University of Nebraska - Lincoln

DigitalCommons@University of Nebraska - Lincoln

1-1-2006

Computation of a face attractiveness index based on neoclassical canons, symmetry, and golden ratios
Kendra Schmid
University of Nebraska - Lincoln

David Marx
University of Nebraska - Lincoln, [email protected]

Ashok Samal
University of Nebraska - Lincoln, [email protected]

Follow this and additional works at: http://digitalcommons.unl.edu/statisticsfacpub Part of the Statistics and Probability Commons
Schmid, Kendra; Marx, David; and Samal, Ashok, "Computation of a face attractiveness index based on neoclassical canons, symmetry, and golden ratios" (2006). Faculty Publications, Department of Statistics. Paper 12. http://digitalcommons.unl.edu/statisticsfacpub/12

This Article is brought to you for free and open access by the Statistics, Department of at DigitalCommons@University of Nebraska - Lincoln. It has been accepted for inclusion in Faculty Publications, Department of Statistics by an authorized administrator of DigitalCommons@University of Nebraska - Lincoln.

Department of Statistics Technical Reports, University of Nebraska, 2006.

Computation of a face attractiveness index based on neoclassical canons, symmetry, and golden ratios Kendra Schmid1, David Marx1 and Ashok Samal2
1

Department of Statistics, 2Department of Computer Science and Engineering University of Nebraska-Lincoln Lincoln, NE 68588-0115

Abstract Analysis of attractiveness of faces has long been a topic of research. Literature has identified many different factors that can be related to attractiveness including symmetry, averageness, sexual dimorphism and adherence to the Golden Ratio. In this research we systematically analyze the role of three factors: symmetry, conformance to Neoclassical Canons and the Golden Ratio in the determination of attractiveness of a face. Unlike many researchers, we focus on the geometry of a face and use actual faces in standardized databases for our analysis. Our results are in agreement with the literature in that males and females generally agree on which faces are viewed as attractive and which are not. However, there are some differences in the criteria used by males and females to determine attractiveness. Using statistical analyses, we have developed a model to predict the attractiveness of a face using its geometry. The results show that our model is accurate with low residual errors.

Key Words: Face Attractiveness, Face Recognition, Neoclassical Canons, Face Symmetry, Golden Ratio

1. Introduction A popular axiom concerning physical attractiveness is: Beauty is in the eye of the beholder. Research in the area of facial perception has identified many different factors that contribute to a face being considered attractive. Armstrong (2004) suggests that beauty cannot be defined by one single principle. Rhodes (2006) focuses on averageness,

symmetry, and sexual dimorphism and their link to facial attractiveness. Little et al. (2000) suggests that self-perceived attractiveness influences ones opinion of the attractiveness of others, and DeBruine (2004) shows both males and females prefer faces that resemble their own.

In this paper, we develop a quantitative method for measuring facial attractiveness using a combination of several factors that have been deemed significant in previous research. Many previous studies have used composite faces (combinations of several faces) or faces that are altered in some other way to study the effects of symmetry and averageness on attractiveness (Kowner 1996, Langolis 1990, Langolis et al. 1994, Little & Hancock 2002, OToole et al. 1999, Perrett et al. 1998, Perrett et al. 1999, Rhodes 2006, Rhodes et al. 1998, Rhodes et al. 1999, Rhodes & Tremewan 1996, Swaddle & Cuthill 1995). In contrast, we use the actual faces compiled from a standard face recognition database for our analysis. We then determine the location of important landmarks in the face (Farkas 1994, Farkas & Munro 1987). In all, we use 29 landmarks on each face as described by Shi et al. (2006) to take physical measurements. Using these landmarks, we compute the values of three factors: Neoclassical Canons, symmetry, and golden ratios. These faces are presented to a set of human subjects to determine their perceived attractiveness using a partially balanced incomplete block design to find which of these or which combination of these is the best predictor of attractiveness. Using a statistical approach, we systematically investigate the relationship between the facial measurements in the images and the attractiveness ratings given by the human participants.

In addition to which measure (canons, symmetry, or golden ratios) best predicts attractiveness, we are able to identify which features play the greatest role in attractiveness and if these features are common across both genders of raters and images. Rhodes (2006) suggests that it may be wise to distinguish same sex ratings from opposite sex ratings. We record the gender of the rater and the image to explore the differences or similarities in how males and females view attractiveness in images of the same and opposite gender.

We describe the image database used for our research in Section 2. We also explain how the measurements are obtained and the design of our human studies experiment. In Section 3 we discuss the methods to compute the attractiveness predictors from face images. The predictors are based on Neoclassical Canons, symmetry, and golden ratios, which are described in literature. However, we provide new approaches to compute these predictors. In Section 4, we present the details of our statistical analyses and the associated results. The features in the face that are the best predictor of attractiveness are described in this section. Finally, we summarize our findings and identify areas for future work in Section 5.

2. Datasets and Experimental Design We begin with an image database containing a set of face images for our experiment and analysis. Using the image database, we compile two datasets for our analysis. The feature dataset consists of the locations of the landmarks in the faces. The attractiveness dataset contains the attractiveness ratings given to the images by the human participants.

We briefly describe the image database and the process of creating the features and attractiveness datasets in this section.

2.1.

The Image Database

The majority of the images used in this research were taken from the Facial Recognition Technology (FERET) Database (Phillips et al. 2000). The FERET Database contains some fourteen thousand facial images that were collected by photographing over one thousand subjects at various poses. The images were collected over the course of fifteen sessions between 1993 and 1996. Some individuals were photographed multiple times, sometimes with more than two years separating their first and last sitting.

Image Selection and Filtering: For our research, we choose only the images that showed a full frontal view of a face with little or no facial expression. All non-Caucasian images and faces with glasses were removed to reduce the variability among images. The remaining images were converted to gray scale to reduce the effects of skin color. At the end we selected a set of 420 unique images with equal numbers of males and females. Figure 1 shows some sample face images from this set.

Figure 1: Sample images taken form the FERET database

In addition to the images from the FERET Database, we used the images of 32 popular movie personalities ranging from the 1930s until the present day (Movie Actor Index). Again, an equal number of male and female faces were used. The personalities were chosen to include only those that were considered to be attractive. Our motivation for including such faces, deemed more attractive than the norm by the society, was to verify our system of rating faces; the ratings given to these faces would be expected to be significantly higher than the ratings given to the faces of non-famous people. Figure 2 shows some sample face images from this set.

