Direct 3D Body Measurement Estimation from Sparse Landmarks

David Bojani

1 a

, Kristijan Bartol

2 b

, Tomislav Petkovi

1 c

and Tomislav Pribani

1 d

Faculty of Electrical Engineering and Computing, University of Zagreb, Unska 3, Zagreb, Croatia

TU Dresden, 01062 Dresden, Dresden, Germany

Keywords:

Anthropometry, Body Measurements, 3D Landmarks.

Abstract:

The current state-of-the-art 3D anthropometry extraction methods are either template-based or landmark-

based. Template-based methods ﬁt a statistical human body model to a 3D scan and extract complex features

from the template to learn the body measurements. The ﬁtting process is usually an optimization process,

sensitive to its hyperparameters. Landmark-based methods use body proportion heuristics to estimate the

landmark locations on the body in order to derive the measurements. Length measurements are derived as

distances between landmarks, whereas circumference measurements are derived as cross-sections of the body

and a plane deﬁned at the desired landmark location. This makes it very susceptible to noise in the 3D

scan data. To address these issues, we propose a simple learning method that infers the body measurements

directly using the landmarks deﬁned on the body. Our method avoids ﬁtting a body model, extracting com-

plex features, using heuristics, and handling noise in the data. We compare our method on the CAESAR

dataset and show that using a simple method coupled with sparse landmark data can compete with state-

of-the-art methods. To take a step towards open-source 3D anthropometry, we make our code available at

https:/github.com/DavidBoja/Landmarks2Anthropometry.

1 INTRODUCTION

Anthropometry is the scientiﬁc study of the measure-

ments and proportions of the human body (Bartol

et al., 2021). With the most recent development of 3D

scanners, now available even in mobile devices (Zhao

et al., 2023), the need for body measurement methods

from 3D data is increasing. Automatic extraction of

body measurements could accelerate the tedious and

time-consuming manual measurement process crucial

for numerous applications such as surveying (Zakaria

and Gupta, 2019), medicine (Donli

c, 2019; Heyms-

ﬁeld et al., 2018), fashion (Zakaria and Gupta, 2019),

ﬁtness (Casadei and Kiel, 2020), and entertainment

(Camba et al., 2016).

The ﬁrst methods to address automatic 3D body

measurement extraction were landmark-based. The

goal of landmark-based methods is to extract body

locations suitable for accurate measurement. These

methods usually use heuristics, such as body pro-

portions (Lu and Wang, 2008), to determine the ap-

https://orcid.org/0000-0002-2400-0625

https://orcid.org/0000-0003-2806-5140

https://orcid.org/0000-0002-3054-002X

https://orcid.org/0000-0002-5415-3630

proximate landmark locations. To further reﬁne the

landmark locations, the change in body curvature

(Markiewicz et al., 2017) or the change in the cross-

section by cutting the body with a plane (Zhong

et al., 2018) is analyzed. Length measurements can

then be estimated using the Euclidean distance or the

Geodesic distance (Xie et al., 2021) between the land-

marks. Circumference measurements can be deter-

mined by cutting the subject with a plane at a de-

sired landmark, and ﬁnding the cross-section (Lu and

Wang, 2008). These measurements are, however, less

accurate in the presence of data noise. Additionally,

such methods assume speciﬁc scanning poses (Zhong

et al., 2018), no severe distortions of the body (Xie

et al., 2021), known gender (Lu and Wang, 2008), etc.

Recent advances in human body measurement

have been driven by statistical human body models

(often referred to as templates) such as SMPL (Loper

et al., 2015). The templates are relevant because

body measurements can be predeﬁned since the se-

mantics of each vertex remain consistent across dif-

ferent poses and shapes (Wasenmuller et al., 2015).

Template-based methods ﬁrst ﬁt a statistical body

model to a 3D scan to ﬁnd the optimal pose and shape

parameters after which the predeﬁned measurements

524

Bojani

c, D., Bartol, K., Petkovi

c, T. and Pribani

c, T.

Direct 3D Body Measurement Estimation from Sparse Landmarks.

