Performance of a 2D image-based anthropometric measurement and clothing sizing system

PIERRE MEUNIER† and SHI YIN‡

† Defence and Civil Institute of Environmental Medicine, 1133 Sheppard Ave West, P.O. Box 2000, Toronto, Ontario, M3M 3B9, Canada. Phone: (416) 635-2093. Fax (416) 635-2104. E-mail: pmeunier@dciem.dnd.ca

‡ VisImage Systems Inc., 1183 Finch Ave,West. Suite 500,  Toronto, Ontario, M2J 3C6, Canada.  Phone: (416) 398-5634. Fax (416)398-2690. E-mail: yin@vis.ca

Abstract

Two-dimensional, image-based anthropometric measurement systems have started to replace the traditional anthropometric tools in applications such as clothing sizing. These automated systems are attractive because of their low cost and the speed with which they can measure size and determine the best-fitting garment. Although these systems appear to be successful in this type of application, not much is known about the precision and accuracy of the measurements they take. In this paper, the accuracy and precision of one such system was analysed. The accuracy was estimated using a database of 349 subjects (male and female) who were also measured traditionally, and the precision was estimated through repeated measurements of both a plastic mannequin and a human subject. The results of the system were compared with those of trained anthropometrists, and put in perspective relative to clothing sizing requirements and short-term body changes. It was concluded that, when properly designed and calibrated, image-based systems can provide unbiased anthropometric measurements that are quite comparable to traditional measurement methods (performed by skilled measurers), both in terms of accuracy and repeatability.

Keywords: anthropometry, 2D-measurement system, automated clothing sizing

  1. Introduction

    2. In spite of highly standardised protocols designed to maximise the degree of repeatability and accuracy of measurements, anthropometric data are not always as reliable as they appear. Many factors come into play during the measurement of human subjects, which can result in the appearance of numerous sources of error. Some of the important sources include posture, identification of landmarks, instrument position and orientation, and pressure exerted by the measuring instrument (). The difficulty in controlling all potential sources of error is such that it has been said that true values are seldom measured in anthropometry (). Accuracy and precision of anthropometric measurements are at the mercy of the measurers who take them. Even if measured by highly trained observers, comparison of two populations may be meaningless (). In a comparative study by ), fifty boys (12 and 13 years of age) were measured independently by experienced observers in two institutes. Both teams of observers were trained to the same measurement techniques and used the same measuring instruments. In spite of this, systematic differences were found in nine of the twelve measurements taken. Pearson correlations between 0.872 (biacromial diameter) and 0.996 (stature) were found between the measurements taken by the two groups. Although the variable with the lowest correlation (biacromial diameter) did not present systematic errors, it suffered from repeatability problems (precision error). The results of these and many more studies show how difficult it is to measure humans, even under controlled conditions and after extensive training of the observers.

    Computerised image-based systems have found a niche in measuring individuals to determine their size of clothing and equipment. They offer rapid body measurements that are quickly translated into clothing size categories based on some fit criteria. Such systems can overcome some of the problems of traditional anthropometry, but can not overcome all sources of error. In image-based systems, the sources of error take the form of perspective distortion, camera resolution, landmarking error, and modelling error (since circumferences are not measured directly). The objective of this paper is to quantify the accuracy and precision of measurements made from two-dimensional images of humans, compare them with those of highly trained anthropometrists, and put the results in perspective in the context of clothing and equipment sizing.

    System description

    The system under review is a PC-based system comprised of two Kodak DC120 colour digital cameras (1280 x 960 pixels) and a blue backdrop embedded with calibration markers (Figure 1). The system takes simultaneous front and side pictures of individuals standing with their arms alongside slightly abducted. By taking both images simultaneously, the exact posture in space is captured, and it is possible to recover the object dimensions in 3D.

    Potential sources of error can be found at each of the steps in the image analysis process illustrated in Figure 2. These sources are: a) pre-processing of the front and side images, b) calibration of the cameras, c) segmentation of the body from the background, d) detection of landmarks, and e) calculation of the anthropometric dimensions.