Greta Garbo

Meg Ryan

Rock Hudson

Keanu Reeves

Figure 2: Sample images of faces known to be attractive.

2.2.

The Feature Point Database

We developed a tool to derive measurements from a set of standard face images using a graphical user interface. A set of 29 important landmarks was identified based on existing literature (Farkas 1994). A reference image showing the location of the features and the test image is presented. The user is prompted to locate the corresponding feature in the test image. The user indicates the location of the feature by a mouse click. Table 1 provides the description of each of these feature points. Using this tool we extracted all

29 feature points from each of the 452 face images. Figure 3 shows the layout of the landmarks on a face image. The feature point database consists of the locations of the feature points for the faces from the FERET database and the faces of famous people.

Feature Point 01 02 03 04 05 06 07 08 09 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29

Feature Description The point on the hairline in the midline of the forehead Highest point on upper borderline in mid portion of left eyebrow Most prominent midline point between eyebrows Highest point on upper borderline in mid portion of right eyebrow Highest point on the free margin of left ear Most prominent lateral point on left side of the skull Highest point on lower border of left eyebrow Highest point on lower border of right eyebrow Most prominent lateral point on right side of the skull Highest point on the free margin of right ear Point at outer right side of the eye. Point at inner right side of the eye Point at inner left side of the eye Point at outer left side of the eye. Lowest point on lower margin left eye Lowest point on lower margin right eye Lowest point of left ear Most lateral point on left side of nose Midpoint of nose Most lateral point on right side of nose Lowest point of right ear Highest point on left side of lip Midpoint on upper lip Highest point on right side of lip Left most point of closed lip Midpoint of closed lip Right most point of closed lip Point on lower border of lower lip or upper border of chin Tip of chin
Table 1: Description of feature points

Figure 3: Feature points on an image

2.3.

Attractiveness Scores Database

In order to get some ground truth on the attractiveness of faces in our database, we compiled their attractiveness scores using human subjects. The human subjects were asked to rate the faces in the database using a ten-point scale. We use these scores to build a database of the attractiveness for the faces.

Design of Experiment: Asking a subject to rate all 452 faces would not only take a long time, but would reduce the quality of the results. Therefore, we chose a partially balanced incomplete block design for this process. The 420 FERET images were split into six groups of 70 images, labeled A, B, C, D, E, and F with each group consisting of 35 males and 35 females. In addition, each group has a total of 30 duplicate images with fifteen male duplicates and fifteen female duplicates, for a total of 100 images per group. Each participant was assigned to rate two of these groups. Including the duplicates

provides a way to check the consistency within each rater. In addition to these 200 images, each subject was asked to rate each of the famous faces. Thus, each subject gave ratings to 232 faces in all.

We chose a partially balanced design instead of a balanced design because the latter would require fifteen raters which would not allow for an equal number of male and female raters. In our design, each rater is assigned two of the six groups of images so that no two groups appear together more than once and so each group is shown to two male and two female raters. Table 2 shows the layout of the design. Participant ID 1 2 3 4 5 6 7 8 9 10 11 12 Gender M M M M M M F F F F F F Groups Shown A, C B, F D, E A, D B, C E, F A, E B, D C, F A, F B, E C, D

Table 2: Partially Balanced Incomplete Block Design

Participant Selection and Data Collection: Twelve participants (raters) were chosen from students and employees at the University of Nebraska-Lincoln and ranged in age from 19 to 61 years. An equal number of male and female participants (six of each) were chosen to rate the faces. Using twelve participants allowed for the partially balanced incomplete block design while maintaining an equal number of male and female raters. The gender and age of each participant were collected and stored in the database.

Each of the six groups of images was shown to four raters, two males and two females. Each rater viewed two of the six groups of images in a random sequence and the experiment was designed so that no raters viewed the same two groups. After viewing the 200 images, each rater viewed the 32 famous images. Each participant rated each face image on a scale from 1 (least attractive) to 10 (most attractive) based on his or her opinion. In addition to the score given by the rater, we record the time taken to give the ratings. After rating the 232 images, the participant is given the option to rate him or herself using the same 10 point attractiveness scale.

The experimental design was repeated three times, using different raters in each instance, for a total of 36 raters. Within each group of twelve raters, each image was shown four times, each duplicated image was shown eight times, and each famous image was shown twelve times.

Data Filtering: After the attractiveness scores were collected, we examined the integrity of the data. We calculated the variance of the ratings given by each user. A variance of zero indicates that a participant gave the same rating to each face and hence the scores of this rater should be ignored. We also examined the raters whose scores had very large or very small variances. A very small variance could indicate a participant going back and forth between two consecutive numbers. A very large variance could indicate a pattern where a participant goes back and forth between two numbers such as 1 and 10, or just goes up and down from 1 to 10. We also examined the time spent to rate the faces by each participant. A very small amount of time to complete the rating process could

indicate a participant just clicking numbers and not actually rating while a very large amount of time might mean that the participant was interrupted during the rating process. If a rater was flagged for any of the above problems, the corresponding ratings (and the time taken to rate) were examined more carefully. The scores that did not have a high level of integrity were removed.

3. Computation of Attractiveness Predictors The main motivation of our research is to examine the attractiveness of a face, Fi, as a function of its face geometry. The geometry of a face is captured by a set of m landmarks. Thus: Fi = {fi1, fi2, , fim}, where each feature point is represented by its two dimensional spatial coordinates in the face. Thus, fij = (xij, yij), 1 i n, 1 j m. Our goal is to determine a function A that maps a face to an attractiveness score. A(Fi) [1,10] To compute the attractiveness, we use three predictors that have been proposed in literature: Neoclassical Canons, Face Symmetry, and Golden Ratios. We discuss each of them below.

3.1.

Neoclassical Canons

Neoclassical Canons view the face in proportions and have been proposed by artists dating back to the renaissance period as guides to drawing beautiful faces (Farkas et al.

10

1985). The basic premise is that portions of an attractive face should follow certain defined ratios. Farkas et al. (1985) summarizes these principles in nine Neoclassical Canons and their variations. Four of the canons deal with vertical measurements, four with horizontal measurements, and one with angles of inclination. Only six of these can be tested from the frontal views of the images. Therefore, only those six canons (listed in Table 3) are used to investigate their relationship with attractiveness of a face.