DOI: 10.5220/0012384000003660

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 19th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications (VISIGRAPP 2024) - Volume 4: VISAPP, pages

524-531

ISBN: 978-989-758-679-8; ISSN: 2184-4321

Figure 1: We use a simple Bayesian ridge model to estimate 11 body measurements (show on the right) using 72 human body

landmarks (shown on the left). We mark with green the visible landmarks on the front of the body, and with red the not visible

landmarks that are located on the back of the body.

are extracted. However, the ﬁtting is very sensitive to

the hyperparameters of the minimization (Loper et al.,

2015) and 3D scan data is often noisy. To increase the

robustness of the body measurements, some methods

extract additional features from the ﬁtted template and

use them to estimate the measurements (Yan et al.,

2020). These features are complex, using multiple

predeﬁned paths or PCA coefﬁcients of the triangle

deformation found during the ﬁtting process (Tsoli

et al., 2014).

We argue that anthropometric measurements can

be accurately learned using only sparse landmark data

instead of complex features or heuristics. Therefore,

we propose to use a simple Bayesian ridge regres-

sion model (Tipping, 2001) that takes landmark co-

ordinates as input, and outputs body measurements.

The motivation behind using the Bayesian regres-

sion model is the underlying assumption that the out-

put variables (the measurements) are normally dis-

tributed, as can be seen in the left part of Figure 2.

Furthermore, the normalized 3D coordinates of the

landmarks also reﬂect a 3D normal distribution, as

can be seen in the right part of Figure 2 (without the

outliers). Using the Bayes regression, we can encode

these distribution priors into the model.

Our approach assumes that accurate 3D body

landmarks are given, and their automatic extraction

from 3D scans is out of the scope of this work. If

3D landmarks are not given, they can be estimated

using existing methods such as (Wuhrer et al., 2010;

Luo et al., 2022; Xie et al., 2021). The key point is

that the competing methods (Tsoli et al., 2014; Hasler

et al., 2009) also assume accurate body landmarks but

use them to ﬁrstly ﬁt a template body to the 3D scan,

and then estimate the body measurements, either us-

ing predeﬁned locations or intermediate features. In

contrast to previous works, we use the landmark data

to directly predict the body measurements.

The main aim of our work is, therefore, to show

that the ﬁtting process and complex feature extraction

can be skipped, and accurate body measurements can

be estimated directly from the landmark coordinates.

The body measurements estimated using only land-

marks are comparable in accuracy to the body mea-

surements obtained by ﬁrst ﬁtting the template mod-

els. To validate our claims, we evaluate our body mea-

surement estimation approach on a public CAESAR

dataset (Robinette et al., 1999).

2 RELATED WORK

The existing approaches for automatic body measure-

ment extraction can typically be divided into three

categories: template-based, landmark-based, and di-

rect methods.

Landmark-based methods make use of the 3D

landmarks on the body to extract the measurements

from the scanned human body. To ﬁnd the landmarks,

(Zhong et al., 2018) slice the 3D scan every 5mm

along the height of the body, and search for changes

in the cross-sections to ﬁnd the landmarks; (Lu and

Wang, 2008) use 2D silhouettes and grayscale im-

age to detect landmarks in 2D, after which they are

reprojected into the 3D space; (Xie et al., 2021)

use the Mean Curvature Skeleton (Tagliasacchi et al.,

2012) to ﬁnd the segmentation of the body, after

which heuristics are used on the segmented parts

to deﬁne the landmarks; (Markiewicz et al., 2017)

Direct 3D Body Measurement Estimation from Sparse Landmarks

525

Figure 2: Distribution of measurements and landmarks. The left part of the Figure shows the distribution in mm for 4 body

measurements from the CAESAR (Robinette et al., 1999) dataset: arm length (shoulder to wrist), chest circumference, hip

circumference, and stature. As expected from body measurements, the arm length and stature seem to follow a normal

distribution; whereas the chest and hip seem to follow a positively skewed normal distribution. The right part of the Figure

shows the distribution for the Nuchale landmark over all the CAESAR dataset subjects. Since the subjects are scanned in a

normalized position, the variation of the landmark location reﬂects the change in human body shape. The Nuchale landmarks

across the subjects seem to follow a 3D normal distribution. This is observed for all the other landmarks as well.

compute the Gaussian curvature on the body to ﬁnd

characteristic points with salient curvatures; (Wang

et al., 2006) assume colored markers on the human

body during the scanning process, which they extract

in post-processing. Once the landmarks are found,

length measurements can be estimated as Euclidean

or Geodesic distances between the landmarks, and

circumference measurements can be estimated by cut-

ting the subject with a plane at a desired landmark

location and ﬁnding the cross-section. Rather than

cutting the human body, (Xiaohui et al., 2018) ﬁnd

a path on the body mesh from a set of landmarks to

determine the measurements.

