Accurate Facial Morphologic Measurements Using

8/3/2019 Accurate Facial Morphologic Measurements Using

http://slidepdf.com/reader/full/accurate-facial-morphologic-measurements-using 1/6

Accurate Facial Morphologic Measurements Usinga 3-Camera Photogrammetric Method

Roberto Deli, MD, DS, PhD,* Eliana Di Gioia, MD,*

Luigi Maria Galantucci, Dr, MS,Þ and Gianluca Percoco, MS, PhDÞ

Objectives: A new, low-cost photogrammetric method has been de-

veloped for facial morphometry applications. To evaluate the system,

tests for the measurement and comparison of three-dimensional vir-

tual faces were carried out in different subjects.

Materials and Methods: Twenty adult white Italian subjects,

10 men and 10 women, of ages ranging from 23 to 37 years, were

included in this study. Three cameras were finely calibrated, and the

point precision vector length was calculated, together with the qual-ity parameters. For each subject, 3 different acquisitions were per-

formed. A tessellated surface was obtained from each point cloud.

The comparison was made by aligning three-dimensional information

from different models. Differences between 2 different models were

estimated by analysis of the distances.

Results: For the cases analyzed, the mean point precision over-

all root-mean-square vector length was 0.07 mm, with a SD of

0.027 mm. The results are reported for the system’s capability of

discriminating between the faces of different people. Results of com-

parisons between facial models of a single person were compared

with those of comparisons between different subjects. Student’s

t -test revealed that the system was able to discriminate among dif-

ferent people, with a P 9 95%. Two sex subgroups were formed: the

mean error between subgroups ranged from 1.65 to 3.43 mm, and

the mean ranged from 1.76 to 2.72 mm.

Conclusions: The experiments confirmed the capabilities and the

accuracy of the proposed photogrammetric system. Facial compar-

ison was performed by analysis of distances on three-dimensional

virtual models.

Key Words: Anthropometry, facial morphometry, photogrammetry

( J Craniofac Surg 2011;22: 54 Y 59)

A detailed facial morphologic measurement acquired using three-dimensional methods is a new tool for soft tissue analysis of

faces and could be used for measurements in orthodontics. Manymethodologies have been proposed in recent years, mainly based onlaser scanning,1 Y 3 structured light,4 and stereophotogrammetry.5 Y 12

As regards the acquisition method, previous studies have al-ready been carried out by the authors, evaluating a particular pho-togrammetric approach named the 3-camera method and comparing

it with laser scanning, used as a means of performing data acqui-sition and recognition of human faces.5,6

Results obtained using the 2 methodologies are analyzed and evaluated to determine whether, even if it is a low-cost methodology,this photogrammetric technique is able to provide valid informationabout facial shape. Data processing is slow with the photogram-metric methodology, but the data acquisition, that is, the time spent taking the images, is very fast (1/5000-s flash time). This is anessential aspect because the object acquired is a live human being,whose involuntary movements of the head could make the work useless.

Although three-dimensional acquisition and recognition sys-tems have undergone major developments in the past few years,further improvements can still be achieved, such as reducing costs(currently very high) and decreasing error in the subsequent process

of creating and measuring a virtual facial model.The ultimate aim of these studies was to facilitate a clinicaldiagnosis for orthodontic/surgical purposes based on measuring not only absolute values (linear-angular distances related to a standard range of measurements) but also relative values (angles, proportions,and changing surfaces) to produce a three-dimensional soft tissuestemplate that can be related to an average normoface, which willdiffer for each population.13

In fact, the photogrammetric reconstruction of three-dimensional models of human faces provides information from two-dimensional images gained by projecting a regular grid of pointsover the face and identifying the correspondence between these pointson different photographs, to determine their spatial localization.