Formula No. 2 4 5 6 7 8

Description Forehead height Nose length Interocular distance Interocular distance Mouth width Face width

= Nose length = Lower face height = Ear length = Nose width = Right or left eye fissure width = 1.5 Nose width = 4 Nose width

Table 3: Description of Neoclassical Canons (Formula number given by Farkas (1985))

As shown in Table 3, some canons use two measurements (e.g. Formula 4 and 8) while others use three (e.g. Formula 2, 6). To consistently measure compliance with the canons (i.e. equality to proposed ratios) with different numbers of features, we use the coefficient of variation. The coefficient of variation is defined as the ratio of the standard deviation of the distances to the mean of the distances. For a canon with three distances, using a ratio would require pair-wise comparisons of these distances, but using the coefficient of variation allows us to incorporate all three distances into one value while adjusting for the size of the face (dividing by the mean). A value of zero for the coefficient of variation says there is no variation in the distances (they are equal). For non-zero values, the larger the value, the more the face differs from the canon. Using this approach, we compute the degree of match with each canon (i.e. coefficient of variation) for all the faces and store it in a database.

11

3.2.

Symmetry

Symmetry for a face is considered to be an important factor for attractiveness (Rhodes 2006). Symmetry has been defined in many different ways (Grammer & Thornhill 1994, Kowner 1996, Little & Jones 2006, Penton-Voak et al. 2001, Perrett et al. 1999, Rhodes et al. 1998, Rhodes et al. 1999, Rhodes 2006, Samuels et al. 1994, Scheib et al. 1999, Swaddle & Cuthill 1995); however, many consider only the symmetry about a vertical axis. We define the axis of symmetry to be located vertically at the middle of the face. We determine this line by fitting the least squares regression line through the seven points measured along the middle of the face (Points 1, 3, 19, 23, 26, 28, 29 shown in Figure 3). We use the following feature pairs (left and right) for our analysis of symmetry. Eyebrows (Points 2 and 4; Points 7 and 8) Eyes (Points 11 and 14; Points 12 and 13; Points 15 and 16) Nose (Points 18 and 20) Ears (Points 5 and 10; 17 and 21) Lips (Points 22 and 24; 25 and 27), Face (Points 6 and 9)

In order to compute the symmetry of a face about the vertical axis, and assuming that there are p symmetric pairs of feature points, a face, Fi, can be represented as: Fi = {sfi1, sfi2, , sfip}, where sfij is a symmetric pair of feature points (left (L) and right (R)) represented by: sfij = < fijL , fijR>, fijL Fi, fijR Fi, 1 I n, 1 j p and each feature point is given by its two dimensional coordinates on the face:

12

fijL = (xijL, yijL) and fijR = (xijR , yijR). If we assume that the line of symmetry is vertical, a feature pair is perfectly symmetric if yijL = yijR and xs- xijL = xijR- xs, where xs=0 represents the vertical line of symmetry.

( xijR , yijR )

( xijL , yijL )

hijR

hijL dijL

( xijC , yijC )

dijR

( xijR , yijL )

l
Figure 4: Line of symmetry (l), a pair of symmetric feature points, and angles

To compute the symmetry of a face, we first compute the symmetry of the individual features. Symmetry can be computed using many different rules. Literature in sexual size dimorphism (SSD) (Smith 1999), for example, has identified many formulas to compare the measurements for males and females. SSD uses differences or ratios of features to help in determining the degree of difference between male and female measurements. In our case, we use some of the same indices to determine the degree of difference between

13

the left and right side of a face. We refer to these as facial symmetry measures (FSM) as they are functions of the perpendicular distance (d) from a given feature point to the line of symmetry as shown in Figure 4. We define four different functions to compute the face symmetry measure as shown below.

Difference:

FSM Diff (d ) = dijL dijR

Ratio:

LN(Ratio):

d FSM LNRatio (d ) = ln ijL d ijR

dijL dijR Adjusted Difference: FSM AdjDiff (d ) = (dijL + dijR ) / 2

These measures, when used in SSD, determine the degree of difference between males and females. In this application, we use them to determine the degree of asymmetry in the left and right sides of a face. For the difference, adjusted difference, and natural log of the ratio, a value of zero implies symmetry; the greater the value the less symmetric the face. For the ratio, a value of one indicates symmetry.

The above measures are useful in measuring degree of symmetry when the face is vertical. However, when the face is titled or rotated about the vertical axis, this measure is not accurate. For example, if one ear is significantly higher than the other but both are the same distance from the line of symmetry, these points would be considered symmetric by using only horizontal distances. For this reason, we hypothesize that
14

incorporating both angle and distance into the measure of symmetry will result in better predictions of attractiveness scores than by using distances alone. Figure 4 shows a pair of feature points (xijL , yijL) and (xijR , yijR) with the line of symmetry, l. The points are considered symmetric if = 0 and dijL=dijR. The angles are calculated as follows:

= tan 1

y ijR yijL . d ijL + d ijR

We dont consider a vertical distance measure, even though like angles, it is not affected by slight tilts in the face. The vertical distance can be calculated by simply subtracting the y-coordinates of the paired feature points and using the absolute value of that difference as a measure of symmetry, |yijL-yijR|. The problem, however, is that it is only a single value for each pair of feature points that will be dependent on the size of the face. When using horizontal distance, there are two measures, dijL and dijR, for each pair of feature points. There is no obvious line of reference for vertical distances like the line of symmetry when using horizontal distances. Furthermore, using vertical measures, in addition to angles and horizontal distances, does not result in increased information, since the three measures are related. Therefore, we use only the horizontal distances and angles to compute the symmetry of a face. Together they measure both the horizontal and vertical symmetries in the face.

3.3.

Golden Ratios

While there is no systematic published study that shows any correlation between attractiveness and proportions in face measurements that approach the Golden Ratio, such

15

relationships have been reported in popular literature (Meisner 2006, Narain 2003). According to these reports, faces that have features with ratios close to the Golden Ratio are thought to be aesthetically pleasing. To determine the validity of this claim, we systematically analyzed all the ratios in the face that can be determined from the set of 29 feature points we have identified in the face. Using all ratios derived from using the pairwise horizontal and vertical distances all 406 possible pair-wise combinations of the 29 feature points results in 659,344 ratios per face. Each ratio was then averaged over all of the images and standardized (subtract the Golden Ratio and divide by the standard deviation of that ratio over all images) to see if it is close to the Golden Ratio. A standardized value of zero indicates that when averaged over the images, that ratio is close to the Golden Ratio. Using only ratios with standardized values less than 0.001 away from zero resulted in a set of 70 ratios. These ratios are likely to be closest to the Golden Ratio. However, a detailed analysis of the features showed that they do not have good intuitive descriptions (e.g. ratio of the horizontal distance from point 2 (top of eyebrow) to 17 (bottom of ear) with the vertical distance from point 4 (top of eyebrow) to 14 (outer corner of eye)). Therefore, they were dropped from further analysis.

