EC-VIP-MC 2003,4th EURASIP Conference focused on

Video I Image Processing and Multimedia Communications, 2-5 July 2003,Zagreb, Croatia

A COLOUR FACE IMAGE DATABASE FOR

BENCHMARKING OF AUTOMATIC FACE DETECTION
ALGORITHMS
Prag Sharma, Richard B. Reilly

Dept. of Electronic & Electrical Engineering,

University College Dublin, Belfield, Dublin 4, Ireland.
prag(ii,ee.ucd.ie, [email protected]

Abstract: With increasing research in the area of face segmentation, new methods for
detecting human faces automatically are being developed. However, less attention is being
paid to the development of a standard face image database to evaluate these new algorithms.
This paper recognizes the need for a colour face image database and creates such a database
for direct benchmarking of automatic face detection algorithms. The database has two parts.
Part one contains colour pictures of faces having a high degree of variability in scale,
location, orientation, pose, facial expression and lighting conditions, while part two has
manually segmented results for each of the images in part one of the database. This allows
direct comparison of algorithms. These images are acquired from a wide variety of sources
such as digital cameras, pictures scanned in from a photo-scanner and the World Wide Web.
The database is intended for distribution to researchers. Details of the face database such as
the development process and file information along with a common criterion for performance
evaluation measures is also discussed in this paper.

Key Words: Face Detection, Colour Face Image Database, Evaluation Performance.

1. INTRODUCTION

Detecting human faces automatically is becoming a very important and challenging task in
computer vision research. The significance of the problem can be easily illustrated by its vast
applications, as face detection is the first step towards intelligent vision-based human
computer interaction. Face recognition, face tracking, pose estimation and expression
recognition all require robust face detecting algorithms for successful implementation.
Segmenting facial regions in images or video sequences can also lead to more efficient coding
schemes [ 11, content-based representation (MPEG4) [2], three-dimensional human face model
fitting, image enhancement and audio-visual speech integration [3]. Although a major area of
interest, many problems still need to be solved, as segmenting a human face successfully
depends on many parameters such as skin-tones under varying lighting conditions, complexity
level of the background in the image to be segmented and application for which the
segmentation is required. Inherent differences due to the existence of different ethnic
backgrounds, gender and age groups also complicate the face detection paradigm.

423
 EC-VIP-MC 2003.4th EURASIP Conference focused on
Video / Image Processing and Multimedia Communications. 2-5 July 2003, Zagreb, Croatia

With so many new applications, the development of faster and more robust face detection
algorithms has become a major area of research over the last few years. Techniques based on
knowledge of rules that capture the relationship between facial features, feature invariant
approaches that tend to define structural features that exist even when the pose, viewpoint and
lighting condition vary, and template matching methods that use several standard templates to
describe a face, are all being thoroughly investigated. A comprehensive survey on the methods
used to detect faces in images can be found in [4]. However, many of the successful claims
reported in the literature either use data sets that are too small or the test images acquired are
not standard images and so can show biased results favouring one method over another. The
comparison difficulties are due to the fact that much less attention has been paid to the
development of a standard face detection database that can be used as a benchmark to test the
performance of these new algorithms.

This paper is organised as follows. Section 2 gives details of the existing face detection
databases and the need for a new database. Section 3 describes the UCD Colour Face Image
Database for Face Detection in detail and Section 4 goes on to discuss the evaluation
measures that can be used compare the performance of one algorithm against another. The
concluding remarks are present in Section 5 .

2. EXISTING FACE IMAGE DATABASES

Although several face image databases exist, most of them are geared towards the evaluation
of face recognition algorithms. To this end, the face images in these databases end up being
taken with a specific set up in order to maintain a degree of consistency throughout each
database. For example, the most notable face image database is the FERET [5] database. In
this database, each greyscale image consists of an individual (head and neck visible only) on a
uniform and uncluttered background in mostly frontal positions. This is typical of many image
databases that are available. For example, the MIT [6] database consists of frontal and near-
frontal view images on a cluttered background. These databases are more useful as training
sets rather then test sets for face detection purposes. A detailed description of the face
recognition databases present can be found at [4].