A process based on photogrammetry consists of 5 steps: (a)acquisition of facial images from different directions, (b) determi-nation of the camera positions and calibration parameters, (c) es-tablishment of a dense set of corresponding points in the images, (d)computation of their three-dimensional coordinates, and (e) gener-ation of a surface model.

As in data acquisition performed with a three-dimensionalscanner, the result of a photogrammetric acquisition is a point cloud,that is, a nonordered set of disconnected points in three-dimensionalspace. After obtaining point clouds, some topological information isneeded to perform the subsequent tasks; one of the most commonlyused methods for retrieving topological information after three-dimensional scanning is tessellation. This converts the given set of

points into a consistent polygonal model (mesh). The operation usu-ally generates vertexes, edges, and faces (representing the analyzed surface) that share only edges. Typically, triangles or quadrilaterals

ORIGINAL ARTICLE

54 The Journal of Craniofacial Surgery &

Volume 22, Number 1, January 2011

From the *Orthodontic Specialization School, Universita Cattolica del SacroCuore, Rome; and †Dipartimento di Ingegneria Meccanica e Gestionale,Politecnico di Bari, Bari, Italy.

Received December 22, 2009.Accepted for publication March 12, 2010.Address correspondence and reprint requests to Luigi Maria Galantucci,

Dr, MS, Dipartimento di Ingegneria Meccanica e Gestionale,Politecnico di Bari, Viale Japigia 182, 70126 Bari, Italy; E-mail:[email protected]

This research was funded by the Italian Ministry of Research and University by the Relevant National Interest Projects ProgramPRIN 2007 awarded to the Politecnico di Bari University(coordinator Prof L.M. Galantucci) and the Universita Cattolicadel Sacro Cuore di Roma (coordinator Prof R. Deli).

The authors report no conflicts of interest.Copyright * 2011 by Mutaz B. Habal, MDISSN: 1049-2275DOI: 10.1097/SCS.0b013e3181f6c4a1

Copyright © 2011 Mutaz B. Habal, MD. Unauthorized reproduction of this article is prohibited.



are used to divide the measured domain into many small two-dimensional elements. In particular, when tessellation is performed using triangles it is called triangulation and can be read and stored

by computer-aided design software using a particular solid-to-layer format.14 However, tessellated point clouds usually contain a high

quantity of data and must be appropriately managed to obtain good results within a reasonable time frame.

The equipment used for the photogrammetric acquisitionis simple and low cost; 3 digital cameras, 1 projector device, and 1 photogrammetric software are enough to obtain the three-dimensional information. The point clouds are subsequently merged (by the stitching process described by Lane and Harrell15) to obtaina single point cloud for each face. Then, tessellated and nontes-sellated point clouds are compared for facial morphologic retrieval

purposes.

MATERIALS AND METHODS

The photogrammetric system consisted of 3 digital cameras:2 Nikon Coolpix 4500 and 1 Nikon Coolpix 990 (Nikon Corp,

Tokyo, Japan). Three-dimensional data were stored and analyzed using the commercial software PhotoModeler 5.0 (Eos Systems Inc,Vancouver, Canada). To cover a wide area of the face,11 the cameraswere positioned on a circumference arc with the subject’s face in thecenter. The central camera was positioned in front of the subject; thelateral ones were positioned at the same high level but at a 30-degreeangle from the horizontal plane8 (Figs. 1 and 2).

The reconstruction of a three-dimensional face model wasobtained from the acquisition of 3 simultaneous images, each one

taken from a different viewpoint. The three-dimensional image ac-quisition of human faces is more critical as compared with imageacquisition of a static object: in fact, it is necessary to freeze motion,that is, to avoid breathing and moving artifacts.12 If the images arecaptured at different instants, there could be errors owing to major movements (changing the head position) or minor movements(muscle activity, skin or hair surface variation, and changes due to

breathing).3,16 Furthermore, any movements could cause errors, dueto shifting of the grid projected onto the face. Therefore, a good three-dimensional reconstruction can actually be performed onlywith simultaneous image acquisition.10

In previous studies, the authors proposed a completely auto-mated version of the three-dimensional photogrammetric scanningsystem, also based on the projection of hybrid and regularly spaced

FIGURE 1. Positions of the cameras. Points A, B, and Ccorrespond to the positions of the digital cameras, inferredduring the orientation process.

