1. Eye Movements Biometrics Where you look shows who you are
Paweł Kasprowski, PhD Silesian University of Technology

2. Outline Eye movements at a glance How to measure eye movements
Why do we measure it?
Eye movements biometrics (EMB) – advantages
EMVICompetition
Summary

3. Physiology of eye movements
Non-uniform picture quality across the visual field
Area of high visual acuity zone
Fovea
Acuity drops outside
Eyes in constant movement

4. Eye movements Oculomotor system Movement Elements Six muscles
Three nerves
Movement
voluntary
involuntary
Elements
Fixations
Saccades
Smooth pursuits
Tremors, drifts

5. Eye movements Oculomotor system Movement Elements Six muscles
Three nerves
Movement
voluntary
involuntary
Elements
Fixations
Saccades
Smooth pursuits
Tremors, drifts

6. Eye movements Oculomotor system Movement Elements Six muscles
Three nerves
Movement
voluntary
involuntary
Elements
Fixations
Saccades
Smooth pursuits
Tremors, drifts

7. Eye movement measurement
First researches: Emile Javal, 1879
First eye-tracker: Edmund Burke Huey, 1897
Different methodologies:
Contact lens
Elektro-oculography (EOG)
Video-oculography (VOG)

8. Contact lenses
Very accurate but very intrusive

9. Elektrooculography (EOG)
Electric potentials measured with electrodes placed around the eyes
Measures rather movement that gaze point

10. Videooculography (VOG)
Detects eyes on face image
Measures reflection of infrared light
Purkinje images 1-4
Non-invasive
May even be covert
Computational power needed

11. Our eye tracker
Frequency: 1 kHz
Accuracy: about 1 degree

12. Eye movement visualization
X-Y temporal graphs

13. Eye movement visualization
Scan paths
Olivier Le Meur,

14. Eye movement visualization
Heat maps
Tommy Strandvall,

15. Eye movement usages Cognitive studies Medicine Usability of interfaces
How our brain works
Education, experience
Medicine
Diagnoses
Usability of interfaces
What attracts attention on an image?
Advertisement, web pages
Eye as a pointer
Gaze instead of mouse?
Adjusting image quality basing of a gaze point

16. Reading Every man reads differently
Reading easy and difficult texts is different
It is possible to find out education level!

17. Yarbus (1967)
The way we move our eyes depends on what we are looking for
Yarbus, A. L. (1967), Eye Movements and Vision, New York: Plenum

18. Gender differences
EyeTrackShop, MRC International

19. Gender differences
EyeTrackShop, MRC International

20. Interface usability
Checking if application, interface, web page or leaflet is readable and understandable
Checking if users pay attention where we want
FengGui,

21. Eye as a pointer Gaze instead of mouse?
Problem: vision is our input not output
Eye pointing may be not comfortable
"Midas touch problem"
Possible usages:
For disabled
Adaptive interface
PIN entering (eye signature)
CEATEC, Fujitsu eye-tracking technology

22. Eye Movement Biometrics
Identification of people based on their eye movements
Another example of biometric identification method
Weird Science: 10 Forms of Biometric Authentication

23. Advantages Physiological elements Behavioral elements Acceptability
Oculomotor Plant
Muscles
Behavioral elements
Depends on stimulus
"Measures" knowledge and experience
Subject must be conscious
Acceptability
Accessibility
Multimodality

24. Acceptability Easy to use Non-intrusive
May be done during normal activity
Possible covert identification

25. Accessibility Expensive for now It changes! Simple VOG eye-trackers
Tablets
Smartphones
Mirametrix, Samsung

26. Multimodal biometrics
Video-based eye trackers may be easily combined with:
Face recognition
Iris scanning
Mouse and keystroke dynamics
Other imaging methods

27. Challenges
Distinctiveness
Repeatability
as in every behavioral method

28. State of the Art Idea of Prof. Ober (1945-2007) First publications
inventor of Ober2 eye tracker
First publications
KASPROWSKI, P., OBER, J Eye movement tracking for human identification, In 6th World Conference BIOMETRICS’2003, London. (Best Poster on Technological Advancement)
KASPROWSKI, P., OBER, J Eye Movement in Biometrics, In Proceedings of Biometric Authentication Workshop, European Conference on Computer Vision in Prague 2004, LNCS 3087, Springer-Verlag.
KASPROWSKI P., OBER J Enhancing eye movement based biometric identification method by using voting classifiers. SPIE Defense & Security Symposium.

29. Research centers Texas State University (Komogortsev)
Jumping point, reading, OPC, CEM, +Iris
SC University, Finland (Bendarik, Kinnunen)
Static point stimulus, task independent
University of Patras, Greece (Rigas)
Static images (faces)
Silesian Univ. of Technology (Kasprowski)
Jumping point, noninteractive

30. EMB Challenges Results are not comparable No established techniques
Different equipment, experiments, methods
No established techniques
Lack of common databases
Our idea: competition similar to FVC

31. Competition Competition uses ready-to-use datasets Competition goals
Eye movement recordings without proper identification
Some training data
Search for the best classification algorithm
Competition goals
Popularize the EMB
Provide single reference point with a set of datasets
Highlight importance of data quality

32. EMVIC 2012 Official competition of: Task: Web page:
IEEE Fifth International Conference on Biometrics: Theory, Applications and Systems (BTAS 2012)
Task:
Assign every test sample to proper user id
Web page:

35. EMVIC Competition Four different datasets Jumping point type
Dataset A (978 recordings of 37 persons)
8 sec, rec. per person
Dataset B (4168 recordings of 75 persons)
8 sec, rec. per person
Dataset C (116 recordings of 29 persons)
Only left eye, 4 rec. per person, horizontal jumps
Dataset D (108 recordings of 27 persons)
Only left eye, 4 rec. per person, random jumps

36. Competition rules Publicly available: Evaluation
Labeled training set
Unlabeled testing set
Evaluation
Participants should correctly label unlabeled samples
Training samples for building classification model
Emvic.org and Kaggle competitions
Only Dataset A for Kaggle

37. Competitors 45 users registered in the main competition
106 submissions
Quatar, Wisconsin, Dehli, Brno, Alaska, Luxemburg…
49 competitors in Kaggle competition
524 submissions (only one daily limit per user)

38. Results Dataset A – 97% Dataset B – 95% Dataset C – 58%
Dataset D – 67%
Question:
Why are the results so different for different datasets?

39. Why so different results?
Number of samples per person
About 26 per person in A, 55 in B, exactly 4 for C,D
Sequences of samples in A, B
Lack of calibration and worse data quality in A, B
Horizontal movements only for C
Random stimulus for D
Last but not least: order of datasets
Competitors started with Dataset A and didn't have enough time for C and D

40. Summary Eye movement biometrics has a lot of advantages
Vision is the main human being's input
It is not mature enough
We need more generally available and comparable datasets

41. Thank you for your attention
Paweł Kasprowski
Drinks reception is waiting!
Thank you for your attention