Template-based methods ﬁt a statistical human

body model (template) to a 3D scan and extract com-

plex features from the template to learn the body mea-

surements. The ﬁtting process can be done by opti-

mizing over the body model parameters, and reﬁned

by non-rigid deformation (NRD) (Yan et al., 2020;

Tsoli et al., 2014; Li and Paquette, 2020; Wasen-

muller et al., 2015) which minimizes several loss

components, such as the data term, landmark term,

smoothness term and normal term. Similarly, the ﬁt-

ting process can be done using a deep learning model

(Kaashki et al., 2021; Kaashki et al., 2023), where

a 3D-CODED (Groueix et al., 2018) architecture is

used to infer the ﬁtted body. After the template body

has been ﬁtted, the measurements can be learned,

transferred, or estimated from 3D landmarks. To learn

the measurements (Tsoli et al., 2014; Li and Paquette,

2020; Yan et al., 2020) use 3D points and features ex-

tracted from the ﬁtted body to learn the measurements

using different models, such as the ElasticNet (Zou

and Hastie, 2005), SVR (Chang and Lin, 2011) and

PLS (Geladi and Kowalski, 1986) models. To trans-

fer the measurements (Kaashki et al., 2021; Kaashki

et al., 2023; Wasenmuller et al., 2015) predeﬁne the

body measurement paths on the template body, which

can be transferred onto the scan by ﬁnding the nearest

neighbor of each path point from the template to the

scan. Similarly, to estimate the measurements from

landmarks, (Wang et al., 2014; Gonzalez Tejeda and

Mayer, 2020) transfer the landmarks from the tem-

plate body to the scan, and ﬁnd the measurement as

the previously described landmark-based methods.

Differently from the template-based and

landmark-based methods, direct methods

(

Skorv

ankov

a et al., 2022; Probst et al., 2017)

learn the body measurements directly from the

frontal partial 3D scans. (

Skorv

ankov

a et al., 2022)

uses a variation of the PointNet (Qi et al., 2016)

architecture whereas (Probst et al., 2017) uses

gradient-boosted trees to predict local measurements,

which are weighted in order to compute the ﬁnal

measurements.

3 METHODOLOGY

3.1 Dataset

The CAESAR dataset (Robinette et al., 1999) is com-

prised of 4396 subjects scanned in the standing pose

with 73 annotated body landmarks. The landmarks

can be seen in Figure 3. The subjects were manu-

ally measured, and we use a subset of 11 measure-

ments following common practice (Tsoli et al., 2014),

as seen in Figure 1. We pre-process the dataset in

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

526

Figure 3: Landmarks used in this work deﬁned in the CAESAR dataset. For more details, we refer the readers to (Robinette

et al., 1999).

order to remove the subjects with missing landmarks

or body measurements. We normalize the landmarks

by centering them using the Substernale landmark lo-

cated in the middle of the chest (see Figure 3). After

normalization, we remove the Substernale landmark

from the dataset since it does not hold any more in-

formation.

Finally, we end up with 1879 male subjects and

1954 female subjects, each with 72 landmarks and 11

body measurements. Following (Tsoli et al., 2014),

we randomly sample 200 scans for each sex as a test

set and use the remaining 1679 and 1754 for training.

3.2 Method

We use the Bayesian ridge regression model (Tipping,

2001) to learn the 11 body measurements listed in

Figure 1 given the 72 body landmarks sparsely scat-

tered over a scan. The Bayesian regression is used

to include regularization parameters in the estima-

tion procedure by introducing uninformative priors

over the hyper-parameters for the precision (inverse

of variance) of the weights w and error variance σ

a linear model.