FIGURE 2. Scanning system.

FIGURE 3. Example of a tessellated surface obtained from adense point cloud.

FIGURE 4. Mean error (A) and SD (B) in tessellated-tessellated(white columns) and tessellated Y point cloud (white columns)comparisons of the subject male subject 9. Units are incentimeters.

The Journal of Craniofacial Surgery & Volume 22, Number 1, January 2011 Morphologic Photogrammetric Measurements

* 2011 Mutaz B. Habal, MD 55




grids,8 to reproduce the whole human face. Using this method, it is possible to acquire information about the characteristics of thesubject’s soft tissues and to make exact measurements.

In the current study, during the acquisition sessions, thesubjects were asked to assume a natural head position and a naturalexpression of the face. Three acquisition sessions were performed

for each person in a completely dark environment; the study popu-lation consisted of 20 people of ages ranging between 23 and 37 years, 10 men and 10 women (in the male group, there were2 homozygous twins). The first and second acquisitions were bothmade by projecting 4316 points onto the face.17

The photographs related to the same person were taken on thesame day. Therefore, although time could be a factor affectingresults,12 in this case, it did not need to be taken into account. After a previous calibration and orientation of the cameras, the photo-grammetric software PhotoModeler was used to process data related to the reference points and calculate their spatial localization.

For each person analyzed, 2 models related to the full surfaceface (one for each acquisition session) were created on the basis of the 4316-point grid 17 (example in Fig. 3).

RESULTSTo achieve a high-accuracy photogrammetric project, a

camera calibration process is needed. This is used to insert camerainformation about focal length, lens distortion, format aspect ratio,and principal points. The cameras were calibrated, obtaining a mean

point precision vector length with an overall root-mean-square(RMS) equal to 0.08 pixels, and a point precision vector length with

FIGURE 5. Mean error (A) and SD (B) in tessellated-tessellated(black columns) and tessellated Y point cloud (white columns)comparisons of the subject female subject 3. Units are incentimeters.

FIGURE 6. Three-dimensional color map comparison between one male subject’s face and all the other male subjects’ faces.

Deli et al The Journal of Craniofacial Surgery & Volume 22, Number 1, January 2011

56 * 2011 Mutaz B. Habal, MD




an overall RMS equal to 0.085 mm. PhotoModeler can establish thequality parameters of the project. The accuracy of the system wastested in a previous work.18 For the cases analyzed, the mean point

precision overall RMS vector length was 0.07 mm, with an SD of

0.027 mm (maximum, 0.118 mm; minimum, 0.025 mm).Two kinds of comparison were performed: the first one was

made between 2 tessellated models, and the second one between atessellated model and a point cloud. The latter was done to test thevalidity of the results when a tessellated model is compared with a

point cloud, so as to speed up the entire process by eliminating theneed for tessellation of each point cloud. The comparison was made

by aligning the three-dimensional information from different mod-els. Models of the entire facial surface were compared using thesame commercial software (http://www.geomagic.com/en/products/ studio/) used for the tessellation process. Once the best solution for the alignment of models had been chosen, differences between 2different models were estimated by performing an analysis of dis-tances. This analysis yielded 2 indicative numerical parameters: (1)mean value of distances between points belonging to the 2 models

and (2) SD, calculated on all distances between the models.Results of comparisons between facial models of a single

person were compared with those of comparisons obtained betweenfacial models belonging to different subjects. In this way, it was

possible to test the robustness of the comparison procedure in dis-criminating whether 2 three-dimensional facial models belong to thesame person.