We also analyze a set of ratios defined by Meisner (2006) and Narain (2003) as being equal to the Golden Ratio and as related to attractiveness of a face. With the points available in this study, there are seventeen ratios used to explore their relationship to attractiveness. Table 4 describes these ratios and identifies the points used for each, where x or y refers to the x-coordinate or y-coordinate of the points and the numbers indicate which points from Figure 3 were used in calculating the ratio.

16

Ratio Number 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17

Numerator Points y10-y21 y10-y21 x15-x16 x15-x16 x25-x27 y23-y29 y23-y29 x12-x13 x12-x13 x18-x20 x18-x20 x12-x11 y23-y28 y1-y29 y19-y29 x18-x20 x25-x27

Denominator Points x12-x13 x18-x20 x12-x13 x18-x20 x12-x13 x12-x13 x18-x20 x12-x11 y23-y28 x12-x11 y23-y28 y19-y26 y19-y26 x17-x21 y26-y29 y19-y26 x18-x20

Description Ear length to Interocular distance Ear length to Nose width Mideye distance to Interocular distance Mideye distance to Nose width Mouth width to Interocular distance Lips - chin distance to Interocular distance Lips chin distance to Nose width Interocular distance to Eye fissure width Interocular distance to Lip height Nose width to Eye fissure width Nose width to Lip height Eye fissure width to Nose mouth distance Lip height to Nose mouth distance Length of face to Width of face Nose chin distance to Lips chin distance Nose width to Nose mouth distance Mouth width to Nose width

Table 4: Golden Ratios obtained from Meisner 2006 and Narain 2003

4. Analyses and results

We begin with the examination of a set of general questions about the attractiveness of human faces. First the variability in the raters as a function of both the gender of the rater and the gender of the face is examined. Then we examine if the self-perceived attractiveness has any effect on the ratings given by the rater as proposed by Little et al. (2001). Finally the relationship between the time taken by the rater and the ratings given to the faces is analyzed. Later we examine in depth the roles of the three predictor variables: Neoclassical Canons, symmetry and the Golden Ratio, used in this research. We conclude this section by analyzing how the three predictor variables can be combined to develop a predictive model to determine attractiveness. For all our analysis we use the SAS statistical analysis software (SAS Institute 2003).

17

4.1 Do males and females rate faces differently?

It has been reported in literature that males and females generally agree on attractiveness (Langlois et al. 2000). In order to examine this systematically, we carried out an analysis of variance (ANOVA) to determine if there is a difference in ratings given by men and ratings given by women. We also wanted to examine if the gender of the face had an impact on the rating, and if the ratings given by males and females were consistent for male and female faces. In this analysis, the response (dependent) variable is the average rating (AR) of the image by each participant. The ratings of duplicate images were averaged for each rater. The following statistical model was used for the analysis. ARijkl = Si + P( S )ij + Gk + I (G ) kl + ( S * G )ik + eijkl i = 1,2 j = 1,,18 k = 1,2 l = 1,,116, where S is the effect due to gender of the participant, P(S) is the random effect due to participant, G is the effect of image gender, I(G) is the random effect due to image, S*G is the interaction effect due to the gender of the participant and gender of the image, and e is residual error. Central conclusions of our analysis are summarized below: There is no significant interaction effect between the gender of the face and the gender of the rater. This means that the way male and female participants rated the faces did not differ depending on the gender of the image (p = 0.5024). There was a slight difference in how men and women rated faces overall (p = 0.0571), with males rating faces higher than females. Female faces are rated significantly higher than male faces (p = 0.0004) by both male and female raters.

18

Table 5 summarizes the attractiveness scores given by the raters. The ratings are separated by gender of the faces and the participants. The overall averages (a) are the average ratings given or received overall by males and females. For example, the female images were rated at an average of 4.8997 when the gender of the rater is not considered and the average rating of all images by all participants was 4.7308.

Female
Image

Male Overall

Participant Female Male 4.3597 (n) 5.1044 (n) 7.4656 (f) 7.4179 (f) 4.5887 (a) 5.2106 (a) 4.0845 (n) 4.7025 (n) 7.0915 (f) 7.1750 (f) 4.2952 (a) 4.8283 (a) 4.2221 (n) 4.9034 (n) 7.2786 (f) 7.2965 (f) 4.4419 (a) 5.0195 (a)

Overall 4.7321 (n) 7.4418 (f) 4.8997 (a) 4.3934 (n) 7.1333 (f) 4.5618 (a) 4.5628 (n) 7.2876 (f) 4.7308 (a)

Table 5: Summary of attractiveness ratings (n): Non-famous (FERET) faces, (f): Famous faces, (a): All the faces

When the faces were separated into famous (f) and non-famous (n), the results were fairly consistent. For the non-famous faces, the ratings given by males and females did not differ depending on the gender of the face (p = 0.3221). This is also true for the famous faces (p = 0.4951).

For the non-famous faces, the difference in ratings given by males and females was significant (p = 0.0473) with males giving higher ratings overall. For the famous faces however, there was no difference in the ratings given by male and female participants (p = 0.9543).

19

For the non-famous faces, there was a difference between ratings received by male and female faces (p < 0.0001), with female faces rated higher overall. For the famous faces, females were still rated higher than males but this difference was not significant (p = 0.1242).

These results suggest that males and females view attractiveness the same when looking at images of known attractive faces, but do not agree on attractiveness when looking at images of non-famous faces. Furthermore, female faces are rated higher than male faces by both the same and opposite gender of raters. Famous females are not rated significantly higher than their male counterparts while non-famous female faces are.

4.2 Do the male and female raters exhibit the same variability when rating faces?

One of the objectives was to determine if females and males exhibit the same amount of variability in rating faces. We also wanted to examine if the amount of variability was different depending on the gender of the faces and if the differences were consistent across genders of the faces and raters. The dataset for this analysis consisted only of those faces that were rated twice by the same rater resulting in 120 ratings by each participant. We computed the variance for each rater and each face gender as the variance in ratings of the same face compounded over all 30 sets of duplicate faces given to each rater. Thus, we have two variances per subject (72 variances in all), one for male faces and one for female faces. We found the response variable (variance) to follow a lognormal

20

distribution, so the GLIMMIX procedure in SAS was used (Schabenberger 2005). The analysis is summarized below.

