1. Quantification of Facial Asymmetry for Expression-invariant Human Identification
Yanxi Liu
The Robotics Institute
School of Computer Science
Carnegie Mellon University
Pittsburgh, PA USA

2. Acknowledgement
Joint work with Drs. Karen Schmidt and Jeff Cohn (Psychology, U. Of Pitt).
Students who work on the data as research projects: Sinjini Mitra, Nicoleta Serban, and Rhiannon Weaver (statistics, CMU), Yan Karklin, Dan Bohus (scomputer science) and Marc Fasnacht (physics).
Helpful discussions and advices provided by Drs. T. Minka, J. Schneider, B. Eddy, A. Moore and G. Gordon.
Partially funded by DARPA HID grant to CMU entitled:
“Space Time Biometrics for Human Identification in Video”

3. Human Faces are Asymmetrical
Left Face
Right Face

4. Under Balanced Frontal Lighting (from CMU PIE Database)

5. What is Facial Asymmetry?
Intrinsic facial asymmetry in individuals is determined by biological growth, injury, age, expression …
Extrinsic facial asymmetry is affected by viewing orientation, illuminations, shadows, highlights …

6. Extrinsic Facial asymmetry on an image is Pose-variant
Left face
Original Image
Right Face

7. Facial Asymmetry Analysis
A lot of studies in Psychology has been done on the topics of
attractiveness v. facial asymmetry (Thornhill & Buelthoff 1999)
expression v. facial movement asymmetry
Identification
Humans are extremely sensitive to facial asymmetry
Facial attractiveness for men is inversely related to recognition accuracy (O’Toole 1998)
Limitations: qualitative, subjective, still photos

8. Motivations
Facial (a)symmetry is a holistic structural feature that has not been explored quantitatively before
It is unknown whether intrinsic facial asymmetry is characteristic to human expressions or human identities

9. The question to be answered in this work
How does intrinsic facial asymmetry affect human face identification?

10. DATA: Expression Videos Cohn-Kanade AU-Coded Facial Expression Database
Neutral Peak
joy
anger
disgust

11. Sample Facial Expression Frames
Total 55 subjects. Each subject has three distinct expression videos of varied number of frames. Total 3703 frames.
Neutral
Joy
Disgust
Anger

12. Face Image Normalization
Inner canthus
Face Midline
Philtrum
Affine
Deformation based on 3 reference points

13. Quantification of Facial Asymmetry
1. Density Difference: D-face
D (x,y) = I(x,y) – I’(x,y)
I(x,y) --- normalized face image,
I’(x,y) --- bilateral reflection of I(x,y) about face midline
2. Edge Orientation Similarity: S-face
S(x,y) = cos(Ie(x,y),I’e(x,y))
where Ie, Ie’ are edge images of I and I’ respectively,  is the angle between the two gradient vectors at each pair of corresponding points

14. Asymmetry Faces
An half of D-face or S-face contains all the needed information. We call these half faces Dh, Sh,Dhx, Dhy, Shx,Shy AsymFaces.
Original D-face S-face

15. Facial Asymmetry as a Biometric
Inner canthus
Philtrum
Spatiotemporal facial asymmetry of expression videos
Subj 85
Subj 10
Expression Videos from Cohn-Kanade Database
Original
D-face
S-face
joy
anger
disgust
neutral
Neutral Peak

16. Asymmetry Measure Dhy for two subjects each has 3 distinct expressions
forehead
forehead
chin
chin
Joy anger | disgust
Joy | anger | disgust

17. spatial
temporal
Forehead -- chin
Forehead -- chin
Forehead -- chin

18. spatial
temporal
Forehead -- chin
Forehead -- chin
Forehead -- chin

19. spatial
Forehead -- chin
Forehead -- chin
Forehead -- chin

20. Evaluation of Discriminative Power of Each Dimension in SymFace Dhy
Variance Ratio
Bridge of nose
forehead
chin

21. Most Discriminating Facial Regions Found

22. Experiment Setup
55 subjects, each has three expression video sequences (joy, anger, disgust). Total of 3703 frames. Human identification test is done on ----
Experiment #1: train on joy and anger, test on disgust;
Experiment #2: train on joy and disgust, test on anger;
Experiment #3: train on disgust and anger, test on joy;
Experiment #4: train on neutral expression frames,test on peak
Experiment #5: train on peak expression frames,test on neutral
The above five experiments are carried out using
(1) AsymFaces, (2) Fisherfaces, and (3) AsymFaces and FisherFaces together.

23. Sample Results: Combining Fisherfaces (FF) with AsymFaces (AF) (Liu et al 2002)
Data set is composed of 55 subjects, each has three expression videos.
There are 1218 joy frames, 1414 anger frames and 1071 disgust frames. Total number of frames is 3703.

24. All combinations of FF and AF features are tested and evaluated quantitatively

25. Complement Conventional Face Classifier
107 pairs of face images taken from Feret database.
It is shown that asymmetry-signature’s discriminating power
demonstrated
(1) has a p value << from chance
(2) is independent from features used in conventional classifiers, decreases the error rate of a PCA classifier by 38% (15%  9.3%)

26. Quantified Facial Asymmetry used for Pose estimation

27. Summary
Quantification of facial asymmetry is computationally feasible.
The intrinsic facial asymmetry of specific regions captures individual differences that are robust to variations in facial expression
AsymFaces provides discriminating information that is complement to conventional face identification methods (FisherFaces)

28. Future Work
(1) construct multiple, more robust facial asymmetry measures that can capture intrinsic facial asymmetry under illumination and pose variations using PIE as well as publicly available facial data.
(2) develop computational models for studying how recognition rates is affected by facial asymmetry under gender, race, attractiveness, hyperspectral variations.
(3) study pose estimation using a combination of facial asymmetry with skewed symmetry.