    Theoretical assessment of error

    The error of a measurement is defined as the difference between the measured value and the true value of the item being measured. Errors can be catalogued as either random (precision error) or systematic (bias error). Precision is defined as the difference in values obtained when measuring the same object repeatedly. It has an average value of zero. Accuracy is the difference between the measured and true values. Bias error, which occurs in the same way on each measurement, affects the accuracy of a measurement while random error affects precision.

    The concept of error is useful, but it implies knowledge of the true value of what is being measured. Since any measurement contains error, the pure error can not be calculated. However, it can be estimated. Precision error can be estimated by taking a large number of readings on an individual and using a statistical model to determine the expected spread of values at a given probability level. Bias error, on the other hand, requires comparison of measurements with a more accurate method/instrument. This is difficult to do in anthropometry, given that the best available method is one that contains non-negligible error itself.

    A rough estimate of measurement error can be made from a theoretical perspective, using camera resolution as the starting point. Since the cameras used in this system have 1280 by 960 pixels covering an area that is approximately 2.5 m by 1.8 m at the subject, the corresponding spatial resolution is about 2.0 mm/pixel. Assuming a segmentation error of plus or minus one pixel, direct measurements requiring two points (i.e. for breadths, depths, and heights) will likely fluctuate within ±Β˜Ÿ 2 mm of the true value (1 pixel x 2 mm/pixel). The maximum error, which is obtained when both points err in making the dimension too small or too large, would put the result within ±Β˜Ÿ 4 mm of the true value (2 pixels x 2 mm/pixel).

    Circumferences can not be measured directly using only front and side pictures and must therefore be calculated using some form of mathematical model. The choice of model depends on the cross-sectional shape being measured, which varies among individuals. In two-dimensional systems, circumference measurement error depends on the accuracy of the model as well as of the breadth and depth measurements used in the calculation. Assuming a cylindrical object (i.e. no modelling error) and the same logic as above, the results would be likely to fluctuate within ±Β˜Ÿ 6 mm (p x (d1 – d2) = p x 2 mm) of the true value.

  2. Methodology

    3. Accuracy assessment

    The accuracy of the image-based system was assessed by comparing image-based measurements with manual measurements taken by anthropometrists during the 1997 survey of the Canadian Land Forces (). Six dimensions were selected because of their relevance to clothing sizing, which is the main purpose of the system. These were: stature, neck circumference, chest circumference, waist circumference, hip circumference, and sleeve length (spine-wrist).

    The test sample consisted of a subset of 349 subjects (95 females and 254 males) from the survey that had been measured both with traditional methods and with the image-based system. The image capture was performed within 90 minutes of the traditional measurements to avoid the effects of daily body variations. T-tests were performed to compare the means of all dimensions. Waist circumference was excluded from this comparison due to the difference in measurement definition between the two methods.

    Precision assessment

    The precision of the image-based system was determined by performing repeated measurements on a full size plastic mannequin as well as on a human subject. All image capture and analysis sequences were performed in succession (every minute or so) such that camera calibration and lighting conditions were relatively constant. The mannequin was used in order to exclude variations due to breathing movement and postural differences from picture to picture. The human subject was instructed to stand with the arms slightly abducted along the side the body during picture taking, and to move away from the platform between measurements. Thus, the precision estimates obtained this way contain variability coming from postural differences, breathing movement, and repositioning from one set of images to the other.

  3. Results

    4. Accuracy

    The means and standard deviations for the subjects measured manually and digitally are listed in Table 1. No significant difference was found between the means for either males or females. Table 1 also lists the Pearson correlation coefficients between manual and 2D image measurements.

    Females

    Males

    Measurement

    Mean

    Std.Dev.

    Correlation

    Mean

    Std.Dev.