There was no significant interaction between the gender of the rater and the gender of the face (p = 0.7168), i.e. the variability with which females and males rated images did not differ depending on the gender of the image. Although the difference is not significant (p
2 = 0.8318) tended to have somewhat higher variability in their = 0.1658), females ( F 2 = 0.5854). ratings than males ( M

In addition to the variances, the F statistics were computed to compare variability within images (variability in rating the same face) to variability between faces (variability in rating all faces). An F statistic of one would indicate the rater exhibits the same amount of variability both within and between faces. An F statistic larger than one indicates higher variability between faces than within faces meaning that the participant was quite consistent when rating the same face compared with his or her consistency of rating all faces. An F statistic smaller than one would indicate higher variability within faces than between faces, i.e., the rater was not consistent when rating the same image. Figure 5 shows the distribution of F statistics for the 32 raters in our study. It can be clearly seen that with the exception of two, the raters exhibited less variability in rating the same face than the variability with which they rated all faces.

21

F Statistics
14 12 Frequency 10 8 6 4 2 0 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 F Value

Figure 5: Distribution of F statistics for rating variability

We also examined if the F statistics were different for males and females and if there was an interaction between the gender of the face and the gender of the participant. The F statistics were calculated for each rater and face gender combination resulting in 72 total observations. These F statistics were also found to follow a lognormal distribution.

Results found no significant interaction effect (p = 0.8815), meaning that the difference in ratios of between face variability to within face variability for male and female raters is the same for both male and female faces. In addition, no significant effect was found due to the gender of the face (p = 0.7505) or the gender of the rater (p = 0.2219). Overall, we found no difference in the consistency with which males and females rated faces. Furthermore, males and females exhibit the same amount of variability in their ratings regardless of the gender of the faces.

22

4.3 Does the self-perception of attractiveness affect ratings?

Regression analysis was used to examine the relationship between a participants self rating and his or her average rating of others. If a person chose not to do a self rating, he or she was not included. The resulting dataset contained data from 22 (12 male and 10 female) of the 36 participants. For this analysis, the average rating of all the images rated by a participant is the response variable and the participants self rating is the explanatory variable.

A positive relationship was found between the self ratings of participants and their average rating of others (intercept = 2.898, b = 0.38, p = 0.0041, R2 = 0.3437). This indicates that as an individuals perception of his or her own attractiveness increases, so does his or her average rating of others. Separate analysis for males and females yielded similar results. Both had positive linear relationships, although the relationship was not significant for females (intercept = 3.1484, b = 0.30, p = 0.156, R2 = 0.234). Thus, for each unit increase in self rating by a female, the average rating of others increases by 0.30. The linear relationship between self rating and rating of others for male participants was stronger (intercept = 3.2116, b = 0.359, p = 0.049, R2 = 0.334). In males, each unit increase in self rating results in an increase of the average rating by 0.359.

23

Self Rating vs. Rating of Faces 8 Avg Rating of Faces 7 6 5 4 3 3 4 5 6 Self Rating 7 8 9 Female Male Linear (Female) Linear (Male)

Figure 6: Self rating vs. average rating given to others

Figure 6 shows average ratings given by a subject as a function of his or her selfattractiveness scores, with the males and females shown distinctively. The linear trend indicates that as ones perception of his or her own attractiveness increases so does his or her opinion of the attractiveness of others. The plot also seems to show the males self ratings are more skewed toward the right (higher scores) while the females self ratings are spread out. This might indicate that males view themselves as more attractive than females view themselves. Males rated themselves at an average of 6.833 while females rated themselves at an average of 6.0, although the difference is not significant (t20 = 1.46, p = 0.1605).

4.4 Is attractiveness related to speed of rating?

There is a significant relationship between the time it took to rate a face and the rating given to it. However, this relationship is dependent on the gender of the rater (p = 0.0016). For each additional second a female spent rating an image, the rating decreased 24

by 0.0135 points, although it is not significant (p = 0.3194). For males, as time spent rating increased the rating significantly increased (p = 0.0072). For each additional second males spent looking at an image, the rating they gave increased by an average of 0.0408 points. These trends did not depend on the gender of the face.

4.5 Relationship between Neoclassical Canons and Face Attractiveness

Of the six Neoclassical Canons described in Section 3.1, five had a significant relationship with attractiveness. Only Formula 7 (mouth width = 1.5 nose width) showed no relationship (p = 0.1412). If the canons are a true predictor of attractiveness, one would expect the scores to decrease as the coefficient of variation increases. This was true for all but one of the five significant canons. For Formula 5 (interocular distance = nose width), the attractiveness scores decreased as the coefficient of variation increased for male images, but the scores actually increased for female images (p = 0.0028). This suggests that female faces are viewed as more attractive when they have smaller noses and/or a larger distance between their eyes than proposed by the canon. For Formulas 2, 4, 6, and 8 the attractiveness scores decreased significantly as the proportions of the face deviated from the proportions defined by the canons (p = 0.0009, p = 0.0014, p < 0.0001, and p = 0.0064, respectively).

4.6 Relationship between Symmetry and Face Attractiveness

In Section 3.2, four measures to compute the symmetry in a face were presented. The first task was to determine which of the four measures had the strongest relationship with attractiveness. We also wanted to determine if adding angle symmetry significantly

25

increased the ability to predict attractiveness score. Finally we identify the pair(s) of points that play significant roles in the attractiveness of a face.

Face Symmetry Measures. Table 6 summarizes our analysis of the four face symmetry measures. It shows that the difference symmetry measure, which measures the difference in distances from the symmetric points to the line of symmetry, has the strongest relationship with attractiveness. The difference measure has the highest R2 value overall. Thus, the measure is able to explain more of the variation in attractiveness score than any of the other measures. When we examined how the four symmetry measures performed with the data separated based on the gender of the rater and the gender of the face, the difference measure had the highest R2 value in each instance. R2 Adjusted Rater/Image Difference Difference Ratio Ln(Ratio) All/All 0.0513 0.0572 0.0410 0.0493 Female/Female 0.0655 0.0917 0.0634 0.0644 Female/Male 0.0810 0.0878 0.0629 0.0762 Male/Female 0.0566 0.0798 0.0544 0.0558 Male/Male 0.0868 0.0897 0.0580 0.0820
Table 6: Summary of the performance of symmetry measures

Role of Angle Symmetry. When the angle symmetry measures are added to the difference symmetry measures and its relationship to attractiveness was evaluated, there was a slight increase in the R2 values. However, the increase was very small and hence our conclusion is that adding angles to symmetry calculations has no significant benefit in the evaluation of attractiveness. Therefore, it was not included for rest of the analysis.