Two examples of the resulting data are reported, one for eachsex, in Figures 4 and 5. The results were similar for all the other subjects. Mean error and SD resulting from both kinds of comparisonare shown in each histogram.17 The horizontal lines represent themaximum value obtained when comparing models of the same per-son (value related to tessellated-tessellated comparison of models

belonging to male subject 7). Data from tessellated-tessellated comparisons correspond to the blue columns, whereas results of

tessellated Y point cloud comparisons are shown in yellow.

DISCUSSION

The validity of the system for analyzing facial morphometryis described for its capability of discriminating between different faces. It is noticeable that the minimum error of comparison wasalways obtained for facial models belonging to the same person.This error is comparable with the mean shell deviation found in

Kau et al,19 where a laser imaging system was used. Moreover, thevalues obtained in comparisons of models belonging to 2 different subjects greatly exceeded in each case the maximum value obtained in comparisons of models related to the same person, marked bythe horizontal red line (Figs. 4 and 5). This demonstrates the validityof the comparison methodology for measuring facial morphologicfeatures. This seems to be partially in contrast with findings byBozic et al20 in which the mean linear facial differences between 4sex-specific subgroups of 2 white European groups ranged from0.36 to 1.51 mm, but in that work, a laser scanner system was used,which is less precise than photogrammetry (T0.5 vs 0.07 mm).

No significant differences were found between the resultsobtained with the 2 kinds of comparison (tessellated-tessellated and tessellated Y point cloud). The 2 techniques yielded very similar error values. As a consequence, because more time is needed to process

2 tessellated models, the tessellated Y point cloud comparison could prove to be more convenient to gain faster results.

In Figure 6, colored maps of the comparison between onemale and the other male subjects is shown; green areas indicate agood superimposition with a distance of less than 0.5 mm betweenthe clouds, whereas the other colors define greater distances.Comparing these results with Figure 7, where 2 point clouds be-longing to the same people are compared, it is evident that thesystem is able to discriminate among different subjects correctly.

The system’s ability to distinguish whether subjects were of the same sex was investigated, by computing the mean values of themean error and SD values for each kind of comparison: male-male,female-female, and male-female.

Tests were performed on the facial models of the entire face because such results were considered more reliable in view of the

previous considerations. Table 1 shows the mean values of comparison errors between tessellated models. Values deriving from thecomparison between a tessellated model and a point cloud yield much the same kind of results.

The values obtained in comparisons between male subjectsshow 20% larger errors than the values obtained in comparisons

between female subjects. This result is even more important whenconsidering the mean values of the mean error and SD valuesobtained in comparisons between models belonging to the same

person, whose values seem to be independent of their sex.It can be observed that the system is able to identify larger

differences between male facial models, in agreement with theresults found in Baik et al,1 where a test result showed that allrecognition systems were better able to identify men than women.The data obtained are comparable with those in Seager et al,21 where

a photogrammetric system was used to compare facial morphologies

FIGURE 7. Three-dimensional color map comparisonbetween 2 point clouds obtained for the male subject inFigure 6.

TABLE 1. Mean Values of Mean Error (Millimeters) and SD (Millimeters) Obtained in Comparisons Between Sexes(Comparison Between Tessellated Models)

Mean Value of

Mean Error, mm

Mean Value of Mean Error

in Comparisons Related to

the Same Person, mm

Mean Value

of SD, mm

Mean Value of SD in

Comparisons Related to

the Same Person, mm

Female-female 1.65 0.45 Female-female 1.76 0.48

Male-male 2.06 0.40 Male-male 2.19 0.50

Male-female 3.43 0.43 Male-female 2.72 0.49






from 2 population groups, even if a less accurate system was used in that study. Major enhancements are also gained as compared with recent studies22 where the same software is used but with amanual approach that leads to a much lower accuracy of the digitiza-tion than that obtained with the automated system described in the

