Abstract
This is a report on age recognition in faces by Mara Chin-Purcell. Sixty-four faces of ages 0 to 69 were compiled and made into a survey. The survey was given to 58 subjects. The data was compiled into six spreadsheets and made into six major graphs. Topics looked at were Overall Error, Error by Subject Age, Average Error vs. Subject Age, Average Error vs. Face Age, Estimation Error Over All Faces, and Male vs. Female Accuracy. The results showed that people were more accurate at guessing the younger ages, probably because there is a smaller age range to choose from in younger children. People were also more accurate at guessing the ages of older people than they were at guessing the ages of middle-aged adults.
Males were shown to be generally more accurate than females. Also, the Overall Error sheet showed that subjects were more accurate at guessing the ages of people younger than the subject than people older than the subject, disproving my hypothesis that people would be more accurate around their own age than at other age groups.
Table of Contents
Abstract..................................................................................................... 2
Statement of Purpose…………………………………………………….4
Literature Review....................................................................................... 5
Hypothesis................................................................................................. 7
Materials.................................................................................................... 7
Procedure.................................................................................................. 8
Results....................................................................................................... 9
Analysis Results..................................................................................... 9
Graphical Results................................................................................. 12
Conclusion............................................................................................... 20
Bibliography............................................................................................. 21
Appendix................................................................................................. 22
Statement of purpose
The reason I am doing this experiment is to determine if people are more accurate at guessing the age of faces closer to their own age and also to discover how accurately people guess ages in general. For example, are 50-year-old people better at guessing ages than 7 year olds? What about 30-year-olds? The test faces will span about the same ages as the test subjects. Probably the test faces' age range will be from 5 years to 65 years. The test faces will be acquired from a database of images used in an age study from the University of Cyprus.
The survey will generate a large data set; I will be able to get answers to ask several questions of this database. Like, are people more accurate for younger or older faces? There can be many different interpretations of my data.
Literature Review
My topic is about age perception in faces. My plan is to have subjects from a wide range of ages guess at the ages of people shown in a series of pictures (aged from 1 to 69). I hope to discover how the age of the subject influences their perception of age in other people. My hypothesis is that they will be able to guess age in people closer to their age more accurately.
Through my research I have not found another study exactly like mine, but there have been some that have been close. Several studies have been done on age and attractiveness. Kathleen M. Korthase and Irene Trenholme did such a study [1]. The purpose was to determine if older faces were judged as less attractive to younger faces. Sixty subjects in four groups of 15 were studied: adults, male and female, and adolescents, male and female. They were given a set of male faces and a set of female faces to look at and they were asked to arrange them in order of attractiveness for each set. They were also asked to guess the age of the pictures. The researchers found that as perceived age increased, perceived attractiveness decreased. Differences between the perceptions of adults and adolescents were not marked, but there was more consensuses between younger females about male attractiveness and female age. All subjects were more in agreement about female attractiveness than male attractiveness.
Another age and attractiveness study was done by the same researchers on children’s perception of age and attractiveness [2]. Forty third and fourth grade students were asked to rate photos on age and physical attractiveness. The results were consistent with the research on adults [1], showing that the basis for age and attractiveness is already set by 3rd grade and maybe even before then.
A research overview by Diane S Berry [3] investigated age-related facial changes and social perception. She found that the more prominent your cranium (forehead) is, the younger you are perceived as, and also the less seriously you are taken. She also states that as you age, your forehead size diminishes, and your chin size increases. In females, a younger look increases cuteness and warmness perceptions. In males, the narrower, smaller chin was associated with honesty and kindness. More angular aces are perceived as older, and also stronger than curved ones. For eyes, the larger eyes with round edges and pupils are characteristic of infants and give a face a younger look. Eyebrows of the young are lighter, higher on the face and fainter than that of older people. As you age, you age, your skin goes from the baby softness of an infant to rougher, less flexible skin of old age. Wrinkles appear because of this lost flexibility, the more wrinkles a face has, the older the face is generally perceived as. Feature length (the distance between the eyes and mouth and nose) is shorter in younger faces than in older faces. Overall facial babyishness has been shown to make women more attractive and appear more honest, caring and naïve. In men, facial babyishness results in the same effects, but not attractiveness. ‘Baby’ faced men and women are often perceived as more innocent or kind than non-‘baby’ faced people. As a result, ‘baby’ faced people often get less severe punishments in court and are not readily perceived as being guilty.
