Available online at www.sciencedirect.

com

ScienceDirect
Procedia Computer Science 42 (2014) 372 – 377

International Conference on Robot PRIDE 2013-2014 - Medical and Rehabilitation Robotics and
Instrumentation, ConfPRIDE 2013-2014

Automatic craniofacial anthropometry landmarks detection and

measurements for the orbital region
Salina Mohd Asiab*, Nor Hidayah Ismaila, Roshahida Ahmadc, Effirul Ikhwan Ramland,
Zainal Arif Abdul Rahmana*
a
Dept of Oro-Maxillofacial Surgical and Medical sciences, Faculty of Dentistry, University of Malaya, Kuala lumpur, Malaysia
b
School of Computer and Communication,University Malaysia Perlis,Tkt 2 Bgn KWSP, Jalan Bukit Lagi, 01000 Kangar, Perlis, Malaysia
c
Dept. of Paediatric Dentistry and Orthodontic, Faculty of Dentistry, University of Malaya, Kuala Lumpur, Malaysia
d
Faculty of Computer Science And Technology Information, University Of Malaya, Kuala Lumpur Malaysia

Abstract

Facial landmarks detection is undoubtedly important in many applications in computer vision for example the face detection and
recognition. In craniofacial anthropometry, consistent landmarks localization as per standard definition of the craniofacial
anthropometry landmarks is very important in order to get accurate craniofacial anthropometry data. In this article we
demonstrated an automatic detection of craniofacial anthropometry landmarks at the orbital region. 3D images of 100
respondents’ were photogaphed using Vectra-3D in controlled environment. Craniofacial measurements of 30 3D images were
measured using VAM software. Two data sets of left and right eyes positive training data were created to train 'en' and 'ex' haar
cascade classifiers. These classifiers were used to detect and locate the inner (en) and outer (ex) eye corners. We automatically
measured the left and right eye fissures length (en-ex), the intercanthal (en-en) and the biocular (ex-ex) width. Statistical analysis
was performed on the measurements taken by Vectra 3D and by our software with paired t-test and calculated the ICC indices.
We observed quite amount of false positive detections. We removed the false positive and predicted the eye corners. Our
classifiers able to detect and locate the ‘en’ and the ‘ex’ in 59 out of 60 test images. Our results show accurate detection of ‘ex’
and ‘en’ craniofacial landmarks as per standard definition. The paired t-test showed that all four (4) measurements are no
significant difference with the p values on 95% confidence level are above 0.05. The ICC indices for the measurements were
from 0.4 to 0.78. In conclusion, our trained enHaar and exHaar cascade classifiers were able to automatically detect the 'en' and
'ex' craniofacial anthropometry landmarks in controlled environment. The measurements were clinically no significant
differences with the mean different were less than 1 mm in both eye fissures and intercanthal except the biocular width ( 1.16
mm). The consistency of the measurements between the two methods are good for the intercanthal width and moderate for the
biocular width and for both eye fissure lengths.
© 2014 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
© 2014 The Authors. Published by Elsevier B.V.
(http://creativecommons.org/licenses/by-nc-nd/3.0/).
Peer-review under
Peer-review underresponsibility
responsibilityofof
thethe
Center forfor
Center Humanoid Robots
Humanoid and and
Robots Bio-Sensing (HuRoBs).
Bio-Sensing (HuRoBs)
Keywords: Craniofacial Anthropometry; facial landmarks; Haar Cascade Classifier.

1877-0509 © 2014 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY-NC-ND license
(http://creativecommons.org/licenses/by-nc-nd/3.0/).
Peer-review under responsibility of the Center for Humanoid Robots and Bio-Sensing (HuRoBs)
doi:10.1016/j.procs.2014.11.076
 Salina Mohd Asi et al. / Procedia Computer Science 42 (2014) 372 – 377 373

1. Introduction

Craniofacial Anthropometry is a standard facial measurement used in a wide range of applied and research
contexts, hence, reliable and accurate data are very important and cannot be compromised [1-2]. The use of
computer technology such as Vectra 3D stereo-photogrammetry system and other similar systems helped the
researchers in acquiring the data fast, reliable and accurate [3-5]. Unfortunately, the systems require human
interaction in locating the facial landmarks. Consistency in locating the landmarks is an important parameter in
contributing the measurement errors as discussed in [2]. Toma and Zhurov studied the intra and inter examiners
error in reproducing the facial landmarks on the 3D image at 3 planes x, y and z [6]. Their results showed 11% of
inter examiners and 17% of intra examiners reproduced landmarks with more than 1 mm difference.
Dauglas had attempted to locate the eye landmarks automatically in her study of Fetal Alchoholic Syndrom
(FAS). She proposed an automatic measurements of the orbital region to compare the measurements of normal and
the FAS group. She used genetic algorithm and eye template to automatically locate the eye landmarks in her study
[7]. Mutsugawa and Douglas extensive study on the used of stereophotogammetry craniofacial anthropometry as a
tool in FAS screening has been accepted as a standard procedure to screen FAS in children[8-9]. Automatic locating
and extracting the facial landmarks proposed by Douglas is an alternative method to reduce the inter and intra
examiner errors in craniofacial anthropometry.

