introduction identity data set and face representation associate-predict model switching...
TRANSCRIPT
An Associate-Predict Model for Face
Recognition
CVPR 2011
Qi Yin1,3 Xiaoou Tang1,2 Jian Sun3
1. Department of Information EngineeringThe Chinese University of Hong Kong
2. Shenzhen Institutes of Advanced TechnologyChinese Academy of Sciences, China
3. Microsoft Research Asia
Outline
Introduction
Identity Data Set and Face Representation
Associate-Predict model
Switching Mechanism
Experimental Results
Introduction
Appearance-based for face recognition
Inevitable obstacle
Associate-Predict model
The studies of brain theories
Introduction
Identity Data Set and Face Representation
Identity data set
Face representation
Identity data set
200 identities from the Multi-PIE data set
7 pose
4 illumination
Face representation
Representation at the facial component level
12 facial components
Face F = (f1, f2, ..., f12) › fi for each component
Associate-Predict model
Appearance-prediction model
Likelihood-prediction model
Appearance-predictionmodel
Two input faces
Setting : SA , SB › A and B are facial components
Select the specific face image setting is equal to SB
› component A’ from this image
Appearance-predictionmodel
dA = |fA' − fB|› distance between the components
dB = |fB' − fA|
Final distance between A and B: 1/2 (dA + dB)
Appearance-predictionmodel
Adaptive distance dp
αA and αB : weight
After the “appearance-prediction” on all 12 facial components , we can obtain a new composite face
Likelihood-prediction model
Using classifier measure the likelihood of B belonging to A
Positive training samples› Input face › the K most alike generic identities
Switching Mechanism
Implement this switching mechanism› facial components : A and B › settings : SA = { PA , LA } and SB = { PB , LB }
Categorize the input pair into two classes› “comparable” › “not comparable” › based on the difference of SA and SB
Switching Mechanism
Comparable class › {|PA − PB| < 3 } and {|LA − LB| < 3 }
Not comparable class › the rest situations
Switching Mechanism
The final matching distance dsw
› da : the direct appearance matching › dp : the associate-predict model
Experimental Results
Experiments on the Multi-PIE and LFW data sets
Basic comparisons
Results on benchmarks
Basic comparisons
Holistic vs. Component
Basic comparisons
Positive sample size› number of positive samples is 1 +
28*k › “1” is the input sample › K is the selected number of top-
alike associated identities
Basic comparisons
K = 3 as the default parameter
Basic comparisons
Switching mechanism› the switch model can effectively improve
the results on both benchmark
Results on benchmarks
Multi-PIE benchmark
LFW benchmark
Multi-PIE Benchmark
LFW benchmark