A very close study to mine was a study of “Perception of Age in Adult Caucasian Male Faces” by D. Michael Burt and David I. Perrett [4]. One hundred forty seven male faces between the ages of 20 and 62 were compiled. Seven groups of age were separated, and blended together in both shape and color to produce a prototype of that age span. They were shown to 40 subjects. The blended faces were consistently perceived as younger than individual faces, and the difference was increased with increased age. Blends of young and old faces did increase the perceived age from the younger subjects as expected.
From my literature review I learned several things. A good sample size of subjects and images would be about 40 to 50 people. I also learned what the current research is in facial recognition and that it is interesting how many studies have been done on things close to my topic. I have also learned some analysis techniques, data presentation and about important ways to carry out my study.
Hypothesis
My hypothesis is that subjects will be able to guess faces ages more accurately if the face is closer in age to the subject. I think this because people might spend more time with people that are of their age, and they might become more finely tuned to people of that age. For example, a person who is 13 might think a man who is, say, 50, looks older than they are because they are used to being around young teens.
Materials
1. Surveys, with at least 50 faces, with many copies of the same survey
2. Paper
3. Computer
4. Permission from teacher or official
5. Database of photos of many different ages and genders
6. Printer or Printing shop
7. Access to many subjects of different ages and genders
8. Microsoft Excel or another graphing program
9. Picture editing program
Procedure
1. Create the survey:
a. Pictures of faces were assembled from a database made available from the University of Cyprus [7].
b. All the pictures were image processed using the program Picasa [6] to balance the contrast. Some of the pictures were color, and some were black and white. The color pictures were made black and white to remove any bias for or against color pictures.
c. Sixty-four faces were chosen from a total of 1008 faces. All blurred, and flawed photos were passed over in favor of clearer ones. The goal was to choose one picture for each year of age from 0 to 69.
d. The face images were arranged in random order.
e. Sixteen pictures were placed 4 by 4 on a page and were all made 400 pixels high.
f. A cover page describing the experiment was created, asking for the age and gender of the subject. The human subject permission form was also made part of the cover page.
g. About 80 photocopies were made for the test.
2. Give survey:
a. The survey was given to about 50 subjects.
b. The range of subjects was from 7 to about 80.
3. Analysis
a. The data was gathered and entered into the computer
b. Graphs were created showing the accuracy of subject’s guesses
Results
Analysis
My data set has 3712 data points. There were 18 men and 40 women for a total of 58 subjects all together and each subject made 64 age guesses. In the different age groups there were 9 6-15 year-olds, 16 16-25 year-olds, 9 26-35 year-olds, 6 36-45 year-olds, 10 46-60 year-olds, and 6 people 60 and up. [See appendix A for histogram of subject ages]. There was mostly one face for every age in the survey, but there were some older ages that were not present.
Six spreadsheets were created to analyze the data. Spreadsheet 1 is Raw Data. [Appendix B] This was the place where data was entered after being collected. There were columns to record for each subject for survey number, initials, age, and gender. The subjects were sorted by their age, from 7 to 84. From column E to BP, there was the data collected. Across the top was the actual age of the picture and an index number for each picture. The pictures occurred in the order they were on the survey (randomized). On the bottom, below the guesses of subjects, there was a row that calculated the average guess of age for each face on the survey. Below that was the amount off the average guess was from the actual. This spreadsheet was the building block of the whole project.
Spreadsheet 2 is Diff or difference. [Appendix C] The goal of this spreadsheet is to show the number of years subjects were off in either a negative and positive direction, and also to calculate the standard deviation and the average net error. The subjects and pictures were arranged the same as on Raw Data, but instead of the age guessed, there was the number of years off the person was from the actual face age. For example, if the subject guessed 25 when the person was 18, the value would have been -7. At the right end of the data there were 2 columns with calculations on the data for each subject. Average Net Error took all of the data and added it, positive and negative to find how many years over or under the actual age of the picture a subject guessed on average. Standard Deviation took the square of the error numbers and averaged them to find the average deviation of years a person guessed off per face.
Spreadsheet 3 is Absolute. [Appendix D] The purpose of this spreadsheet was to take the Diff data and make it all positive, thus making a positive or negative error the same. From these errors it was then possible to calculate the total absolute error and the average absolute error for each subject. On this spreadsheet, was the same data as on the Diff sheet, but without the negatives. This showed the number of years off a person was regardless of whether the guess was high or low. Here there was also the calculation of Total Absolute Error and Average Absolute error. Total Absolute Error simply added all of the absolute errors per person. An average score was around 450. Then the Absolute Error was taken and divided by 64 to find the Average Absolute Error per face. An average score was about 6.5. The data was also taken and divided into male and female errors. The average error (as in Raw Data) was found for each face. Then the female and male guesses were averaged separately for each face. All this data was put into graph #6.