Since all the face databases mentioned above are designed for the evaluation of face
recognition algorithms, they are not ideal candidates for the evaluation of face detection
algorithms. These databases do not provide the challenges that face detection algorithms can
encounter in real applications: such as poor image quality, presence of multiple faces and
faces with different orientations (up-right and rotated). To this end, some researchers have
compiled image databases to date. Sung and Poggio created two databases for face detection
[ 6 ] .The first set consists of frontal and near-frontal mugshots of different people in varying
lighting conditions whereas the second set contains high and low quality images with most
faces appearing in frontal and upright positions. All images are greyscale. Rowley et al. [7]
created a face database with most images containing more than one face on a cluttered
background. The face size varies significantly but most faces are upright and frontal and do

424
 EC-VIP-MC 2003,4th EURASIP Conference focused on
Video / lmaqe Processing and Multimedia Communications, 2-5 July 2003, Zagreb. Croatia

not contain colour information. To detect faces that are frontal but have a rotation in the image
plane, a separate database was compiled [8]. Schneiderman and Kanade further extended the
database to include faces with facial expression and in profile view [9]. While all the above
databases contain only greyscale images, Kodak has compiled a colour face image database
with a variety of face sizes and backgrounds [lo]. Table 1 summarizes the above mentioned
face detection databases.

Data Set Description

. -
and Poggio) [6] different people and the second set contains 23 images with 149
faces in complex backgrounds. Most faces are frontal and upright.
All images are greyscale.
CMU Test Set 130 images with a total of 507 frontal faces. Face sizes vary. All
(Rowley et al.) [7],[8] images are greyscale. Also contains 50 images with a total of 223
faces with 95% of the rotated at an angle of more than 10 degrees.
All images are greyscale.
CMU Profile Test Set 208 images with varying facial expressions and in profile view. The
(Schneiderman and images are greyscale.
Kanade) [9]
Kodak Data Set [IO] 80 images with 90% of the faces in frontal view. Wide variety of
resolutions and face sizes. The images are in colour.

As more face detection algorithms are being developed, new approaches that use multiple
features such as skin colour, size, shape and presence of facial features are being developed. A
typical approach starts with skin-colour based region detection [ll], [12]. This is primarily
due to the processing of colour information has proven to be much faster than the processing
of other facial characteristics. Moreover, if an effective colour model can be created that
adapts to varying lighting conditions, colour can be found to be invariant to changes in size,
orientation and partial occlusion of the face. Because of these advantages, more and more
algorithms are now using colour as a segmentation tool, [4]. However, most experimental
results reported using new algorithms often use very different test sets [I 11, [ 121. These test
sets are not standard and often use personal images that are not available to the research
community and often not in colour. The MIT and CMU databases contain greyscale images
only whereas the Kodak dataset has most of the faces in frontal view only. In order to compare
results fairly, a face database of colour images that can challenge face detection algorithms is
needed.

3. THE UCD COLOUR FACE IMAGE DATABASE FOR FACE DETECTION

The details of the UCD face detection database are described in this section. The database has
two parts. The first part contains a 100 colour images of faces with variations in background

425
 EC-VIP-MC 2003,4th EURASIP Conference focusedon
Video I imaae Processina and Multimedia Communications. 2-5 July 2003.Zaqreb, Croatia

3D Pose Orientation Occluded Structural

Faces Components
Frontal Intermediate Profile Upright Rotated Beard Glasses/
Moustache Sunglasses
182 91 26 242 50 39 18 26

426
 EC-VIP-MC 2003,4th EURASIP Conference focused on
Video I Image Processing and Multimedia Communications, 2-5July 2003, Zagreb. Croatia