present work.Errors resulting from comparisons between male and female

subjects were approximately 40% larger than those obtained be-tween male subjects. This shows the system’s ability to distinguish

individuals of different sexes.Moreover, the mean value of the mean errors were tested for

normality, and differences between the groups measured were ana-lyzed using Student’s t -test, considering a limit value of P = 0.05.In particular, paired t -tests carried out on the mean value of themean errors obtained comparing 2 different point clouds belongingto the same person, subdivided into sex subgroups, revealed no sig-nificant differences between sexes ( P 9 0.05), pointing out that thedigitizing system can recognize both men and women. This canseem to be in contrast with the results found in Baik et al,1 but although it is stated in that study that the differences betweenwomen are higher than the differences between men, the t -test reveals that when the laser scanner digitizes the same person, themean errors are similar for men and women, thus allowing thesystem to perform facial recognition on women and men.

Another t -test was made to find the mean value of the meanerrors between 2 groups, consisting of one-to-one comparisons re-lated to (1) 2 point clouds for the same person and (2) 2 point cloudsfor 2 different people. The test revealed that the probability of 2sample groups belonging to the same population is lower than 5%(limit value of P = 0.05), so the system is able to discriminatedifferent people with a P 9 95%.

Among the 20 subjects participating in the test, there were2 male homozygous twins: male subjects 6 and 8. Important dif-ferences can be seen between the 2 subjects as regards comparisonof the entire facial surface. At the same time, the mean error and the SD obtained in the comparison between male subjects 6 and 8 are definitely lower than those obtained in all the other tests madeto compare data from 2 different subjects (Table 2). On the other hand, the comparison done only on landmarks extracted from the

three-dimensional tessellated model was not able to distinguish between the 2 twins. In fact, the mean error results lower whencomparing male subjects 6 with 8 than when comparing male sub-

jects 6 and 6 and male subjects 8 and 8.

CONCLUSIONS

& The proposed acquisition and reconstruction methodology can provide three-dimensional virtual models of the entire face.

& The performance of the digital photogrammetric system designed for facial features measurement purposes was successfully as-sessed and demonstrated its validity as compared with other ex-

periences already reported in the literature.& The system is accurate (T0.07 mm).

& The digitizing system is able to discriminate between different people with a P 9 95% and allows facial recognition of womenand men to be performed.

& The only requirement is that the subject adopts a natural ex- pression while the photographs are taken.

ACKNOWLEDGMENT

The authors thank Dott. Ing. Nicola Mastronardi for con-

tributing to the execution of the experimental tests.

REFERENCES

1. Baik H, Jeon J, Leeb H. Facial soft-tissue analysis of Korean adults withnormal occlusion using a 3-dimensional laser scanner. Am J Orthod

Dentofacial Orthop 2007;133:612 Y 620

2. Kau C, Richmond S, Zhurov A, et al. Reliability of measuring facialmorphology with a 3-dimensional laser scanning system. Am J Orthod

Dentofacial Orthop 2005;128:424 Y 430

3. Kovacs L, Zimmermann A, Brockmann G, et al. 3D recording of thehuman face with a laser scanner. Plast Reconstr Surg 2006;59:1193 Y 1202

4. )arnick ] J, Chorvat D Jr. Three-dimensional measurement of facewith structured-light illumination. Meas Sci Rev 2006;6:1 Y 4

5. Ayoub AF, Siebert P, Moos KF, et al. A vision-based three-dimensional

capture system for maxillofacial assessment and surgical planning. B J Oral Maxillofac Surg 1998;36:353 Y 357

6. Ferrandes R, Galantucci LM, Percoco G. Optical methods for reverseengineering of human faces. 4th International CIRP Design Seminar,May 16 Y 18, Cairo, Egypt, 2004;6B:1 Y 12

7. Galantucci LM, Percoco G, Dal Maso U. Coded targets and hybrid gridsfor photogrammetric 3D digitisation of human faces. Virtual Phys