26

Significant Feature Points. To determine the contribution of the symmetry pairs towards attractiveness of a face, we used a stepwise regression analysis to reduce the number of variables in the model. One would expect the relationships to be negative, that is, as the difference in distances of the two points from the line of symmetry increases, the attractiveness score decreases. In addition to the stepwise procedure, we are able to further reduce the number of variables by eliminating those that have a positive relationship with attractiveness, leaving us with five pairs of symmetry points. Both male and female raters find the symmetry of the nose (points 18 and 20) and mouth (points 25 and 27) as an important part of attractiveness when viewing male and female images (p = 0.0025, p = 0.0604). The symmetry of the upper tips of the lips (points 22 and 24) is also important for both genders of raters and images. For female images the attractiveness score increases by about 0.1 for every unit increase in the difference (p < 0.0001), but this pair is left in the model because it has a negative relationship with attractiveness for male images.

4.7 Relationship between Golden Ratios and Face Attractiveness

In Section 3.3, 17 ratios from popular literature (Meisner 2006, Narain 2003) were identified that would be included in this study to determine their relationship with attractiveness. If measurements of a face being close to the Golden Ratio is a predictor of attractiveness, the scores should decrease as the ratios in a face deviate from the ideal value. Of the seventeen ratios described in Section 3.3 six showed this relationship. Five of the six ratios that follow this trend are summarized in Table 7. The sixth is described following the table.

27

Ratio No. 2 5 6 7 17

bFemale -1.55 -1.56 -2.10 -3.66 -4.50

bMale -0.64 -1.56 -2.10 -3.23 -3.80

p 0.0040 0.0020 < 0.0001 0.0151 0.0030

Table 7: Regression coefficients for Golden Ratios

The rating given to a face is inversely proportional to the distance of ratio 2 in a face to the Golden Ratio. However, the ratings given by females decrease by a significantly larger amount than those given by males (p = 0.0040). The same is true for the ratio of lip to chin distance to nose width (ratio 7, p = 0.0151) and mouth width to nose width (ratio 17, p = 0.0030).

Both male and female raters rate faces as more attractive as ratios 5 and 6 (mouth width to interocular distance, p = 0.0020; lip to chin distance to interocular distance, p < 0.0001, respectively) approach the Golden Ratio. This trend is the same for both genders of faces.

As the ratio of the length of the face to the width of the face (ratio 14) gets closer to the Golden Ratio, both male and female faces are viewed as more attractive (p = 0.0077). However, female faces that deviate from the Golden Ratio have significantly lower ratings by female raters than by male raters. Male images that deviate from the Golden Ratio have the same decrease in attractiveness score when rated by males or females.

A number of the ratios described in Table 4 had a significant relationship with attractiveness even though the measurements from our face images are not close to the Golden Ratio. In addition, the attractiveness scores increase as the measurements of

28

these ratios get farther away from the Golden Ratio. For almost all of these ratios, the increase is significantly higher when the faces are viewed by female raters than when they are viewed by male raters. These ratios provide additional insights into understanding of the role of face geometry to attractiveness. Table 8, summarizes the results for these ratios. Ratio No. 3 4 8 10 16 bFemale 5.00 4.63 4.04 4.57 0.98 bMale 4.37 3.52 3.49 3.74 0.98 p 0.0433 <0.0001 0.0044 <0.0001 0.0130

Table 8: Ratios strongly related to attractiveness

The attractiveness score is highest when ratio 15 (nose to chin distance to lips to chin distance) is slightly larger than the Golden Ratio indicating that a smaller chin is more attractive (p = 0.0067). Unlike the ratios described in table 8, the increase in attractiveness score for this ratio is dependent upon the gender of the face. Small chins are significantly more important in female faces (b = 4.87) than in male faces (b = 1.63).

Images are viewed as more attractive when the nose width is approximately equal to the nose to lips distance (ratio 16) than when it is close to the Golden Ratio which suggests that smaller noses are preferred (p = 0.0130).

Images are viewed as more attractive when the distance between the middle of the eyes is much larger than the interocular distance or the nose width (ratios 3 and 4, respectively). This is more evidence that smaller noses are preferred to larger ones (p < 0.0001, p < 0.0001).

29

The attractiveness scores are highest when ratios 8 and 10 (interocular distance to eye width and nose width to eye width) are around one. These features are attractive when they are equal in size as suggested by the Neoclassical Canons (p < 0.0001, p < 0.0001).

4.8 Combining Multiple Measures to Predict Attractiveness

Using the Neoclassical Canons, difference symmetry measures, and golden ratios found to have a significant relationship with attractiveness we are able to develop a model for predicting attractiveness. Using the variable selection procedures described in Sections 4.5-4.7, we determined each variables relationship with attractiveness. Only the measurements that were found to have a significant, negative relationship with attractiveness during these preliminary stages were considered for this part of the model building process. The initial combined model, which will be referred to as the optimized model, contained sixteen predictor variables out of the original 78 (6 canons, 55 symmetry, 17 golden ratios) variables. The R2 value for this model was 0.1923, compared to an R2 of 0.2433 when all 78 variables were used.

Because we had previously observed differences in the way males and females view the attractiveness of certain features in the same and opposite gender images, separate models were created for each of the gender combinations. By separating into four different models, each was able to predict attractiveness better than the overall optimized model using all sixteen variables. A stepwise procedure, with entry and exit significant

30

levels set to 0.05, was implemented to obtain the most parsimonious models. The following table summarizes the results.