    Correlation
    Stature:
    Manual

    163.2

    6.1

    0.98

    174.8

    6.4

    0.99
    2D system

    163.2

    6.2

    174.8

    6.5
    Neck circumference:
    Manual

    32.9

    1.8

    0.88

    39.5

    2.3

    0.94
    2D system

    32.9

    1.6

    39.5

    2.2
    Chest circumference:
    Manual

    95.6

    8.7

    0.95

    102.4

    8.3

    0.94
    2D system

    95.7

    8.4

    102.4

    7.8
    Hip circumference:
    Manual

    102.7

    9.1

    0.98

    100.5

    7.2

    0.94
    2D system

    102.6

    8.9

    100.4

    6.8
    Sleeve length:
    Manual

    79.9

    3.4

    0.79

    87.6

    3.5

    0.76
    2D system

    80.0

    2.7

    87.5

    2.6

    Table 1. Accuracy results

    Precision

    Table 2 summarises the results of thirty-five measurements of a plastic mannequin.

    Variable

    Mean

    Range

    Std.Dev.

    1.96 Std.Dev.
    Stature

    182.20

    0.27

    0.07

    0.13
    Neck circumference

    35.96

    0.51

    0.13

    0.26
    Hip circumference

    94.65

    1.24

    0.32

    0.63
    Waist circumference

    85.59

    0.90

    0.27

    0.54
    Chest circumference

    95.98

    1.28

    0.31

    0.61
    Sleeve length

    83.11

    4.29

    1.10

    2.15

    Table 2. Mannequin repeatability results (cm) .

    The results of ten measurements of an individual are shown in Table 3.

    Variable

    Mean

    Range

    Std.Dev.

    1.96 Std. Dev.
    Stature

    181.70

    0.46

    0.16

    0.32
    Neck circumference

    36.87

    0.58

    0.19

    0.38
    Hip circumference

    97.83

    1.14

    0.39

    0.77
    Waist circumference

    87.33

    1.51

    0.49

    0.95
    Chest circumference

    96.42

    1.57

    0.57

    1.11
    Sleeve length

    88.70

    3.56

    1.02

    2.01

    Table 3. Human repeatability results (cm).

  4. Discussion

Accuracy

The overall results did not indicate the presence of large systematic errors in the image-based system when compared to the manual measurements. This is not surprising since the indirect measurement models were fine-tuned using those data. However, there was evidence of differences between measurement methods, especially with respect to the spread of results of neck circumference and sleeve length. In both cases, the spread of digital image results was somewhat smaller than for the manual measurements. The small difference in neck measurement spreads may have been due to differences in landmark identification criteria as well as differences in means of measurement. In the manual method, accuracy may be affected by improper positioning of the measuring tape and skin compression, whereas in image-based measurement, accuracy may be affected by unreliable landmarking and mathematical modelling.

The difference in standard deviations is even greater for the sleeve length measurement. Because of the automatic landmarking algorithms, postural variations, and wrist and shoulder landmark detection inconsistencies affect the accuracy of sleeve length. Review of the survey images confirmed the presence of inconsistent hand postures (some in pronation, some in supination), arms that were not vertical, and bent elbows. These can be remedied by providing subjects with better instructions, and different algorithms to deal with postural variations. In fact, recent trials have shown that a better control of the arm and hand position across subjects has improved the reliability of this measurement considerably.

Precision

The theoretical assessment of the random measurement error made earlier suggested that an error of the order of ±Β˜Ÿ 0.2 cm and ±Β˜Ÿ 0.6 cm could be expected on direct and indirect measurements respectively. As shown in Table 2, the results of repeatability tests performed on the plastic mannequin showed the actual errors to be slightly smaller, e.g. within 0.13 cm of the mean for stature, 95% of the time. Where the mannequin’s shape attributes were true to life (i.e. except for hinged joints, non-standard posture and unnatural shapes), reliable landmark positions were obtained. Hinges at the shoulder, elbow and wrist hindered the repeatability of sleeve length measurements. Fluctuations in this measurement in particular were unavoidable because the landmark detection software was developed to recognise real human shape. Other than for the neck, which was better, circumferences were found to be within 0.63 cm of the mean, 95% of the time (Table 2). Neck circumference exhibited significantly better repeatability due, in part, to special attention paid to it during the development and the fact that it is relatively easy to locate and measure.