Abbreviation
en endocathion
en exocanthion
en-ex eye fissure length in mm
en-en inter canthal length in mm
ex-ex biocular width

* Corresponding author. Tel.: +6019-4776123, fax: +604-9885462.

E-mail address salina@unimap.edu.my;

1.1. The Cascade Classifier

Viola and Jones object detection framework is robust and accurately detect face in image. Viola and Jones used
Haar-like features proposed by Papageorgiou in their algorithm to detect a face. These features encode differences in
average intensities between two rectangular regions, and they are able to extract texture without depending on
absolute intensities. The contrast variances between the pixel groups are used to determine relative light and dark
areas. Viola and Jones used three kinds of features, a two rectangular feature, a three rectangular features and a four
rectangular feature as shown in figure 1. The sum of the pixels which lie within the white rectangles are subtracted
from the sum of pixels in the grey rectangles. Two-rectangle features are shown in (A) and (B). Figure (C) shows a
three-rectangle feature, and (D) a four-rectangle feature. [10]

Figure 1.Example rectangle features

They proposed an efficient system for evaluating these features which is called an integral image to reduce the
exhaustive calculation of the rectangular regions. To construct the object classifier they used adaboost algorithm to
374 Salina Mohd Asi et al. / Procedia Computer Science 42 (2014) 372 – 377

construct a strong classifier from the weak classifiers which of small number of features selected are potentially
described the object. Its result is more robustness and computationally efficient [10-11]. Xia et al had used the Viola
and Jones framework to detect the micro facial features, the facial landmarks. They demonstrated a feature based
method to detect the eye corners. They trained two cascade classifiers of a pair of eyes images and successfully
detect both inner and external eye landmarks [12].
In this study we will demonstrate the use of HAAR Cascade Classifier to detect the inner and external eye
corners. We trained the classifier with our own data which is limited amount. The objectives of the study were to
develop HAAR cascade classifiers for the inner and external eye corners, to locate the eye corners and to measure
the craniofacial anthropometry for the orbital region the left and right eye fissures, inter-canthal and biocular width.

2. Method and materials

We photographed 100 normal adults aged between 18 to 30 years old to get 3D facial images. We used
VECTRA-3D stereo-photogrammetry from Canfield Imaging System of Canfield Science Inc.. The 3D images of
the respondents were captured under a standard clinical lighting. The system consists of four cameras and two
cameras of each side. Respondents were seated at rest position on a chair 90 cm in front of the cameras. We
calibrated the system according to the system requirement before taking the subjects three dimensional images.
We converted and exported the 3D images in Vectra-3D to get 3D-2D facial images which we refer as 2.5D
images. We divided the 2.5D facial images into seventy (70) images for training data and thirty (30) images for
testing data. We measured the 30 test images for the left and right eye fissures lengths, the inter-canthal width and
the biocular width by using the Visualizing and Measurement Software (VAM). VAM is part of the Vectra-3D
module.

1.2. The training dataset

We prepared the positive images by firstly extract the eyes from the frontal facial images. Then we prepared two
sets of positive samples each sample to train the inner (en) and external (ex) eye corner classifiers. Each set contains
300 images of size 200x 200 pixel. On each positive sample we marked the corner models. The inner eye (en) was
3/4 of the object height and 3/4 of the object width. The external eye corner (ex) was 2/3 of the object height and 3/4
of the object width. Example of the dataset as shown in fig 2.1 and the object model of en and ex eye corners are
shown in fig 2.2
Negative samples were in general images that do not contain the en and ex objects model. To improve the
accuracy of eye corner classifiers, we create the negative image by extracting any region of the size of 200 x 200
pixels from the face area which do not include the eye corners. We used Adaboost algorithm to boost these weak
classifiers consisting of selected Haar-like features and construct the cascade classifiers. In this study we constructed
the enHaar and exHaar cascade classifiers.

