Graphics Reference
In-Depth Information
Eventually, the distance between the probe and the gallery is measured by averaging values of
d i in Equation 5.4 over all pairs of matching facial curves.
It should be noted that the matching scheme does not exploit any specific assumption about
the correspondence of keypoints and facial curves to meaningful anatomical parts of the face.
As a consequence, the matching scheme can be used without any change to support matching
between a partial scan and a full scan of the face, thus enabling the recognition of faces with
missing parts and/or occlusions.
5.3.4 Comparison of State-of-the-Art Methods
In the following sections, we report the 3D face recognition results for several approaches
performing face identification/verification on the 3D face scans. Results are first reported for
the FRGC v2.0 data set for neutral and expressive scans, then the results on the GAVAB
data set are shown with the performance of methods capable of managing both expression
variations and large pose changes of the face, this latter one resulting in acquisitions with
missing parts.
Comparative Evaluation on FRGC v2.0 Data Set
For the first evaluation, the results scored by state-of-the-art 3D face recognition methods
on the FRGC v2.0 data set were presented (see Section 5.2). Facial scans are categorized,
according to the classification provided in the FRGC protocol, as showing neutral expression,
small expression, and large expression. The gallery consists of the first scans of each subject
in the database (466 scans in total), with the remaining scans forming the probe set. Using the
categories mentioned earlier, three recognition experiments are considered: (1) neutral versus
neutral, (2) neutral versus non-neutral, and (3) neutral versus all. In these experiments, the first
label indicates the gallery scans (neutral), whereas the second one refers to the probe scans
used in the experiment (i.e., neutral, non-neutral, and all). Overall recognition at rank-1 of
state-of-the-art methods are presented in Table 5.4, thus reporting about the effectiveness in
identification scenario.
Table 5.4 Comparison of rank-1 recognition rates on the FRGC
v2.0 data set for the state-of-the-art methods
Method
rank-1 RR
Spreeuwers (2011)
99.5%
Wang et al. (2010)
98.3%
ter Haar and Velkamp (2010)
97%
Berretti et al. (2010b)
94.1%
Queirolo et al. (2010a)
98.4%
Faltemier et al. (2008a)
97.2%
Kakadiaris et al. (2007b)
97%
Drira et al. (2012)
97%
 
Search WWH ::




Custom Search