Applications - 3D Face Modeling, Analysis and Recognition - page 149

Table 5.3 3D static and dynamic (3D+time) face data sets for 3D facial expressions recognition

Database Format Subjects Scans Expressions Image Resolution Year

BU-3DFE static 100 2500 7 Color 1040 × 1329 2006

Bosphorus static 105 4666 6 Color 1600 × 1200 2008

ZJU-3DFE static 40 360 4 Color - 2006

ADSIP dynamic 10 210 7 Color 601 × 549 2009

BU-4DFE dynamic 101 606 6 Color 1024 × 681 2008

Hi4D-ADSIP dynamic 80 3360 14 Color 1024 × 1728 2011

and at the Bogazi¸i University (Bosphorus database) (Savran et al., 2008), have contributed to
push the research on 3D facial expression recognition and classification. Recently, some 4D
databases that include dynamic sequences of 3D scans along the time have been also released
for public use (Matuszewski et al., 2011, 2012; Yin et al., 2008). These 4D data are mainly
intended for studying facial expressions which are intrinsically related to the time variation of
the face.
More details on the individual data sets are reported in the following paragraphs, starting
with 3D static data sets and then presenting the most recent 4D dynamic data sets.

The Face Recognition Grand Challenge (FRGC) database

The Face Recognition Grand Challenge version 2.0 (FRGC v2.0) database (Phillips et al.,
2005) is the largest 3D database in terms of subjects enrolled and the most widely used in
order to compare methods for 3D face recognition. However, it does not provide categorized
facial expressions; hence it is less useful for the purpose of 3D facial expression recognition.
It includes 3D face scans partitioned in three sets, namely, Spring2003 set (943 scans of 275
individuals), Fall2003 and Spring2004 (4007 scans of 466 subjects in total). The Spring2003
set is also referred to as FRGC v1.0 an is used for training, set up, and tuning of methods,
whereas the Fall2003 and Spring2004 are referred to as FRGC v2.0 and are used for testing.
Face scans are given as matrices of points of size 480

640, with a binary mask indicating
the valid points of the face. Because of different distances of the subjects from the sensor
during acquisition, the actual number of points representing a face can vary. Individuals have
been acquired with frontal view from the shoulder level, with very small pose variations.
About 60 percent of the faces have neutral expression, and the others show expressions of
disgust, happiness, sadness, and surprise. In particular, according to a classification originally
performed at Geometrix (Maurer et al., 2005) and subsequently suggested by the FRGC
organizers, there are 2,469 neutral expression scans, 796 small cooperative expression scans,
and 742 large non-cooperative expression scans. Some scans have also occlusions because of
hair. Guidelines suggest using the Spring2003 set for training and the remaining two sets for
validation.
It is also interesting to note that a 2D color image of the subject is also provided for each
3D scan. In fact, acquisitions are performed using the Minolta Vivid 910 laser scanner that
permits a 2D color image capturing at the same time of the 3D acquisition (actually, the 2D
image is acquired just with a very short time delay, because different sensors are used for 2D

×