Graphics Reference
In-Depth Information
technology. One typically gets holes in the scanned data in locations of eyes, lips, and outside
regions. For instance, scans of people with open mouths result in holes in the mouth region.
Moreover, when the subject is noncooperative and the acquisition phase is unconstrained, these
result in variation in pose and unforeseen extraneous occlusions. Many studies have treated
pose and expression variations, but only a few have tried to focus on solving the problem of
occlusions even if a face can easily be partially hidden by objects like glasses, hats, scarves,
or a hand, hair, and beard. Occluded parts represent wrong information, which can degrade
recognition accuracy, so locating and removing occlusions on face quickly and automatically
are challenging tasks.
In the biometrics literature, recognition is a general term which includes: (1) face
verification
(or
authentication
) and (2) face
identification
(or
recognition
).
Face verification (“Am I who I say I am?”) is one-to-one match that compares a query
face image against a template face image whose identity is being claimed. To evaluate the
verification performance, the verification rate (the rate at which legitimate users are granted)
versus false accept rate (the rate at which imposters are granted access) is plotted, called the
receiver operating characteristic (ROC) curve.
•
Face identification (“Who am I?”) is one-to-many matching process that compares query
face image against all the template images in a face database to determine the identity of the
query face. The identification of the test image is done by locating the image in the database
that has the highest similarity with the test image. To evaluate identification performance the
cumulative matching characteristic (CMC) curve is used. This curve displays the cumulative
identification rates as a function of the rank distribution. This provides an indication of how
close one may be to getting the correct match if the rank-one match was incorrect.
•
One application of the verification task is access control where an authorized individual is
seeking access to a secure facility and presents to the system his or her identity. Here, a one-
to-one matching is performed: The 3D image for this individual is acquired, preprocessed, and
finally compared with an enrollment acquisition already incorporated in the system database.
If the similarity is greater than a defined threshold, the subject is granted access, otherwise
access is denied.
The classic application of the identification task is for identify the presence of suspect
people in database of recorded identity. In this case the one-to-many matching is performed
by first acquiring the 3D image of the individual, then preprocessing the scan to enhance the
quality and extract appropriate shape descriptors and finally comparing against a gallery of
already enrolled subjects. Depending on the specific scenario and on the size of the gallery,
the computational time is typically an additional issue in this case.
5.3.1 Challenges of 3D Face Recognition
When acquired in non-controlled conditions, scan data often suffer from the problem of missing
parts because of self-occlusions or laser-absorption by dark area. Actually, the 3D face needs
more than one scan to be fully acquired. Especially when the pose is not frontal as illustrated
in Figure 5.8
b
, the resulting scan is said to be 2.5D and not full 3D. However, this 2.5D scan is
roughly approximated by 3D scan by 3D face recognition community researchers. Moreover,