Prototyping 2008;3:167 Y 176

8. Di Gioia E, Deli R, Galantucci LM, et al. Reverse engineering and photogrammetry for diagnostics in orthodontics. J Dent Res 2008;87:1620

9. Winder RJ, Darvann TA, McKnight W, et al. Technical validationof the Di3D stereophotogrammetry surface imaging system. B J Oral

Maxillofac Surg 2008;46:33 Y 37

10. Ayoub AF, Xiao Y, Khambay B, et al. Towards building a photo-realisticvirtual human face for craniomaxillofacial diagnosis and treatment planning. Int J Oral Maxillofac Surg 2007;36:423 Y 428

11. Khambay B, Nairn N, Bell A, et al. Validation and reproducibility of a high-resolution three-dimensional facial imaging system. B J Oral

Maxillofac Surg 2008;46:27 Y 32

12. Deli R, Di Gioia E, Galantucci LM, et al. Automated landmarksextraction for orthodontic measurement of faces using the 3 cameras photogrammetry methodology. J Craniofac Surg 2010;21:87 Y 93

13. Kau CH, Hunter LM, Hingston EJ. A different look: 3D facial imagingof a child with Binder syndrome. Am J Orthod Dentofacial Orthop2007;132:704 Y 709

14. Remondino F. From point cloud to surface: the modeling and visualization problem. International Archives of Photogrammetry,Remote Sensing and Spatial Information Sciences, Vol XXXIV-5/W10,2003. Available at: http://www.photogrammetry.ethz.ch/general/ persons/fabio/tarasp_modeling.pdf. Accessed November 4, 2009

TABLE 2. Mean Error (Millimeters) and SD (Millimeters) for Comparison Between Twins

Subjects

Comparison

Tessellated Y Point Cloud

Comparison

Tessellated-Tessellated

Comparison

Between Landmarks

Mean Error SD Mean Error SD Mean Error

Male 6 Y male 6 0.53 0.41 0.52 0.63 1.71

Male 8 Y male 8 0.25 0.38 0.27 0.40 1.57

Male 6 Y male 8 1.14 1.23 1.05 1.18 1.55

Deli et al The Journal of Craniofacial Surgery & Volume 22, Number 1, January 2011

58 * 2011 Mutaz B. Habal, MD




15. Lane C, Harrell W. Completing the 3-dimensional picture. Am J Orthod Dentofacial Orthop 2008;133:612 Y 620

16. Sforza C, Peretta R, Grandi G, et al. 3D facial morphometry inskeletal class III patients: a non-invasive study of soft-tissue changes before and after orthognathic surgery. B J Oral Maxillofac Surg 2007;45:138 Y 144

17. Galantucci LM, Ferrandes R, Percoco G. Digital photogrammetry for facial recognition. J Comput Inf Sci Eng 2006;6:390 Y 396

18. Galantucci LM, Percoco G, Ferrandes R. Accuracy issues of digital photogrammetry for 3D digitization of industrial products. Rev Int Ingegnerie Numerique 2006;2:29 Y 40

19. Kau CH, Richmond S, Savio C, et al. Measuring adult facialmorphology in three dimensions. Angle Orthod 2006;76:773 Y 778

20. Bozic M, Kau CH, Richmond S, et al. Facial morphology of Slovenian and welsh white populations using 3 dimensional imaging.

Angle Orthodontist 2009;79:640 Y 645

21. Seager CC, Kau CH, English JD, et al. Facial morphologies of anadult Egyptian population and adult Houstonian white populationcompared using 3D i maging. Angle Orthod 2009;79:991 Y 999

22. de Menezes M, Rosati R, Allievi C, et al. A photographic system for the three-dimensional study of facial morphology. Angle Orthod 2009;79:1070 Y 1077



Accurate Facial Morphologic Measurements Using

Documents

Transcript of Accurate Facial Morphologic Measurements Using