Figure 2 Example of training data set for inner-en (above) and external-ex (below) for left and right eye corners

3/4 2/3

3/4 3/4

Figure 3 The landmarks position of en (left) and ex (right) on the detected regions
 Salina Mohd Asi et al. / Procedia Computer Science 42 (2014) 372 – 377 375

2.3 Landmark detection and auto measurement

We applied the opencv eye classifier to search the eye at the face image. Then we used the extracted eye to
detect the en and ex. To detect the eye corners we search the input image for objects using the enHaar and exHaar
cascades classifiers. The enHaar classifier was applied to find inner eye corners while the exHaar was used to look
for external. The process of eye corner detection is that one shifts the search window at multiple resolution and
scales and checks each location using the classifier cascade.
Due to the small amount of the data used in the training session resulted in increasing of false positive
detections. To overcome this problem we developed a simple algorithm to predict the eye corners by first eliminate
the false detection by using the below algorithm.

To detect inner eye (en) To detect outer eye (ex)

- Get the boundry point, bPoint = image width/3, height/3
- Eliminate detections above the bpoint
If right eye
- Eliminate detections on the right of bPoint
- Get the most left point
Else if left eye
- Eliminate detections on the left of bPoint
- Get the most right point
- Get the boundry point – bPoint = image width/2, height/2
- Eliminate detections above the bpoint
If right eye
- Eliminate detections on the left of bPoint
- Get the most right point
Else if left eye
- Eliminate detections on the right of bPoint
- Get the most left point

Figure 4.The algorithm used to remove the false positive detections

Then the detected landmarks will be used to measure distance between the related landmarks in pixel. To
measure the intercanthal the software will measure the distance between inner corners of left and right eyes, the
biocular distance is between the outer corners of both eyes and the eye fissure length is distance between the inner
and the outer corners.

2.4 Statistical Analysis

Statistical analysis was performed using GNU PSPP. For each pair difference, normality of the distribution was
checked by plotting the histogram. Paired t test was performed to all measurements to investigate the significant
differences that exist between the Vectra 3D and auto measurement methods (measurement done by our software).
The standard descriptive analysis listed the numeric differences between the two methods in millimetre including
the mean difference and its range. The intra class correlation coefficient (ICC) was calculated for each distance to
study the degree of consistency or agreement between the paired measurements with a single index. The agreement
between Vectra 3D image system and auto measurement method are categorized as poor agreement (0.0 to 0.30),
moderate agreement (0.31 to 0.60) and good agreement (0.61 to 1.00).

3. Experimental Result

We tested our method on 60 left and right eye images. We implemented our method using OpenCV 2.4.1. Fig 3.1
includes the example detection result of the ex and en craniofacial landmarks.

Figure 5 Detection of en and ex of the left eyes

376 Salina Mohd Asi et al. / Procedia Computer Science 42 (2014) 372 – 377

Figure 6 Detection of en and ex of the right eyes

We detected ex craniofacial landmark at 59 out of 60 left and right eye images and detected ex craniofacial
landmark at 59 out of 60 left and right eye images. There was one (1) case of miss detection of en and ex at the left
and right eye. Due to this we were able to measure 29 sets of measurements for the left and right eye fissure length
(ex-ex), inter-canthal length (en-en) and biocular width (ex-ex). The landmarks detection quality is very good as per
craniofacial anthropometry landmark definition. The paired t-test analysis results are as shown in the table below.

Table 1. The mean differences, p-value and ICC between the two method of measurements

Measurement Mean diff (t) p-value ICC

Right Eye Fissure (ex-en) 0.89 0.11 0.40

Left Eye Fissure (ex-en) 0.71 0.20 0.40
Intercanthal (en-en) 0.57 0.35 0.78
Biocular (ex-ex) 1.16 0.37 0.55

Overall the measurements are not significant difference where the p-values on 95% confidence level are above
0.05. The mean differences for all measurements are below one (1) mm except the biocular measurement which is
1.16 mm. Clinically, the accepted mean difference is below than 1 mm but for some researchers below than 2 mm is
acceptable.[3] However the ICC indices for both eyes fissure lengths are moderate agreement between the two
methods. The ICC indices for both eye fissures are 40% consistency while the ICC indices show the intercanthal has
good agreement between the two methods (78%) and for the biocular the ICC index is 55% which is moderate
agreement.

