In: Wagner W., Székely, B. (eds.): ISPRS TC VII Symposium - 100 Years ISPRS, Vienna, Austria, July 5-7, 2010, IAPRS, Vol. XXXVIII, Part 7B 
In: Wagner V 
ACCURACY OF 3D FACE RECOGNITION FRAMEWORKS 
V. Bevilacqua ^ b , M. Caprioli c , M. Coltellino b , M. Giannini a,b ,G. Mastronardi a,b , V. Santarcangelo b 
a DEE (Dipartimento di Elettrotecnica ed Elettronica) Polytechnic of Bari 
Via Orabona, 4-70125 Bari, Italy 
b e.B.I.S. s.r.l. (electronic Business In Security) Spin-Off of Polytechnic of Bari 
Via Pavoncelli, 139-70125 Bari, Italy 
c DVT (Dipartimento di Vie e Trasporti) Polytechnic of Bari 
Via Orabona, 4-70125 Bari, Italy 
m. caprioli@poliba. it 
KEY WORDS: CAD, 3D Geometrie Modelling, Image processing. 
ABSTRACT 
This paper represents a survey of the state of art reached in 3D Face Recognition frameworks and show some different approaches 
developed and tested by its authors. We have designed a strong algorithm that is based on genetic algorithms, Principal Component 
Analysis (PCA) and face geometry assumptions, for head pose normalization of 3D scanned face models. Experiments conducted on 
the GavaDB database show a 100% success rate in correctly those models that “stare” at the camera (with a perfect alignment in 83% 
of the cases). A previously developed algorithm for nose-tip detection based on an adapted Khoshelham GHT has been used to create 
an automatic repere’s points detection system with the purpose of obtaining a biometric system for AFR (Automatic Face 
Recognition) using 3D Face templates. Subsequently two different methodologies, based respectively on an unsupervised self 
organizing neural network (SOM) and upon a graph matching, have been implemented to validate the performance of the new 3D 
facial feature identification and localization algorithm. Experiments have been performed on a dataset of 23 3D faces acquired by a 
3D laser camera at eBIS lab with pose and expression variations. Then an optimization of the search of the points ALS and ALD of 
the nose and a new graph approach for the recognition base on several new points has been implemented. Experiments have been 
performed on a dataset of 44 faces, acquired by a 3D laser camera at eBIS lab with pose and expression variations, scoring a result of 
100% of recognition. 
1. INTRODUCTION 
In Computer Vision object or shape detection in 2D/3D images 
is very hard to solve because shapes can be subject to 
translations, can change by color, can be subject to scale and 
orientation, can endure occlusions and moreover data 
acquisition can introduce high levels of noise. Always more and 
more researchers work on 3D face processing including 
modelling and recognition. Usually, a 3D face is a group of 
high dimensional vectors of the x, y and z positions of the 
vertexes of a face surface. Face recognition based on 3D has the 
potential to overcome the challenging problems caused by the 
expression, illumination variations. However, many 3D face 
recognition approaches, especially the feature based ones 
require a robust and accurate facial feature localization. To the 
best of our knowledge, most of the methods do not use 
benchmark datasets to evaluate their results. Romero et al. 
presented the first work on benchmark datasets based on FRGC 
database. They manually marked landmarks of eleven facial 
features including nose tip and eye comers. With those marked 
feature locations, the results of automatic feature identification 
can be measured and evaluated. This paper focuses on a 
different task and proposes two different steps where the former 
concerns the automatic identification and localization of a 3D 
Nose facial features, and the latter is the validation of the 
correct data obtained by means two independent 3D face 
recognition tasks. Experiments was performed by using an ASE 
format dataset of 23 images acquired in by a 3D laser scanner 
based on structured light with a resolution of 640 by 480. The 
applied 3D reconstruction method is based on analysis of two 
images, obtained illuminating sequentially the target surface, 
with two sinusoidal fringe patterns shifted in phase by 180° and 
the obtained data are a 3D clouds of point in ASE format. Due 
to the short integration time, images may be acquired without 
room darkening (normal ambient illumination). Optical 
methods used for measurement of 3D geometry are based on 
two general principles, one related to the almost direct 
measurement/analysis of the light flight time (lidars, 
interferometers and holography) and the second related to 
geometrical triangulation. In the last group the viewing angles 
of a point of the object can be calculated from its position in the 
registered image/images or from the applied system settings 
(scanning devices). Single image acquisition with a projected 
fringe pattern of high spatial frequency is another kind of 
compromise towards one-shot entire scene 3D photography. 
This type of acquisition is usually related to the Fourier analysis 
of the image. As the main drawback of this approach is its 
sensitivity to the quality of the registered image (fringe pattern), 
which directly influences the precision of the phase unwrapping 
procedures as well as the final accuracy. Another weak point of 
this method is its intrinsic low resolution and artefacts caused 
by the extensive filtering applied in the spatial frequency 
domain, and in particular, in zones close to the analyzed lobe 
borders. As nose tip is the most prominent feature of the face, 
most of the previous work perform nose tip detection and uses 
the nose tip as the foundation to detect other features. However, 
many previous facial feature identification algorithms use an 
assumption that the nose is the closest point to the camera or 
device which acquires the 3D data. Although this supposition is 
true in most cases, there is no 100% guarantee due to the noise. 
Various pose rotations and the complex situation of hair and 
clothes could make some places closer than the nose. Some 
algorithms make use of the corresponding 2D texture 
information 1 
within the s< 
texture and 
example, in : 
FRGC, the 2 
with the 3E 
Mastronardi 
localisation 
automatic n 
obtaining a 
Recognition 
on a databa: 
by the GAV 
at the Unive 
show their r 
tip by mear 
results obta 
dataset with 
3D nose tip 
code deveh 
format clou 
Santarcange 
Mastronardi 
Zaccaglino, 
algorithm tc 
different ap; 
all the five i 
K„ 2007) a 
which face 
graphs, eacl 
geometry o 
Local Fea 1 
automatical 
each indivii 
matching b< 
the trained i 
on the OR 
FERET fac< 
2. F 
2.1 Stereo 
In the first' 
F., Nitti, D 
stereo-mate 
search mat 
on correspc 
selected N 
points on t 
are researc 
epipolar li 
proposed: 
cameras (o 
images ft 
calibration 
case 3D ri 
these step: 
from mat 
visualizatic 
86