Three-Dimensional Face Recognition Using Surface Space Combinations

Three
-
Dimensional Face Recognition Using
Surface Space Combinations

Thomas Heseltine, Nick Pears, Jim Austin

Advanced Computer Architecture Group

Department of Computer Science
-

University of York

www.cs.
york
.ac.uk/~tomh

tom.heseltine@cs.york.ac.uk

2

Introduction

•
Face recognition offers several advantages over other biometrics

•
Covert operation.

•
Human readable media.

•
Public acceptance.

•
Data required is readily available
–

police databases etc.

But…

•
Growing interest in biometric authentication

•
National ID cards, Airport security (MRPs), Surveillance.

•
Fingerprint, iris, hand geometry, gait, voice, vein and face.

3

Limitations of 2D Face Recognition

•
Lighting conditions.

•
Different lighting conditions for enrolment
and query.

•
Bright light causing image saturation.

•
Head orientation.

•
2D feature distances appear to distort.

•
Image quality.

•
CCTV, Web
-
cams etc.

•
Facial expression.

•
Changes in feature location and shape.

•
Partial occlusion

•
Hats, scarves, glasses etc.

System effectiveness is highly dependant on image capture conditions.

Face recognition is not as accurate as other biometrics.

Error rates that are too high for many applications in mind.

Result:

4

A Possible Solution…


3D Face Recognition


•
Newly emerging 3D cameras allow sub
-
second generation of 3D face models

•
Using 3D face models for recognition
potentially provides the following benefits:



•
Use of geometric depth information rather than colour and texture


Invariant to lighting conditions

•
Ability to rotate face model in 3D space


Invariant to head angle

•
3D models captured to scale


Absolute measurements invariant to camera distance

5

3D Face Data


•
Generated using a stereo vision camera
enhanced by light projection.

•
Stored in OBJ file format.

•
Approximately 8000 points on a facial surface.

•
Greyscale texture mapped.

Wire
-
mesh

Texture

Polygons

Lighting

6

The Fishersurface Method

•
Developed in previous work

•
[Heseltine, Pears, Austin. Three
-
Dimensional Face Recognition: A Fishersurface Approach].

•
Adaptation of the fisherface method to 3D face data.

•
[Belhumeur, Hespanha, Kriegman, Eigenfaces vs. Fisherfaces: Face Recognition using class specific linear projection].

•
Uses PCA + LDA to create a surface space projection matrix


Orientate 3D face models to face directly forwards.


Convert to depth
-
map representation (60 by 90 pixels).


Train on 300 depth maps of 50 different people.


Projected depth maps compared using Euclidean or cosine
distance metrics.

7

8

Test Database

•
Little publicly available 3D Face data, so we collect our own 3D face database:

•
Database now consists of over 5000 face models of over 350 people.

•
Large range of expression, orientation, gender, ethnicity, age.

•
We take a subset of this database (1770 models) for training and testing.

•
300 3D models of 50 people for training

•
1470 3D models of 280 people for testing

9

Error Rates

•
Error curves produced for all surface
representations.

•
EER taken as a single comparative
value.

•
A large range of error rates produced.

10

Surface Space Analysis Using FLD












c
i
x
i
c
i
i
i
i
m
x
m
m
d
1
2
1
1
2
)
(
)
(
Fisher’s Linear Discriminant calculates the ratio of
between
-
class and within
-
class scatter, providing
an indication of discriminating ability.

11

Combining Surface Space Dimensions

However, even the worst representations
produce a surface space with some highly
discriminatory dimensions.

Some surface representations
perform better than others.

•
Extract “best” dimensions
from all surface spaces

•
Incorporate into a single
combined surface space

•
Dividing each element by its within
-
class standard deviation
effectively weights each dimension evenly.

12

Test Procedure

13

3D Combination Results

9.3% EER on the blind test set
(11.5% single)

8.2% EER on the full test set
(11.3% single)

7.2% EER on test set used to calculate
dimension combinations
(11.6% single)

Face space dimensions are
selected from a wide range of
systems and combined to form a
single unified 3D face space.

Using the cosine metric results in
combining more surface space
dimensions.

Questions?


Thomas Heseltine, Nick Pears, Jim Austin

Advanced Computer Architecture Group

Department of Computer Science
-

University of York

www.cs.
york
.ac.uk/~tomh

tom.heseltine@cs.york.ac.uk

Three
-
Dimensional Face Recognition Using

Surface Space Combinations