R2 R2 (Optimized) (Reduced) Female/Female 0.2378 0.2335 Rater/Image Female/Male Male/Female Male/Male 0.2162 0.2106 0.2053 0.2097 0.2088 0.2013

No. variables in Reduced Model 8 8 11 10

Table 9: Summary of each model after stepwise variable selection

Each of these four models performed slightly better than the model which included all raters and all faces. In addition, we were able to eliminate up to half of the sixteen variables without incurring much reduction in the R2 values. While each of the four models is slightly different, there are some commonalities between them, as shown in Table 10.
Variables in Final Models Rater/Image Canon Formulas Symmetry Pairs Female/Female 6, 8 22-24 Female/Male 2, 6 7-8, 18-20, 22-24 Male/Female 2, 4, 5, 6, 8 22-24 Male/Male 2, 4, 6, 8 18-20, 22-24, 25-27

Ratio Nos. 5, 6, 7, 14, 17 5, 6, 7 2, 5, 7, 14, 17 5, 6, 7

Table 10: Canon formulas, symmetry pairs, and golden ratios in the final models

All raters view the equality of the width of the eye and the interocular distance (canon formula 6) as attractive for both genders of images. In addition, male and female raters viewed the attractiveness for both genders of faces as higher when the ratio of lip to chin distance and width of the nose (ratio 7) was closer to the Golden Ratio.

31

Female raters preferred the ratio of lip to chin distance with interocular distance (ratio 6) to be less than the Golden Ratio no matter the gender of the image. This would suggest that female raters view a smaller chin and/or larger distance between the eyes as more attractive.

Male raters viewed the equality of the ear length and nose length (canon formula 4) as attractive regardless of the image gender. They also gave higher ratings when the nose width was not quite equal to one fourth of the face width (canon formula 8). From this information it seems male raters prefer a more slender face and/or a smaller nose.

Female images were rated higher when the mouth width to interocular distance (ratio 5) and ratio of the length to width of the face (ratio 14) were slightly less than the Golden Ratio. They were rated higher when the ratio of the mouth to the nose (ratio 17) was proportional to the Golden Ratio. In addition, ratings of female faces were higher when the upper tips of the lips (points 22 and 24) were slightly asymmetric which could support the claim that fuller lips are more attractive in females (Rhodes 2006). Overall, larger distances between the eyes and/or smaller mouth width along with face length to width in proportion less than the Golden Ratio are seen as attractive in female images.

For male images, symmetry of the upper tips of the lips (points 22 and 24) and symmetry of the nose (points 18 and 20) is viewed as attractive. The face being divided into equal vertical thirds (canon formula 2) is an attractive trait in men. The attractiveness scores are higher when the ratio of the mouth to the interocular distance (ratio 5) is proportional

32

to the Golden Ratio and the ratio of lip to chin distance with interocular distance (ratio 6) is less then the Golden Ratio. The latter of the two ratios was also viewed as important to attractiveness by female raters.

Even though the R2 values did not seem very high, we were able to explain between onefifth and one-quarter of the variation in attractiveness ratings using various Neoclassical Canons, symmetry measures, and golden ratios. This is actually quite good given the large amount of variation in the attractiveness scores. The models used produce predicted values that are generally close to the actual attractiveness scores which is evidenced by the small residual values for any rater gender and image gender combination. The studentized residuals were all between -1.48 and 1.45, well inside the usually acceptable range of 2, verifying that our models for predicting attractiveness work rather well. Table 11 shows the actual and predicted attractiveness scores for the images in Figures 1 and 2. Image A B C D Greta Garbo Meg Ryan Rock Hudson Keanu Reeves Actual Score Predicted Score 6.0833 5.2083 6.0000 5.5833 7.0278 8.4167 7.8333 7.3333 5.9379 5.2798 5.8239 5.6342 7.0005 8.0032 6.1887 6.8778

Table 11: Observed and predicted attractiveness scores

For most of the images, the prediction equation works well. As shown, the attractiveness score for Rock Hudson was quite underestimated. The image shown in Figure 2 shows

33

the face tilted and somewhat rotated which is a possible explanation as to why the model underestimated his attractiveness.

Summary and Future Work

The goal of this study was to determine a predictive model for attractiveness based on Neoclassical Canons, symmetry, and golden ratios. In contrast with much of the previous work, our study used landmarks and geometry based means for computing symmetry and had people rate actual faces instead of composite or altered faces. We also include both images of the general population and images of known attractive faces. In addition we identify both the gender of the rater and the image as to compare the ratings given to the same and opposite genders, as suggested by Rhodes (2006). While men and women do generally agree on overall attractiveness, male raters tend to give higher scores than their female counterparts. In addition, we find that male and female raters use somewhat different criteria for determining the attractiveness of a face. Female faces were rated higher by both male and female raters which supports feminine traits being viewed as attractive (Cunningham 1986, Cunningham et al. 1995, Rhodes 2006), but goes against the idea that ratings reflect a sexual attractiveness toward faces of the opposite gender (Cunningham et al. 1990). Our study on attractiveness is centered around the geometry of the face using a set of landmarks. This facilitates understanding roles of individual symmetric feature pairs and proportions in the attractiveness of a face. Our study is consistent with Rhodes (2006) in concluding that smaller chins in females are more attractive. We also find that smaller noses, a larger distance between the eyes, and smaller widths of the mouth are desirable traits for females. Symmetry does not seem to

34

play as important a role in attractiveness as the proportions defined by the Neoclassical Canons and golden ratios. This is demonstrated by the small proportion of symmetry predictor variables, as compared to the proportions of canons and golden ratios that were selected by the stepwise procedures to be included in the final models. Only three of the eleven difference symmetry measures were in any of the four reduced models, while five of six canons and six of seventeen golden ratios were included in at least one of the four models.

While the results presented in this paper provide strong insights into the role that different aspects of face geometry play in attractiveness, this research can be extended in many different directions. Attractiveness is a complex aspect of a face and involves many other issues, for example, Rhodes (2006) and others have studied the effects of averageness on the attractiveness of faces. We are interested in exploring this issue using a landmarkbased approach rather than composite face images. A secondary motive for including the images of famous people was to see if the perception of attractiveness changes over time. Our famous images included two male and two female faces from each of the past eight decades which would allow us to determine if a relationship with attractiveness exists due to the age of the rater and the time period during which the person was famous.

References

Armstrong, J. 2004 The secret power of beauty: Why happiness is in the eye of the beholder. London: Allen Lane.

35

Cunningham, M. R. 1986 Measuring the physical in physical attractiveness: quasiexperiments on the sociobiology of female facial beauty. J. Personal. Soc. Psychol. 50, 925-935.

Cunningham, M. R., Barbee, A. P., & Pike, C. L. 1990 What do women want? Facialmetric assessment of multiple motives in the perception of male facial physical attractiveness. J. Personal. Soc. Psychol. 59, 61-72.