More concretely, given a set of N subjects

, y

}

n=1

, where x

∈ R

1×216

is the set of ﬂattened

72 body landmarks and y

∈ R

1×1

is the i-th body

measurement, we assume a linear model:

= x

· w + ε

, (1)

where w ∈ R

216×1

are the weights of the linear model,

and ε

∼ N (0, σ

) is the vector of normally dis-

tributed errors. Then, the output measurement follows

a Normal distribution p(y

) = N (y

| x

w, σ

). To

use the ridge regularization, the model weights w are

encoded using the normal distribution:

p(w | α) =

∏

i=1

N (w

| 0, α

−1

) (2)

where α

are the hyper-parameters for the precision of

each weight w

, modeling the strength of the prior. To

model these hyper-parameters, the Gamma distribu-

tion is used as a suitable conjugate prior for the preci-

Direct 3D Body Measurement Estimation from Sparse Landmarks

527

Table 1: MAE in mm on the CAESAR dataset for male subjects. AE denotes the allowable error. Results from (Tsoli, 2014).

Measurement (Anthroscan, 2014) (Hasler et al., 2009) (Tsoli et al., 2014) Ours AE (Gordon et al., 1989)

Ankle Circumference 13.66 5.72 5.56 6.24 4

Arm Length (Shoulder to Elbow) 13.99 12.66 13.32 6.95 -

Arm Length (Shoulder to Wrist) 14.49 13.76 12.66 9.93 -

Arm Length (Spine to Wrist) 14.71 11.81 10.40 11.38 -

Chest Circumference 13.96 15.21 13.02 18.24 15

Crotch Height 11.01 9.77 8.36 1.17 10

Head Circumference 5.51 7.46 5.59 8.92 5

Hip Circ. Max Height 16.50 18.89 19.05 17.33 -

Hip Circumference, Maximum 7.90 12.57 10.66 22.87 12

Neck Base Circumference 21.57 13.33 13.47 12.40 11

Stature 5.86 7.98 6.53 5.75 10

Average 12.65 11.74 10.78 11.02

sion of the Normal distribution:

p(α) =

∏

i=1

Gamma(α

| a, b). (3)

To model the error variance σ

, the Gamma distribu-

tion is chosen as a conjugate prior for the precision:

p(β) = Gamma(β | c, d) (4)

where β ≡ σ

−2

. Higher values of the parameters

a, b, c and d, indicate a stronger prior belief about the

corresponding precision.

Bayesian inference proceeds by computing the

posterior over all the unknowns, given the anthropom-

etry data y



, . . . y



p(w, α, σ

| y

) =

p(y

| w, α, σ

)p(w, α, σ

)

p(y

)

. (5)

Equation 5 cannot be solved in full analytically

because of the normalizing integral p(y

). Therefore,

(Tipping, 2001) resort to rewriting the posterior as:

p(w, α, σ

| y

) = p(w | y

, α, σ

)p(α, σ

| y

), (6)

and using approximations, to ﬁnally summarize the

problem as maximizing the marginal likelihood p(y

α, σ

)p(α)p(σ

). These are maximized using a gradi-

ent descent approach on the log marginal likelihood.

Finally, given a new subject with landmarks x

∗

predictions are made in terms of the distribution:

p(y

∗

| y

) =

p(y

∗

| w, α, σ

)p(w, α, σ

| y

)dwdαdσ

(7)

To implement the Bayesian ridge regression, we

use the Scikit-learn (Pedregosa et al., 2011) library

with the default values of 1e−6 for all the four pa-

rameters a, b, c and d. We model each measurement

separately, and each sex separately, resulting in

2 × 11 = 22 models.

4 EXPERIMENTS

A great challenge with comparing and evaluating 3D

anthropometric methods are the limited open-source

datasets and code implementations. Most methods

create private small scale dataset (Zhong et al., 2018;

Kaashki et al., 2021; Lu and Wang, 2008) in order

to test their method with several human body scans,

and do not share it with the community because of

privacy issues. Additionally, most of the methods do

not share their implementations (Zhong et al., 2018;

Probst et al., 2017; Tsoli et al., 2014; Lu and Wang,

2008) with the community, making a thorough com-

parison between different methods very hard.

We compare our method with one commercial

solution denominated as Anthroscan (Anthroscan,

2014) and two template-based methods (Tsoli et al.,

2014; Hasler et al., 2009), which share their results

on the proprietary CAESAR dataset (Robinette et al.,

1999) available for purchase. To make a step to-

wards more open-source 3D anthropometry research,

we share the exact 3D scans we evaluate our method

on and make our implementation available to the re-

search community.