Spreadsheet 4 is Error By Groups [appendix E], containing the graphs showing how accurate certain groups of people were at guessing the ages of different groups of picture ages. There were 5 groups of subjects (as above), and each group had a row. In the columns, arranged by actual age, were the pictures in the survey, from 0 to 69. Taking data from the Absolute page, it calculated the average guess for people aged 6 to15 for a particular face, say for example the 2 year-old face. Then it took the average guess for the 16 to 25 year olds for each picture, and so on. Graph #3 and #4 are column graphs of this data.
Spreadsheet 5 is Data by Subject Age [appendix F], where the survey number, initials, age, and gender were taken from Raw Data. Average net error, standard deviation and average absolute error were taken from Diff and Absolute. A scatter graph was made from all this data and trend lines were put in for each of the data sets. The goal of this spreadsheet was to see if any trends existed as function of subject age.
Spreadsheet 6 was Age Diff [for age difference] [Appendix G]. This was where all the data was pulled together into a single spreadsheet, with the goal of looking for a trend in errors as a function of the difference between the face age and subject age. To do this each data point (a guess) needed an associated value for the age difference between the face’s age and the subject’s age. This required a single long list where each data point had a coordinate that referred back to the data table in Absolute. To do this, in column A was the index that simply counted the data points 0-3712. Column B had a formula to count 0 for 64 rows, then 1 for 64 rows, then 2 and so forth (up to 58). This was counting the subjects. Column C cycled by 64, it would count to 63, and then start back at 0 again. This was counting the pictures in the survey. Column D was the age difference between the face and subject. For example, subject 0 was 7 yrs, and picture 0 was 0 Yrs, so the age difference was –7. Column E was the error that each person had on each face (from Absolute). This list of the data was made into a graph [graph #1]. A trend line added shows that you are more accurate at guessing the ages of people younger than you than older than you.
Spreadsheet 7 [Appendix H] uses the same data and analysis as spreadsheet six, except the children, teenagers, and adults are separately split up into different columns. The revised graph is graph #2.
Graphical Results
There were 3712 data points in this study. Naturally, there were many ways to view the data. Choices had to be made as to the best ways to split and analyze the data. To get an overview the data is first looked at all together in graphs #1 and #2, then split into different views such as gender differences in guessing.
Graph #1: This chart encompasses all of my data points (3712). The middle line is the y-axis representing error per face. The x-axis is the difference off the subject’s age. For instance, if a subject was 23, and the face was 45, the x value would be 22. If a subject was 67, and the face was 55, the value would be –22. The trend line put in is a polynomial, and could curve, but it doesn’t. This shows that people guess ages more accurately if the people are younger than themselves. If my hypothesis were correct, the line would have looked like a U. This graph disproves that theory.
The data might be influenced though by the actual ages of the faces. The next graph splits out the faces by age groups; child, teenage and adult.
Graph #2: This graph takes graph #1 and breaks it down into adult, teen and child age ranges for the face pictures. The children’s trend line dips down (i.e. more accurate) when the subject age is 30 years older than the picture (-30). The same x and y-axes are present in this graph as were on graph #1. Teenage faces were most accurately guessed by people 10-20 years older than them. Adults stayed relatively steady, but the line dipped slightly around the subject’s own age and slightly older (10-15 yrs).
This relates to my hypothesis by showing that there are ‘U’s in the graph, but not in the places that I would expect. The children’s and teenager’s bottom of the ‘U’ is about 20-30 years older than them, prime child bearing age.
In the next graph we look at grouping ages and those grouped age’s guesses to groups of picture ages. We will look for more patterns that can be used to reach a conclusion
Graph #3: This is a column graph showing groups of people (age 6-15 for example) and how well they guessed the ages of the pictures of different age groups. For instance, people of ages 16-25 on average guessed the ages of pictures aged 26-36 8 years off their real age.
Notice that that the children and teenagers have the lowest error rates. Not many other patterns can be seen.
