Graphics Reference
In-Depth Information
depth images derived from 3D face models for the purpose of 3D face recognition. A similar
idea is used in Ohbuchi and Furuya (2004) to perform 3D object retrieval by visual similarity,
but in this case, points of a sampling grid are used and SIFT descriptors are computed for
them. In Berretti et al. (2010d), SIFT feature descriptors computed in correspondence with
facial landmarks of depth images are used to classify 3D facial expressions.
Partial Face Matching Using Keypoints and Facial Curves
In the following, we shortly summarize the approach for partial face matching proposed in
Berretti et al. (2011b). The approach uses SIFT keypoints to detect relevant and stable interest
points on depth images of the face and
facial curves
to model the depth of face scans along
the surface path connecting pairs of SIFT keypoints. In doing so, distinguishing traits of a face
scan are captured by the SIFT descriptors of detected keypoints as well as by the set of facial
curves identified by each pair of keypoints. Facial curves of gallery scans are also associated
with a measure of saliency so as to distinguish those that model characterizing traits of some
subjects from those that are frequently observed in the face of many different subjects. In the
comparison of two faces, SIFT descriptors are matched to measure the similarity between pairs
of keypoints identified on the two depth images. Spatial constrains are imposed to avoid outliers
matches. Then, the distance between the two faces is derived by composing the individual
distances between facial curves that originate from pairs of matching keypoints.
The use of keypoints of the face is advantageous with respect to using landmarks in the
case of partial face matching. In fact, just few landmarks can be detected automatically
with accuracy, and in the case of side scans, just a few of them are likely to be visible.
On the contrary, keypoints are not constrained to specific points of the face, and many of
them can be detected also on just a part of the face. According to this, the approach in
Berretti et al., (2011b) does not exploit any particular assumption about the position of the
keypoints on the face surface. Rather, the position of keypoints is expected to be influenced
by the specific morphological traits of each subject. In particular, assuming that the process
of keypoint detection incorporates a measure of the scale associated with each keypoint, the
assumption that detected keypoints correspond to meaningful landmarks is relaxed and the
more general assumption of within-subject
repeatability
is exploited: The position of the most
stable keypoints—detected at the coarsest scales—does not change substantially within facial
scans of the same subject. In particular, the approach in Berretti et al., (2011b) relies on the
detection of a number of keypoints on the 3D face surface and the description of the 3D face
surface in correspondence to these keypoints as well as along the paths connecting pairs of
keypoints. The SIFT are used to perform keypoints detection and description. SIFT keypoints
are extracted at scales of increasing
, so as to obtain a minimum of
N
keypoints for scan.
Then, the top
N
keypoints—starting from the highest
σ
σ
values—are retained and used as the
base for face description.
To extract depth images and detect SIFT keypoints, 3D face scans first undergo some
preprocessing (Berretti et al., 2010c). First, 3D faces are cropped using a sphere of radius 100
mm centered on the nose tip (the nose tip has been detected using the approach in Mian et al.
(2007b)). After this, face scans are transformed to depth images considering a frontal view
of the scan. To this end, the pose of the scans have been normalized using a solution on the
basis of an iterative
principal component analysis
(PCA) (Mian et al., 2007b). In addition,