4. PERFORMANCE EVALUATION MEASURES

In order to evaluate face detection methods fairly, it is essential to use standard and
representative datasets. Similarly, in order to compare the performance of several face
detection methods some standard metrics need to be defmed. Researchers often use detection
rates, false positives, false negatives etc. to compare results but this becomes more difficult
when different definitions are used for these metrics. Generally speaking, detection rate is
defined as the number of faces correctly detected to the number of faces determined by a
human. A false positive is where an image region is declared to be a face but it is not and a
false negative is where a face is not detected at all. However, in order to evaluate these
measures, the number and location of the faces present in the database need to be determined
by a human. To this end, part 2 of the UCD face database comes with hand-segmented results
for each image in the database as shown in Figure 2. Thus, false positives and negatives can
now be automatically evaluated by comparing the location of the detected regions to the hand
segmented results, and then using an accuracy measure to confirm a ‘correctly detected’ face.
A spreadsheet with details of the faces present in each image of the database is also provided
(See Figure 1).

Fig. 2. Sample images from Part 1 and Part 2 of the UCD Colour Face Image Database.

5. CONCLUDING REMARKS

In this paper, we descrihe the development of a colour image database for face detection. The
database has been developed with the intention of providing a common benchmark for
existing and new algorithms that are being developed. It contains colour images that provide
‘real world’ challenges to face detection algorithms by including faces with a large variety in
size, shape, orientation, expression and images that have varying lighting conditions,
resolution and backgrounds. It is available to all researchers working in the field and can be
downloaded by contacting the authors or visiting the site http:lldsp.ucd.ie/-prag.Further work

427
 EC-VIP-MC 2003.4th EURASIP Conference focused on
Video I Image Promssina and Multimedia Communications. 2-5Julv 2003. Zagreb. Croatia

is being camed out in developing a face detection algorithm using a multiple feature
approach.

REFERENCES:

[ l ] D. Chai, K. N. Ngan, Foregroundbackground Video Coding Scheme, Proc. IEEE Int.

Symp. Circuits Syst., Hong Kong, June 1997, vol. 11, pp. 1448-1451.
[2] F. Pereira, MPEG-4: Why, What, How and When?, Signal Processing: Image
Communication. Vol. 15,2000, p271-279.
[3] C. C. Chibeluschi, F. Deravi, A Review of Speech-Based Bimodal Recognition, IEEE
transactions on Multimedia, Vol. 4, No. 1, March 2001.
[4] M. H. Yang, D. J. Kriegman, Detecting Faces in Images: A Survey, IEEE Transactions on
Pattern Analysis and Machine Intelligence, Vol. 24, No. 1, January 2002.
[5] P. J. Phillips, P. Rauss, S. Der, FERET (Face Recognition Technology) Recognition
Algorithm Development and Test Report, Technical Report ARL-TR 995, U.S. Army
Research Laboratory.
[6] K-K. Sung, T. Poggio, Example-Based Learning for View-Based Human Face Detection,
IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 20, No. 1,pp. 39-51,
January 1998.
[7] H. Rowley, S. Baluja, T. Kanade, Neural Network-Based Face Detection, IEEE
Transactions on Pattern Analysis andMachine Intelligence, Vol. 20, No. 1, pp. 23-38 January
1998.
[8] H. Rowley, S. Baluja, T. Kanade, Rotation Invariant Neural Network-Based Face
Detection, IEEE Con$ Computer Vision and Pattern Recognition. pp. 38-44, 1998.
[9] H. Schneiderman,T. Kanade, A Statistical Method for 3D Object Detection Applied to
Faces and Cars, Proc. IEEE Conference on Computer Vision and Pattern Recognition, Vol. 1,
pp.746-751,2000.
[IO] A. C. Loui, C. N. Judice, S. Liu, An Image Database for Benchmarking of Automatic
Face Detection and Recognition Algorithms, Proc. IEEE Conference on Image Processing,
Vol. 1, pp.146-150, 1998.
[ 111 R-L. Hsu, M. Abdel-Mottaleb, Face Detection in Colour Images, IEEE Transactions on
Pattern Analysis and Machine Intelligence, Vol. 24, No. 5, May 2002.
[I21 C. Garcia, G. Tziritas, Face Detection Using Quantized Skin Colour Regions Merging
and Wavelet Packet Analysis, IEEE Transactions on Multimedia, Vol. 1, No. 3, September
1999.

428