We use the mean absolute error (MAE) metric to

compare our estimated measurements with the ground

truth ones. The MAE for a single measurement is

computed as:

MAE =

∑

i=1

− y

est

| (8)

where N is the number of subjects, y

is the

ground truth measurement for subject i and y

est

the estimate measurement for subject i. We compare

the MAE for each measurement separately and report

them in Table 1 and Table 2.

As can be seen from Table 1 and Table 2, each

method has certain advantages towards speciﬁc mea-

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

528

Table 2: MAE in mm on the CAESAR dataset for female subjects. AE denotes the allowable error. Results from (Tsoli,

2014). We bold the best result for each measurement.

Measurement (Anthroscan, 2014) (Hasler et al., 2009) (Tsoli et al., 2014) Ours AE (Gordon et al., 1989)

Ankle Circumference 7.55 6.59 6.19 5.87 4

Arm Length (Shoulder to Elbow) 11.26 8.42 6.65 6.32 6

Arm Length (Shoulder to Wrist) 11.67 10.42 10.05 7.36 -

Arm Length (Spine to Wrist) 13.19 13.40 11.87 9.84 -

Chest Circumference 12.43 13.02 12.73 17.26 15

Crotch Height 7.45 7.53 5.50 0.78 10

Head Circumference 7.44 7.45 5.91 8.42 5

Hip Circ. Max Height 17.05 18.96 18.59 20.99 -

Hip Circumference, Maximum 7.47 16.15 12.35 21.40 12

Neck Base Circumference 21.03 16.35 15.43 13.84 11

Stature 5.60 10.21 7.51 5.52 10

Average 11.10 11.68 10.25 10.69

surements. The commercial solution Anthroscan

(Anthroscan, 2014), achieves the lowest MAE on the

hip circumference measurements for both female and

male subjects. The reason for this might be that

the measurements are directly extracted from dense

scan data, without using a body template or relying

on accurate landmark data. However, the method

achieves the worst average results when comparing

all the measurements, making it less robust for hu-

man shape estimation. The template-based method

from (Hasler et al., 2009) performs slightly better (on

average) than Anthroscan. However, it does not seem

to have any advantages towards any speciﬁc measure-

ment. (Hasler et al., 2009) ﬁts the template of the

SCAPE (Anguelov et al., 2005) body model onto the

scan, and uses a linear model to predict the measure-

ments from the body model vertices. This kind of

approach, however, depends on a very sensitive, ini-

tial ﬁtting process, since the measurements are then

extracted from the template.

To address these issues, (Tsoli et al., 2014) ﬁts

a BlendSCAPE (Hirshberg et al., 2012) body model

to the 3D scan and extracts complex features from

the template to learn an ElasticNet (Zou and Hastie,

2005) linear model to predict the measurements. By

using more sophisticated features, such as a set of pre-

deﬁned paths, PCA coefﬁcients of the body model ﬁt-

ting, and limb lengths, (Tsoli et al., 2014) achieves

better results than the previous methods because, in

the end, it considers a much greater number of local

and global features. As can be seen from the Tables,

they achieve the lowest average errors.

We argue that dense 3D data and an additional

set of features are not necessary to accurately esti-

mate the body measurements and, therefore, we pro-

pose using only the 3D landmark data to estimate the

measurements. As shown in Table 1 and 2, using

only sparse landmark data, our method achieves com-

parable results to the state-of-the-art methods while

simplifying the measurement protocol. The method

achieves the best performance on the length measure-

ments, such as arm lengths, crotch height, and stature

for both sexes. Intuitively, these measurements ben-

eﬁt from using landmarks, which are directly corre-

lated to how they are manually measured. Conse-

quently, our method achieves the lowest error on the

neck base circumference, which is the second hardest

measurement to estimate, judging by the average er-

rors across all of the competing methods. When mea-

sured manually by the experts, the neck base circum-

ference is taken with a beaded chain with an alligator

clip at one end. The chain is placed and clipped to-

gether so it lies at the base of the neck and falls over

the Cervicale landmark (see Figure 3). The length

of the chain to the bead where it is clipped is mea-

sured. This measuring protocol differs severely from

the others, making the measurement harder to esti-

mate. Our method seems to beneﬁt, again, from well-

placed landmarks around the neck area, allowing a

better estimate of the measurement.