However, if we group the data groups the other way you see many more interesting patterns:
Graph #4: This graph uses the same data as Graph #3, but it shows the data differently. An interesting thing is that not only is it easier for people to guess younger ages but it is easier for people to guess old ages than for people to guess middle aged people. (Note the wave, down-up-down, pattern of the graph). Notice how the children’s error keeps climbing, however, while everyone else’s error seems to ‘hump’. Older people are the worst at guessing the ages of young people. This is probably because they are not around them as much. Also, note that the people most accurate at guessing the ages of the hardest age group to guess (36 to 45) are the people of that age.
Graph #5: This graph scatter plots the Standard Deviation, the Average Absolute Error, and the Average Net Error. Trend lines were added to all three data sets. Only the Average Net Error shows any really any difference between ages.
Although the younger subject’s net error is lower, it doesn’t mean much, since their absolute errors are just as scattered as older subject’s errors. The younger subjects just did better at having the high and low guesses cancel each other.
Graph #6: This graph illustrates the difference between male and female Average Error. Males are more accurate at guessing ages as a whole, but they are less accurate when it comes to younger children. Females are on average worse at guessing ages, but better for younger children.
Conclusion
My results were very interesting, but they did not necessarily prove my hypothesis. The graph that would have proved it would have been graph #1 or #2. If my hypothesis were correct, the polynomial trend lines would have looked like U’s with the bottom of the U at the y-axis. This would mean that people were more accurate guessing the ages of people around their own age and ages close to them. However, the graphs showed U’s in different places. Interestingly, children were not always more accurate guessing the ages of other children, but middle-aged adults were very accurate at guessing the ages of children. This makes sense, though, because those are the adults of childbearing age who would be around children. Teenagers also were not most accurate at their own age, but the bottom of the ‘U’ for teenage faces was in a different place. People 10 to 20 years older than teens were the most accurate at guessing teen faces. For adult faces, there was not much of a ‘U’, but it did dip slightly for the adults the same age and older. So adults were most accurate at guessing ages of adults their own age, or adults slightly older than them.
Of interest was the difference between men and women’s average guesses. Although men were better overall, women were more accurate for the children. This might have been because women traditionally are around children more than men.
In graph #4 there are many interesting trends. You can notice in the graph (and also in graph #3) that the children’s (6 to 15) guesses tend to keep increasing as the picture ages get higher. The other age groups generally have a ‘hump’ in the middle for middle age. I can only speculate on the reason for the better guesses in the older age groups. It could be because people look characteristically old, but it is harder to guess the age of people in the middle age because their faces don’t change appearance much. As in graph #2, graph #4 shows that people of childbearing age are better at guessing the age of children. Also interesting is the bar that shows that people of the middle age (hardest to guess age) are best at guessing people of their own age. This may be because they are around people of that age more, and therefore become more fine-tuned to the cues to tell the age of middle-aged adults.
When looking at graph #5, you can see that the average net error line seems to indicate that young peoples are better at guessing than older peoples. This is not necessarily the case. It may just be that younger people are better at getting their guesses to cancel each other out. They may not yet have any prejudices towards guessing older or younger than faces consistently yet.
Some of the logical ways to apply this data would be in forensics (police’s line of work). If an eyewitness who was 16 said, “ The man was about 50 yrs old” you would know that this guess is on average 9 years off. Or, if you know that people guess on average about 6 yrs older than the person actually is , you could go looking for a man in his older 50’s.
If I were to take this research a step further, I would try some more studies with age relationships and study certain age groups one at a time.
I hope you have enjoyed and learned something from reading this report.
Bibliography
1. Trenholme, Irene and Korthase, Kathleen. “Perceived age and Perceived Physical Attractiveness.” Perceptual and Motor Skills. 1982: 1251-1258.
2. Trenholme, Irene and Korthase, Kathleen. “Children’s Perceptions of Age and Physical Attractiveness.” Perceptual and Motor Skills. 1983: 895-900.
3. T. R., Alley. Social and Applied Aspects of Perceiving Faces. Hillsdale, NJ: Erlbaum.
4. Burt, Michael and Perrett, David. “Perception of Age in Adult Caucasian Male Faces.” The Royal Society. 1995.
5. Hager, Joseph C. Aging of the Face.
<http://face-and-emotion.com/dataface/facets/aging.jsp> [November 2005]
6. Google, inc. Picasa. <http://www.picasa.com> [November 2005]
7. Lanitis, Andreas. FG-NET Aging Database. <http://sting.cycollege.ac.cy/~alanitis/fgnetaging> [October 12, 2005].
Appendix
The original excel workbook (1.5 MB file)