Overall, it would appear that segmentation and landmark identification errors tend to fluctuate by one pixel on a given direct measurement. The ratio of three between direct and indirect measurement error derived in the theoretical assessment was consistent with the circumference measurements observed in the data, i.e. p x 1 pixel x 0.2 cm/pixel = 0.63 cm.

The results in Table 3 show that, for the most part, repeated measurements of a human subject showed the same basic trend as for the mannequin, i.e. direct measurements were more precise than circumferences, and neck circumference was more repeatable than other circumferences. In most cases, the human results exhibited more variability in measurement than in the case of the mannequin, which was anticipated. The largest difference between mannequin and human subject measurements were for waist and chest circumferences. This can be partly explained by torso movement during breathing (expansion and contraction of the rib cage and abdomen) and differences in posture from picture to picture (arm position, relaxed or tight posture).

Computer versus human repeatability

The results of the repeatability study on a human subject were compared with those of recent large-scale surveys where accuracy and precision were monitored throughout. The first survey was conducted on the Canadian Land Forces personnel in 1997 (). The second survey was conducted on US Army personnel in 1988 (). Repeated measurements were part of the routine during the both surveys, although the methodology was slightly different. In the Canadian Land Forces survey, subjects were re-measured by the same observers within minutes (10 to 90 minutes) of the first measurement (see for details). This can be viewed as the best case scenario in terms of repeatability, since it is assumed that the same observer will measure in the same way given the same landmarks. In the US Army survey, subjects were re-measured within minutes by a second observer. This case can be viewed as the best case scenario for repeatability by different observers, since both observers were highly trained on the dimensions specific to their measuring station.

The technical error of measurement (TEM), which is essentially a form of standard deviation, was used as the basis for comparison. The TEM, or r, was calculated using the following equation:

(1)

Figure 3 shows the TEMs for computer measurements (on a mannequin and on a human) and compares them to those obtained by trained human observers (single () and dual observer results (taken from )). Although the computer measurements contain an additional source of error due to automatic landmarking, the results indicate that the repeatability was similar to the single observer results for stature and neck circumference. The single observer results had the lowest TEMs for all other measurements, however, followed by computer measurements on a mannequin and on a human, followed by the measurements made by two observers.

The differences observed between mannequin and human repeatability results show the effect posture and breathing can have on measurements. Better precision could be obtained by controlling these factors, if required. It should be noted that had the survey results included landmarking error (the survey subjects had the same landmarks during re-measurement), the results could have been more in favour of the computer measurements.

Reliability

state that two pieces of information are sufficient to characterise the reliability of an anthropometric variable: the TEM and the reliability coefficient. The reliability coefficient (R) is an interesting metric in that it compares the variability due to measurement error (r2) against the biological variability of that dimension (sample variance s2). It is computed using the following equation:

(2)

where r is the technical error of measurement, s is the sample standard deviation, n is the number of subjects and k is the number of measurements per subject.

If the measurement error is small compared to the standard deviation of the sample then the reliability of that measurement will be high. Reliabilities above 90 to 95% have been recommended for the selection of variables in a survey (). The reliability coefficients obtained by image-based measurement system were well above that, for the dimensions shown in Table 4.

Reliability

Stature

99.9%

Neck circ.

99.3%

Hip circ.

99.7%

Chest circ.

99.6%

Waist circ.

99.7%

Table 4. Reliability of a 2D image-based measurement system.