On the other hand, our method seems to strug-

gle with the chest and hip circumferences. One rea-

son for this can be attributed to the location of the

landmarks. With manual measurement, the chest cir-

cumference is located on the largest part between the

Thelion and Axilla landmarks. As can be seen from

Figure 3, however, there are no landmarks in that lo-

cation. Similarly, with manual measurement, the hip

circumference is located at the maximum protrusion

of the buttocks, approximately around the Trochante-

rion landmarks. As can be seen from Figure 3, again,

there are no landmarks on the front and back of the

body in those locations, which we know carry a lot of

variation of the human shape. Therefore, our method

needs to infer these measurements from substantially

less information than the competing methods (sparse

landmarks) and use the spatial relationship between

all the other landmarks, to infer the body proportions.

Direct 3D Body Measurement Estimation from Sparse Landmarks

529

The second reason these measurements are harder

to estimate is the assumption of the Bayesian regres-

sion to model the output variable (the measurements)

with the Normal distribution. As can be seen from

Figure 2, the chest and hip circumferences are the

only two measurements that have slightly skewed nor-

mal distributions, making them harder to generalize.

To address these issues, future work should focus on

landmark placement, making them more informative.

Rather than using multiple landmarks on the feet of

the subjects, for example, they could be replaced with

landmarks around the chest and hip area.

Finally, our average measurements for the male

and female subjects are behind the state-of-the-art

method from (Tsoli et al., 2014) by only 0.24 mm

and 0.44 mm respectively. We note, however, that

our average MAE gets skewed by the higher hip and

chest circumference errors. Comparing the median

errors, on the other hand, our method would achieve a

median MAE of 9.93 mm for the male subjects, com-

pared to 10.66 mm from (Tsoli et al., 2014); and a me-

dian MAE of 8.42 mm for the female subjects, com-

pared to 10.05 mm from (Tsoli et al., 2014). There-

fore, addressing the issue of better landmark place-

ment would greatly improve our results.

5 CONCLUSION

We propose a simple 3D body measurement estima-

tion method directly using the landmarks on the body.

Our method shows better or comparable results on

the CAESAR dataset for most measurements, and

achieves the lowest errors on the arm lengths, crotch

height, stature, and neck base circumference, for both

sexes. With our method, we show that extracting

complex features prior to measuring (such as mul-

tiple paths from the ﬁtted template model, or PCA

coefﬁcients of the triangle deformation of the ﬁtted

template model), is not necessary to accurately esti-

mate the body measurements. Additionally, we show

that ﬁtting a template body model to the scan is also

not necessary if given the location of 3D landmarks;

where accurate measurements can be directly esti-

mated from the landmarks.

To further improve the results for certain measure-

ments, such as the hip and chest circumferences, fu-

ture work will need to address the placement of the

3D landmarks, in order to provide the model with ad-

ditional information about the body shape. By cou-

pling the accurate landmark placement with a 3D

landmark extraction algorithm, our 3D body measure-

ment method could be made fully automatic.

ACKNOWLEDGEMENTS

This work has been supported by the Croatian Sci-

ence Foundation under projects IP-2018-01-8118 and

DOK-2020-01.

REFERENCES

Anguelov, D., Srinivasan, P., Koller, D., Thrun, S., Rodgers,

J., and Davis, J. (2005). Scape: Shape comple-

tion and animation of people. ACM Trans. Graph.,

24(3):408–416.

Anthroscan (2014). Human solutions gmbh. anthroscan.

http://www. https://www.human-solutions.com/. Ac-

cessed: 2023-10-06.

Bartol, K., Bojani

c, D., Petkovi

c, T., and Pribani

c, T.

(2021). A review of body measurement using 3d scan-

ning. IEEE Access, 9:67281–67301.

Camba, J. D., De Leon, A. B., de la Torre, J., Saor

ın, J. L.,

and Contero, M. (2016). Application of low-cost 3d

scanning technologies to the development of educa-

tional augmented reality content. In 2016 IEEE Fron-

tiers in Education Conference (FIE), pages 1–6.

