interaction …
Configural Processing
Emotion perception inl femotionless faces
NeutralNeutralAngry
Sad
Neth & Martinez, JOV, 2009.
Stimuli
25%
50% 100%
75%
Neth & Martinez, JOV, 2009.
ExperimentExperiment
Less, same, more.
Configural Processing
Sad
*
**
** *
*80
90
**
50
60
70
LessSame
*
* ** * *
*
20
30
40More
0
10
-100% -75% -50% -25% 0% 25% 50% 75% 100%
Neth & Martinez, JOV, 2009.
Configural Processing
Angry
* **
* *80
90
*
**
*
50
60
70
LessSame
***
* *
***20
30
40More
0
10
-100% -75% -50% -25% 0% 25% 50% 75% 100%
Neth & Martinez, JOV, 2009.
Norm-based Face SpaceSadness
MultidimensionalS 75%
100%
Face Space
- density+ density
50%
75%
+ density 25%
Easier
+ density- density
MEAN
density100%
More difficult
AngerNeth & Martinez, JOV, 2009.
Configural Processing
Neth & Martinez, JOV, 2009.
Computational Space
Neth & Martinez, Vision Research, 2010
Computational Space
Thinner faceThinner face
Wider face
Neth & Martinez, Vision Research, 2010
American Gothic Illusion
Neth & Martinez, Vision Research, 2010
Why Configural Features?
15 x 10 pixels
Why Configural cues?
sad neutral angry
Neth & Martinez, Vision Research, 2010; Du & Martinez, 2011
Proposed Hybrid Model:Recognizing other emotion labelsRecognizing other emotion labels
+ cc + + c+ c2c1 + … + cn
Happily Angrily surprised
g ysurprised
Martinez, CVPR, 2011
Configural Processing = Precise detection of facial featuresdetection of facial features
4 2 pixels3,930
images
4.2 pixelserror
(1.5%)(1.5%)
Ding & Martinez, PAMI, 2010
Face Detection
Features VS contextObservation: Most detections are near the correct location – they are not incorrect, they are imprecise.location they are not incorrect, they are imprecise.
Key idea: Use context information to train where nott d t t f d f i l f t
Ding & Martinez, CVPR, 2008; PAMI, 2010
to detect faces and facial features.
Features VS contextObservation: Most detections are near the correct location – they are not incorrect, they are imprecise.location they are not incorrect, they are imprecise.
Key idea: Use context information to train where nott d t t f d f i l f tto detect faces and facial features.
Ding & Martinez, CVPR, 2008; PAMI, 2010
Features VS contextObservation: Most detections are near the correct location – they are not incorrect, they are imprecise.location they are not incorrect, they are imprecise.
Key idea: Use context information to train where nott d t t f d f i l f tto detect faces and facial features.
Ding & Martinez, CVPR, 2008; PAMI, 2010
Subclass Discriminant Analysisy
Between subclassBetween-subclass scatter matrix:
( ) ( )∑∑C H
Ti
Σ ( ) ( ).1 1∑∑= =
−−=i j
ijT
ijijB p μμμμΣ
Basis vectors:
.Λ= VΣVΣ XB
Basis vectors:
How many subclasses (H):Minimize the conflict, K.
Zhu & Martinez, PAMI, 2006
Precise Detailed Detection
E 6 2 i l (2%) M l 4 2 (1 5%)Error: 6.2 pixels (2%) vs Manual: 4.2 (1.5%)
Ding & Martinez, CVPR, 2008; PAMI, 2010
Detection + non-rigid SfM
Gotardo & Martinez, PAMI, 2011; Gotardo & Martinez, CVPR, 2011.
36
Take Home Messages
• What is the form of the computational space in human face perception? Linear combination of known categories.
+ c2c1 + … + cn
Wh t th di i ? M tl fi l
21 n
• What are the dimensions? Mostly configural.• Precise detection of facial features.
CBCSL
Paulo Gotardo, Shichuan Du, Don Neth, Liya Ding, OnurPaulo Gotardo, Shichuan Du, Don Neth, Liya Ding, Onur Hamsici, Samuel Rivera, Fabian Benitez, Hongjun Jia, Di You.
National Institutes of HealthNational Institutes of HealthNational Science Foundation