Clothing perspective

The ultimate goal of 2D image-based system is to determine the best fitting size of garment for a given individual. Anthropometry is one side of the equation, but clothing size and design is on the other. An idea of how much accuracy and precision is required for clothing size prediction can be obtained by considering some of the factors affecting clothing fit. Some of those factors are:

 
Tolerance (cm)
Trouser waist

±Β˜Ÿ 1.3
inseam

±Β˜Ÿ 1.3
Shirt neck

±Β˜Ÿ 0.3
chest

±Β˜Ÿ 1.3
sleeve

±Β˜Ÿ 1.3

Table 5. Typical manufacturing tolerances (on circumference) for dress trousers and shirts.

 

Size increments (cm)
Trousers stature

7.6
waist

5.1
Shirt neck

1.3
sleeve length

5.1
Jacket stature

7.6
chest

5.1

Table 6. Typical clothing size increments for dress uniforms.

One can see by the way clothing is made that the greatest emphasis is placed on the neck. Other garment dimensions are not as tightly controlled.

Body variation

Anthropometric accuracy and precision must also be balanced against body changes over minutes (breathing), hours (stature), days (weight changes), weeks (waist circumference changes), etc. Several body dimensions can change substantially over short periods. Stature, for instance, has been known to change by 3 to 5 cm in a day depending on the amount of standing, walking and carrying done (). In view of this type of fluctuation, it does not seem reasonable to measure within 0.1 cm a variable that can change by an order of magnitude during the course of the day.

also reported changes in dimensions over time. In those experiments, repeated measurements of one subject were made at various times of day over a number of days by the same observer. The results (Table 7) show that measurements varied significantly, most notably for waist circumference, where 95% of the measurements were within ±Β˜Ÿ 2.1 cm of the mean, but also for chest circumference. Again, one could argue that measurement to within 0.1 cm is unnecessary for a dimension that can vary by an order of magnitude over a few days.

1.96* s.d.

(cm)
Waist circumference

2.1
Chest circumference

1.5
Neck circumference

0.5

Measurement accuracy requirements

The first part of the discussion dealt with the capabilities of the image-based measurement system when compared with skilled human measurement. But the answer to the question "how much measurement accuracy is required?" can only be answered in the context of the application. For clothing sizing, a large part of the answer comes from the manufacturing tolerances. In a sense, the manufacturing tolerances represent the limits of a trade-off between fit of the clientele and cost of the garment. They could be interpreted as an amount of fluctuation in garment dimensions having minimal impact on fit for most of the customers of that nominal size. By extension, it could be said that given a garment size, the same amount of fluctuation in body measurement would also have minimal impact on the fit of a garment.

From a measurement standpoint, it is also important to balance the accuracy against short-term body variations. These variations, which occur naturally, must be accommodated by the clothing regardless of their magnitude in order for the clothing to be acceptable. Thus, using this argument, it would stand to reason that the magnitude of short-term body variations should temper measurement accuracy. A comparison of tables 5 and 7 shows a certain agreement between manufacturing tolerances and the short-term body variations that clothing must accommodate. Hence, it can be concluded that, in a balanced approach, measurement system accuracy should also be consistent with both. Neck circumference accuracy should be highest at ±Β˜Ÿ 0.5 cm, whereas all other dimensions should be within ±Β˜Ÿ 1.5 cm.

  1. Conclusions

When properly designed and calibrated, image-based systems can provide unbiased anthropometric measurements that are quite comparable to traditional measurement methods (performed by skilled anthropometrists) both in terms of accuracy and repeatability. The quality of the results depends, in large part, on the dependability of the automatic landmarking algorithms and the correct modelling of indirect measurements. Once that is achieved, however, this type of system can provide uniform measurement of a population regardless of where, when or by whom, it is operated.

In light of short-term body changes, clothing design, fit, and manufacturing tolerances, it was determined that a body measurement system should be capable to measure neck circumference within ±Β˜Ÿ 0.5 cm in order to be effective, and all other dimensions within ±Β˜Ÿ 1.5 cm. From the accuracy and precision analyses, it was found that the system under evaluation was capable of this performance.

References