Casadei, K. and Kiel, J. (2020). Anthropometric Measure-

ment. StatPearls Publishing, Treasure Island (FL).

Chang, C.-C. and Lin, C.-J. (2011). LIBSVM. ACM

Transactions on Intelligent Systems and Technology,

2(3):1–27.

Donli

c, M. (2019). Three-dimensional analysis of back sur-

face under dynamic conditions in scoliosis diagnos-

tics. PhD thesis, Faculty of Electrical Engineering and

Computing, University of Zagreb.

Geladi, P. and Kowalski, B. R. (1986). Partial least-squares

regression: a tutorial. Analytica Chimica Acta, 185:1–

17.

Gonzalez Tejeda, Y. and Mayer, H. A. (2020). Calvis: chest,

waist and pelvis circumference from 3d human body

meshes as ground truth for deep learning.

Gordon, C. C., Churchill, T., Clauser, C. E., Bradtmiller, B.,

and Mcconville, J. T. (1989). Anthropometric survey

of u.s. army personnel: Methods and summary statis-

tics 1988.

Groueix, T., Fisher, M., Kim, V. G., Russell, B., and Aubry,

M. (2018). 3d-coded : 3d correspondences by deep

deformation. In ECCV.

Hasler, N., Stoll, C., Sunkel, M., Rosenhahn, B., and Seidel,

H.-P. (2009). A Statistical Model of Human Pose and

Body Shape. Computer Graphics Forum.

Heymsﬁeld, S. B., Bourgeois, B., Ng, B. K., Sommer, M. J.,

Li, X., and Shepherd, J. A. (2018). Digital anthropom-

etry: a critical review. European journal of clinical

nutrition, 72(5):680–687.

Hirshberg, D. A., Loper, M., Rachlin, E., and Black, M. J.

(2012). Coregistration: Simultaneous alignment and

modeling of articulated 3d shape. In Computer Vision

– ECCV 2012, pages 242–255. Springer Berlin Hei-

delberg.

VISAPP 2024 - 19th International Conference on Computer Vision Theory and Applications

530

Kaashki, N. N., Hu, P., and Munteanu, A. (2021). Deep

learning-based automated extraction of anthropomet-

ric measurements from a single 3-d scan. IEEE Trans-

actions on Instrumentation and Measurement, 70:1–

14.

Kaashki, N. N., Hu, P., and Munteanu, A. (2023). Anet: A

deep neural network for automatic 3d anthropometric

measurement extraction. IEEE Transactions on Mul-

timedia, 25:831–844.

Li, P. and Paquette, S. (2020). Predicting anthropometric

measurements from 3d body scans: Methods and eval-

uation. In Di Nicolantonio, M., Rossi, E., and Alexan-

der, T., editors, Advances in Additive Manufacturing,

Modeling Systems and 3D Prototyping, pages 561–

570, Cham. Springer International Publishing.

Loper, M., Mahmood, N., Romero, J., Pons-Moll, G., and

Black, M. J. (2015). SMPL: A skinned multi-person

linear model. ACM Trans. Graphics (Proc. SIG-

GRAPH Asia), 34(6):248:1–248:16.

Lu, J. and Wang, M. (2008). Automated anthropometric

data collection using 3d whole body scanners. Expert

Systems with Applications, 35(1-2):407–414.

Luo, S., Zhang, Q., and Feng, J. (2022). Automatic location

and semantic labeling of landmarks on 3d human body

models. Computational Visual Media, 8(4):553–570.

Markiewicz, Ł., Witkowski, M., Sitnik, R., and Mielicka, E.

(2017). 3d anthropometric algorithms for the estima-

tion of measurements required for specialized garment

design. Expert Systems with Applications, 85:366–

385.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer,

P., Weiss, R., Dubourg, V., Vanderplas, J., Passos,

A., Cournapeau, D., Brucher, M., Perrot, M., and

Duchesnay, E. (2011). Scikit-learn: Machine learning

in Python. Journal of Machine Learning Research,

12:2825–2830.

Probst, T., Fossati, A., Salzmann, M., and Gool, L. V.

(2017). Efﬁcient model-free anthropometry from

depth data. In 2017 International Conference on 3D

Vision (3DV). IEEE.

