I. XXXVIII, Part 7B 
In: Wagner W„ Székely. B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Voi. XXXVIII, Part 7B 
;elo b 
lifferent approaches 
rincipal Component 
ments conducted on 
ct alignment in 83% 
; been used to create 
l (Automatic Face 
unsupervised self- 
ince of the new 3D 
faces acquired by a 
ts ALS and ALD of 
ten ments have been 
s, scoring a result of 
1 phase by 180° and 
in ASE format. Due 
?e acquired without 
mination). Optical 
metry are based on 
the almost direct 
ght time (lidars, 
second related to 
the viewing angles 
m its position in the 
ied system settings 
Dn with a projected 
is another kind of 
e 3D photography, 
the Fourier analysis 
this approach is its 
age (fringe pattern), 
e phase unwrapping 
lother weak point of 
ind artefacts caused 
; spatial frequency 
0 the analyzed lobe 
feature of the face, 
) detection and uses 
r features. However, 
1 algorithms use an 
nt to the camera or 
h this supposition is 
tee due to the noise, 
ituation of hair and 
tan the nose. Some 
nding 2D texture 
information to detect the face area first then localize the nose tip 
within the selected 3D face crop. This technique requires 2D 
texture and 3D shape to correspond correctly. However, for 
example, in some face datasets such as the Spring2003 subset of 
FRGC, the 2D texture channel is not always perfectly matched 
with the 3D shape channel. In (Bevilacqua, V., Casorio, P., 
Mastronardi, G., 2008) the authors addressed automatic nose tip 
localisation adapting Khoshelham GHT and describe an 
automatic repere point detection system with the purpose of 
obtaining a biometric system for AFR (Automatic Face 
Recognition) using 3DFace templates. That research was lead 
on a database of 3D-faces in ASE format, the GavaDB, given 
by the GAVAB research group of computer science department 
at the University of King Juan Carlos in Madrid and the authors 
show their results and claim successful localization rate of nose 
tip by means several correspondences in terms of very close 
results obtained by different algorithms but using a limited 
dataset without benchmark evaluation. Then starting from the 
3D nose tip co-ordinates, obtained running the same previous 
code developed to automatically localize nose tip in ASE 
format cloud of points, in (Bevilacqua, V., Mastronardi, G., 
Santarcangelo, V., Scaramuzzi, R., 2010) (Bevilacqua, V., 
Mastronardi, G., Piarulli, R., Santarcangelo, V., Scaramuzzi, R., 
Zaccaglino, P., 2009 ) the authors, propose firstly a new 
algorithm to localize other four 3D nose features, and two other 
different approaches to perform a 3D face recognition by using 
all the five nose points. In (Fazl-Ersi, E., Zelek, J. S., Tsotsos, J. 
K., 2007) a novel 2D face recognition method is proposed, in 
which face images are represented by a set of local labelled 
graphs, each containing information about the appearance and 
geometry of a 3-tuple of face feature points, extracted using 
Local Feature Analysis (LFA) technique. That method 
automatically learns a model set and builds a graph space for 
each individual, then proposes a two-stage method for optimal 
matching between the graphs extracted from a probe image and 
the trained model graphs is proposed and achieves perfect result 
on the ORL face set and an accuracy rate of 98.4% on the 
FERET face set. 
2. PREVIOUS WORKS AND BACKGROUND 
2.1 Stereo Matching 
In the first work (Bevilacqua, V., Mastronardi, G., Melonascina, 
F., Nitti, D., 2006) has been proposed a passive intensity based 
stereo-matching algorithm using a constraint handling GA to 
search matched points. Approach used search correspondences 
on corresponding epipolar lines (not on the whole image), then, 
selected N points on the epipolar line of the first image, N 
points on the corresponding epipolar line in the second image 
are researched, the research is carried out for each couple of 
epipolar lines. Two stereo-matching algorithms have been 
proposed: for generic scenes using images from parallel 
cameras (or rectified images), for 3D face reconstruction using 
images from parallel or non-parallel cameras (camera 
calibration is required to compute epipolar lines). In the last 
case 3D reconstruction process has been implemented using 
these steps: Stereo-matching, calculation of 3D coordinates 
from matched points using triangulation, generation and 
visualization of a 3D mesh. 
Figure 1. 3D Reconstruction through Stereo Matching. 
2.2 Hough Transform 
In previous work (Bevilacqua, V., Casorio, P., Mastronardi, G., 
2008) we adapted Khoshelham GHT in an attempt to apply it on 
3D-Face shaded model for nose-tip detection. In that way it was 
possible to create an automatic repere’s points detection system 
with the purpose of obtaining a biometric system for AFR 
(Automatic Face Recognition) using 3DFace templates. The 
research was lead on a database of 3D-faces in ASE format, the 
GavaDB, given by the GAVAB research group of computer 
science department at the 
University of King Juan Carlos in Madrid. One of the more 
effective solutions for shape detection is the Hough Transform. 
Formulated for the first time in early ‘60s, it originally was able 
to recognize shapes that had analytical description such as 
straight lines, circles and ellipses in 2D intensity images. In 
1981 Ballard (Ballard, D. H., 1981) proposed an extension 
defined Generalized Hough Transform (GHT) for generic shape 
detection by using the R-Table, a table that describes the shape 
to search respect to a reference point that could represent the 
center of the . Many efforts have been done in order to try to 
extend the GHT in three- pattern dimensional images. 
Khoshelham (Khoshelham, K., 2007) proposed an extension of 
Ballard GHT for three-dimensional images constituted by point 
clouds obtained by means of laser- scanner acquisitions, for 
generic applications. 
2.3 3D Head Pose Normalization 
In (Bevilacqua, V., Andriani, F., Mastronardi, G., 2009) we 
have first proposed a model reconstruction scheme for human 
head’s point cloud, as it’s directly returned from scanning 
hardware, consisting of a simple pipelining of existing, fairly 
renowned algorithms. Then we have designed a two-step 
process for model alignment, based on two different error 
measures and their corresponding minimization techniques. In 
particular we presented a software system for fully automatic 
alignment of the 3D model of a human face. Starting with the 
point cloud of a human head, previously segmented from the 
background, pose normalization is attended with a novel, purely 
geometric approach. In order to solve the 6 degrees of freedom 
of this problem, we exploit natural mirror symmetry of human 
faces, then analyse frontal profile shape and finally align 
model’s bounding box with nose tip position. 
Normalizing the position of a human face implies achievement 
of a perfect (or quasi-perfect) alignment of the head 3D model 
87