Graphics Reference
In-Depth Information
4.5.2 Results
We can use the morphed face instances to perform 3D face recognition. For this experiment,
we computed the 953
700 dissimilarity matrices and sorted the
ranked lists of face models on decreasing similarity. From these ranked lists, we computed the
recognition rate (RR), the mean average precision (MAP) and the verification rate at 0.1% false
acceptance rate (VR@0.1%FAR). Aperson is recognized (or identified) when the face retrieved
on top of the ranked list (excluding the query) belongs to the same subject as the query. For 77
subjects in the UND set only a single face instance is available which cannot be identified, so
for this set the RR is based on the remaining 876 queries. The mean average precision (MAP)
of the ranked lists are also reported, to elaborate on the retrieval of all relevant faces, that is, all
faces from the same subject. Instead of focusing on 3D face retrieval application, one could use
3D face matching for imposter detection as well. For an imposter/client detection system, all
face matches with a dissimilarity above a carefully selected threshold are rejected. Lowering
this threshold means that more imposters are successfully rejected, but also that less clients
are accepted. We use the dissimilarity threshold at which the false acceptance rate is 0.1%,
which is also used in the face recognition vendor test. Because the VR@0.1%FAR depends
on similarity values, it is not only important to have relevant faces on top of the ranked lists,
but also that their similarity values are alike and differ from irrelevant faces.
Because the VR@0.1%FAR evaluation measure depends on the acquired similarity values,
there are several ways to influence this measure. Rank aggregation with the use of Consensus
Voting or Borda Count (Faltemier et al., 2008), for instance, reassigns similarity values on
the basis of the ranking. This way one can abstract from the actual similarity values, which
allows for the selection of a different imposter threshold and change the VR@0.1%FAR. Of
course, a rank-based threshold cannot be used in case of a one-to-one face matching, that is, a
scenario in which someone's identity must be confirmed or rejected. The application domain
for rank-based measures is the one-to-many face matching, that is, a scenario in which we
search for the most similar face in a large database.
In case of the face matching based on model coefficients, we assume that caricatures of
an identity lie on a vector from the origin to any identity. If we normalize the lengths of
these vectors, we neglect the caricatures and focus on the identity. This normalization step
also regulates the similarity values and thus influences the VR@0.1%FAR. In Table 4.4, we
report the face-matching results on the basis of the
L
1
and
L
2
distances between coefficient
vectors, before and after length normalization. Remarkable is the significant increase of the
VR@0.1%FAR for the normalized coefficient vectors, whereas the rankings are similar as
shown by the MAPs. Although we show in Table 4.4 only the results for the face model fitting
using seven components, it also holds for the one and four component case. Because the
L
1
distance between normalized coefficient vectors slightly outperforms the
L
2
distance measure,
we use this measure whenever we evaluate the performance of model coefficients.
Face retrieval and verification results based on anthropometric landmarks, contour curves,
and model coefficients are shown in Table 4.5. To each set of face scans we fitted the morphable
face model using one, four, and seven components. Each fitted component produces a 99
dimensional model coefficient vector with a different face instance as a result. The performance
of our face matching depends on both the number of components as well as the applied feature
set. The two main observations are that (1) the coefficient-based method outperforms the
landmark- and contour-based methods, and (2) that the use of multiple components can
×
953, 244
×
244, and 700
×