Qi, C. R., Su, H., Mo, K., and Guibas, L. J. (2016). Pointnet:

Deep learning on point sets for 3d classiﬁcation and

segmentation. arXiv preprint arXiv:1612.00593.

Robinette, K. M., Daanen, H., and Paquet, E. (1999).

The caesar project: a 3-d surface anthropometry sur-

vey. In Second International Conference on 3-D Dig-

ital Imaging and Modeling (Cat. No.PR00062), pages

380–386.

Skorv

ankov

a, D., Rie

cick

y, A., and Madaras, M. (2022).

Automatic estimation of anthropometric human body

measurements. In Proceedings of the 17th Interna-

tional Joint Conference on Computer Vision, Imag-

ing and Computer Graphics Theory and Applications.

SCITEPRESS - Science and Technology Publications.

Tagliasacchi, A., Alhashim, I., Olson, M., and Zhang, H.

(2012). Mean curvature skeletons. Computer Graph-

ics Forum, 31(5):1735–1744.

Tipping, M. E. (2001). Sparse bayesian learning and the

relevance vector machine. J. Mach. Learn. Res.,

1:211–244.

Tsoli, A. (2014). Modeling the Human Body in 3D: Data

Registration and Human Shape Representation. PhD

thesis, Brown University, Department of Computer

Science.

Tsoli, A., Loper, M., and Black, M. J. (2014). Model-based

anthropometry: Predicting measurements from 3d hu-

man scans in multiple poses. In IEEE Winter Confer-

ence on Applications of Computer Vision. IEEE.

Wang, M.-J. J., Wu, W.-Y., Lin, K.-C., Yang, S.-N., and

Lu, J.-M. (2006). Automated anthropometric data col-

lection from three-dimensional digital human models.

The International Journal of Advanced Manufactur-

ing Technology, 32(1-2):109–115.

Wang, Q., Jagadeesh, V., Ressler, B., and Piramuthu, R.

(2014). Im2ﬁt: Fast 3d model ﬁtting and anthropo-

metrics using single consumer depth camera and syn-

thetic data. In 3D Image Processing, Measurement ,

and Applications.

Wasenmuller, O., Peters, J. C., Golyanik, V., and Stricker,

D. (2015). Precise and automatic anthropometric mea-

surement extraction using template registration. In

Proceedings of the 6th International Conference on

3D Body Scanning Technologies, Lugano, Switzer-

land, 27-28 October 2015. Hometrica Consulting - Dr.

Nicola D'Apuzzo.

Wuhrer, S., Azouz, Z. B., and Shu, C. (2010). Semi-

automatic prediction of landmarks on human models

in varying poses. In 2010 Canadian Conference on

Computer and Robot Vision, pages 136–142.

Xiaohui, T., Xiaoyu, P., Liwen, L., and Qing, X. (2018).

Automatic human body feature extraction and per-

sonal size measurement. Journal of Visual Languages

& Computing, 47:9–18.

Xie, H., Zhong, Y., Yu, Z., Hussain, A., and Chen, G.

(2021). Automatic 3d human body landmarks extrac-

tion and measurement based on mean curvature skele-

ton for tailoring. The Journal of The Textile Institute,

113(8):1677–1687.

Yan, S., Wirta, J., and K

ainen, J.-K. (2020). Anthro-

pometric clothing measurements from 3d body scans.

Machine Vision and Applications, 31(1-2).

Zakaria, N. and Gupta, D. (2019). Anthropometry, Apparel

Sizing and Design. The Textile Institute Book Series.

Elsevier Science.

Zhao, H., Chen, J., Wang, L., and Lu, H. (2023). Arkittrack:

A new diverse dataset for tracking using mobile rgb-

d data. In Proceedings of the IEEE/CVF Conference

on Computer Vision and Pattern Recognition (CVPR),

pages 5126–5135.

Zhong, Y., Li, D., Wu, G., and Hu, P. (2018). Automatic

body measurement based on slicing loops. Interna-

tional Journal of Clothing Science and Technology,

30(3):380–397.

Zou, H. and Hastie, T. (2005). Regularization and vari-

able selection via the elastic net. Journal of the Royal

Statistical Society Series B: Statistical Methodology,

67(2):301–320.

Direct 3D Body Measurement Estimation from Sparse Landmarks

531