4. Discussion and Conclusion

We demonstrated the used of HAAR Cascade Classifier in automatic detection of the micro facial feature. The
contributions in this study were new haar cascade classifier namely enHaar and exHaar. The classifiers successfully
detected the en and ex craniofacial landmarks in controlled environment. We are able to detect ‘en’ and ‘ex’ in 59
out of 60 images. The detection quality of en and ex craniofacial landmarks is very good. Our results show accurate
detection of ‘ex’ and ‘en’ craniofacial landmarks as per standard definition. The high detection of false positive is
due the small amount of data during the training session. Due to the small amount of data the outlier features were
sometime included to be the object features. This is because the Adaboost algorithm will select the feature if it is
occur in certain percentage in the training data for example 50% of the data.
The paired t-test showed that all four (4) p-values are more than 0.05 on 95% confidence level, therefore we
accept the null hypothesis there are no significance differences between the measurements by Vectra-3D and the
measurements by the automatic software. The ICC indices for the measurements are from 0.4 to 0.78. In conclusion,
our trained enHaar and exHaar cascade classifiers were able to automatically detect the ‘en’ and ‘ex’ craniofacial
anthropometry landmarks in controlled environment. The measurements were clinically not significant differences
with the mean different are less than 1 mm for the eye fissure lengths and the intercanthal width while for the
biocular width the mean difference is 1.16. For some researchers this value is accepted to be clinically not
significant [3]. The agreements of the measurements between the two methods are good for the intercanthal width
and moderate for the biocular width and for both eye fissure lengths.
In conclusion, our trained enHaar and exHaar cascade classifiers are able to automatically detect the en and ex
 Salina Mohd Asi et al. / Procedia Computer Science 42 (2014) 372 – 377 377

craniofacial landmarks in controlled environment. The detections are good as per standard ‘ex’ and ‘en’ craniofacial
landmarks definition. The used of small amount of trained data resulted in classifiers with high rate of false positive.
In this situation we able to remove the false positive by geometrically model the eye. The measurements between
the two methods more or less the same but the reliability and validity are moderate except the intercanthal
measurement.

Acknowledgements

The authors thank University Of Malaya for the Post Graduate Research Fund (PS168-2010B), Faculty of Dentistry,
University of Malaya, The 3D imaging lab administration for the VECTRA 3D Stereo Photogrammetry Camera that
supported by University of Malaya High Impact Research Grant (UM.C/625/HIR/MOHE/DENT/13).

References
[1] Farkas, L. G., 1994. “Anthropometry of Head and Face,” second ed., Raven Press, 1994
[2] Farkas, L. G., (1996). Accuracy of Anthropometric Measurements: Past, Present, and Future, Cleft Palate-Craniofacial Journal. 33, 10-22.
[3] Asi, S. M., Ismail N. H., Rahman, Z. A. A., (2012) Validity and Reliability Evaluation of Data Acquisition Using Vectra 3D Compare to
Direct Method. (IEEE ICBES2012) Langkawi, Malaysia.
[4] Weinberg, S. M., Scot, N. M., Neiswagner, K., Brandon, C. A. and Marazita, M. L., (2004). Anthropometric precision and accuracy of digital
three-dimensional photogrammetry: comparing the Genex and 3dMD imaging systems with one another and with direct anthropometry,”
Cleft Palate- Craniofacial Journal, 41(5), 507-518.
[5] Wong, J. Y., Oh, A. K., Ohta, E., Hunt, A. T., Rogers G. F., Mulliken J. B. and Deutsch, C. K., (2008). Validity and Reliability of
Craniofacial Anthropometric Measurement of 3D Digital Photogrammetric Images”, Cleft Palate- Craniofacial Journal, 45(3), 232-239
[6] Toma, A. M., Zhurov, A., Playle, R., Ong, E. and Richmond, S. (2009). Reproducibility of facial soft tissue landmarks on 3D laser-scanned
facial images Orthod Craniofac Res 12 :33–42
[7] Douglas TS, Martinez F, Meintjes EM, Vaughan CL, Viljoen DL. Eye feature extraction for diagnosing the facial phenotype associated with
fetal alcohol syndrome. Medical & Biological Engineering & Computing 2003; 41(1):101-06.
[8] Mutsagawa, T. and Douglas, T. S, (2007) Morphomatric analysis of facial landmark data to characterize the facial phenotype associated with
fetal alchohol syndrom, J Anat, 210, 209-220.
[9] Moore, E. S. and Ward, R. E. (2012). Chapter 63 Use of Computerized Anthropometry and Morphometrics to Identify Fetal alcoholic
Syndrome in Preedy, V. R. (ed) Handbook of Anthropometry Physical Measures of Human Form in Health and Diseases. Springer
[10] Viola, P. and Jones, M. J. (2001) Rapid Object Detection using a Boosted Cascade of Simple Features. Proc. of CVPR, pages 511-518, 2001.
[11] Lienhart, R and Maydt, J. (2002) An Extended Set of Haar-like Features for Rapid Object Detection. (IEEE ICIP), Vol. 1, pp. 900-903
[12] Xia, H., Yan, G. and You C., (2009) Feature-Based Eye Corner Detection from Static Images, Proc. of SPIE Vol. 7496, 19(1)