Cunningham, M. R., Roberts, A. R., Barbee, A. P, Druen, P. B., & Wu, C-H. 1995 Their ideas of beauty are, on the whole, the same as ours: consistency and variability in the crosscultural perception of female physical attractiveness. J. Personal. Soc. Psychol.
68, 261-279.

DeBruine, L. M. 2004 Facial resemblance increases the attractiveness of same-sex faces more than other sex faces. Proc. R. Soc. Lond. Ser. B. Biol. Sci. 271, 2085-2090 Farkas, L.G. 1994 Anthropometry of the Head and Face. 2nd ed. Raven Press. Farkas, L. G. & Munro, I. R. 1987 Anthropometric Facial Proportions in Medicine. Illinois: Charles C. Thomas.

Farkas, L. G., Forrest, C. R., & Litsas, L. 2000 Revision of neoclassical facial canons in young adult Afro-Americans. Aesthetic Plastic Surgery 24, 179-184.

Farkas, L. G., Hreczko, T. A., Kolar, J. C., & Munro, I. R. 1985 Vertical and horizontal proportions of the face in young adult North American Caucasians: Revision of neoclassical canons. Plastic and Reconstructive Surgery. 75, 328-337. Grammer, K. & Thornhill, R. 1994 Human (Homosapiens) facial attractiveness and sexual selection: the role of symmetry and averageness. J. Comp. Psychol. 108, 233242.

36

Kowner, R. 1996 Facial asymmetry and attractiveness judgement in developmental perspective. J. Exp. Psychol. Hum. Percept. Perform. 22, 662-675.

Langlois, J. H., Kalakanis, L., Rubenstein, A. J., Larson, A., Hallam, M., & Smoot, M. 2000 Maxims or myths of beauty? A meta-analytic and theoretical review. Psychol. Bull. 126, 390-423. Langlois, J. H., & Roggman, L. A. 1990 Attractive faces are only average. Psychol. Sci.
1, 115-121.

Langlois, J. H., Roggman, L. A., & Musselman, L. 1994 What is average and what is not average about attractive faces? Psychol. Sci. 5, 214-220.

Little, A. C., Burt, D. M., Penton-Voak, & Perrett, D.I. 2001 Self-perceived attractiveness influences human female preferences for sexual dimorphism and symmetry in male faces. Proc. R. Soc. Lond. Ser. B. Biol. Sci. 268, 39-44.

Little, A. C., & Hancock, P. J. B. 2002 The role of masculinity and distinctiveness in judgments of human male facial attractiveness. Br. J. Psychol. 93, 451-464.

Little, A. C., & Jones, B. C. 2006 Attraction independent of detection suggests special mechanisms for symmetry preferences in human face perception. Proc. R. Soc. Lond. Ser. B. Biol. Sci. 273, 3093-3099.

Meisner, G. 2006 The human face. Available on the World Wide Web at http://Goldennumber.net/face. Last Accessed on December 9, 2006.

Movie Actor Index. Available on the World Wide Web at http://www.movieactors.com Last Accessed on December 20, 2006

37

Narain, D.L. 2003 The perfect face. Available on the World Wide Web at http://cuip.uchicago.edu/~dlnarain/golden/activity8.htm. Last Accessed on December 9, 2006.

OToole, A. J., Price, T., Vetter, T., Bartlett, J. C., & Blanz, V. 1999 3D shape and 2D surface textures of human faces: the role of averages in attractiveness and age. Image Vis. Comput. 18, 9-19.

Penton-Voak, I. S., Jones, B. C., Little, A. C., Baker, S., Tiddeman, B., Burt, D. M., & Perrett, D. I. 2001 Symmetry, sexual dimorphism in facial proportions and male facial attractiveness. Proc. R. Soc. Lond. Ser. B. Biol. Sci. 268, 1617-1623.

Perrett, D. I., Burt, D. M., Penton-Voak, I. S., Lee, K. J., Rowland, D. A., & Edwards, R. 1999 Symmetry and human facial attractiveness. Evol. Hum. Behav. 20, 295-307.

Perrett, D. I., Lee, K. J., Penton-Voak, I., Rowland, D., Yoshikawa, S., Burt, D. M., Henzill, S. P., Castles, D. L., & Akamatsu, S. 1998 Effects of sexual dimorphism on facial attractiveness. Nature. 394, 884-887.

Phillips, P. J., Moon, H., Rizvi, S. A., & Rauss, P. J. 2000 The FERET evaluation methodology for face-recognition algorithms. IEEE Trans. Patt. Anal. & Mach. Intell.
22, 1090-1104.

Rhodes, G. 2006 The evolutionary psychology of facial beauty. Annu. Rev. Psychol. 57, 199-226.

Rhodes, G., Proffitt, F., Grady, J. M., & Sumich, A. 1998 Facial symmetry and the perception of beauty. Psychol. Bull. Rev. 5, 659-669.

Rhodes, G., Sumich, A., & Byatt, G. 1999 Are average facial configurations attractive only because of their symmetry? Psychol. Sci. 10, 52-58.

38

Rhodes, G., & Tremewan, T. 1996 Averageness, exaggeration, and facial attractiveness. Psychol. Sci. 7, 105-110.

Samuels, C. A., Butterworth, G., Roberts, T., Graupner, L, & Hole, G. 1994 Facial aestheticsbabies prefer attractiveness to symmetry. Perception 23, 823-831.

SAS Institute. 2003 Online Doc. Version 9. SAS Institute, Inc., Cary, NC.

Schabenberger, O. 2005 Introducing the GLIMMIX procedure for generalized linear mixed models. Paper 196-30, SUGI 30. SAS Institute, Inc., Cary, NC.

Scheib, J. E., Gangestad, S. W., & Thornhill, R 1999 Facial attractiveness, symmetry and cues of good genes. Proc. R. Soc. Lond. Ser. B. Biol. Sci. 266, 1913-1917.

Shi, J. Samal, A., & Marx, D. 2006 How effective are landmarks and their geometry for face recognition? Computer Vision and Image Understanding 102, 117-133. Smith, R. J. 1999 Statistics of sexual size dimorphism. Journal of Human Evolution 36, 423-459.

Swaddle, J. P., & Cuthill, I. C. 1995 Asymmetry and human facial attractiveness symmetry may not always be beautiful. Proc. R. Soc. Lond. Ser. B. Biol. Sci. 261, 111116.

39