WO2020156627A1 - The virtual caliper: rapid creation of metrically accurate avatars from 3d measurements - Google Patents

Abstract

The invention relates to a computer-implemented method for generating full body avatars of a user, comprising the steps: - Obtaining body measurements of the user; - Determining / deriving shape parameters of a virtual body model, based on the measurement data; - Generating an avatar, based on the shape parameters / the virtual body model; and - Outputting the avatar. The body measurements are set manually by the user or determined by a method comprising the steps of: - Obtaining sensor data using one or more electronic sensor devices; - Determining one or more body measurements of the user, based on the sensor data; - Outputting the determined body measurements. The electronic sensor device may be a 3D position-tracking device, e.g. an off-the-shelf game controller.

Description

The Virtual Caliper:

Rapid Creation of Metrically Accurate Avatars from 3D Measurements The present invention relates to methods and systems for the rapid creation of metrically accurate avatars from three-dimensional measurements (e.g. a virtual caliper).

TECHNICAL BACKGROUND

Creating realistic full-body 3D avatars is important for many applications such as virtual clothing try-on (K. Kristensen, N. Borum, L. G. Christensen, H. W. Jepsen, J. Lam, A. L. Brooks, and E. P. Brooks. Towards a next generation universally accessible online shopping- for-apparel system. In International Conference on Human-Computer Interaction, pp. 418- 427. Springer, 2013), ergonomics (N. Badler. Virtual humans for animation, ergonomics, and simulation. In Nonrigid and Articulated Motion Workshop, 1997. Proceedings. IEEE, pp. 28- 36. IEEE, 1997; H. Honglun, S. Shouqian, and P. Yunhe. Research on virtual human in ergonomic simulation. Computers & Industrial Engineering, 53(2) :35o- 356, 2007), medicine (A. Keizer, A. van Elburg, R. Helms, and H. C. Dijkerman. A virtual reality full body illusion improves body image disturbance in anorexia nervosa. PloS one, n(io):eoi6392i, 2016 S. Molbert, A. Thaler, B. Mohler, S. Streuber, J. Romero, M. Black, S. Zipfel, H.-O. Karnath, and K. Giel. Assessing body image in anorexia nervosa using biometric self-avatars in virtual reality: Attitudinal components) S. C.Molbert, A. Thaler, S. Streuber, M. J. Black, H.-O. Karnath, S. Zipfel, B. Mohler, and K. E. Giel. Investigating body image disturbance in anorexia nervosa using novel biometric figure rating scales: A pilot study. European Eating Disorders Review, 25(6):6o7-6i2, 20, telepresence (H. Fuchs, A. State, and J.-C. Bazin. Immersive 3d telepresence. Computer, 47(7):46-52, 2014) and gaming. More particularly, avatars for ergonomic assessment and product design have been an important part of the design process since Badler created“Jack” in the late 90s. Today virtual humans are used in design process to assess the life cycle of a product from production design, evaluation, and validation (H. Honglun, S. Shouqian, and P. Yunhe. Research on virtual human in ergonomic simulation. Computers & Industrial Engineering, 53(2):350- 356, 2007). Kristensen et al. discuss next generation on-line shopping apparel requirements and report that an estimated 15-40% of apparel purchased on-line is returned, because the clothing does not not fit or look right. A virtual and metrically accurate avatar for trying-on clothes before purchase could play an important role in reducing these costly returns/exchanges. Some medical/clinical applications require precise body models to investigate the patient’s body perception such as in Anorexia Nervosa. Some authors have used Make- Human (MakeHuman. http://www.makehuman.org/, 2018) among other avatar creation tools, to investigate self body perception. The assessment of body image disturbance in patient populations relies upon precise tools to quantify distortions. For all applications involving patients, privacy and the protection of self-identifying information (such as images) is essential. Finally, scientists have worked for over 20 years towards establishing a telepresence system that enables a person to be present digitally somewhere else in 3-dimensions both in shape and as an interactive viewer. Fuchs and colleagues describe the greatest challenge to be the real-time demands of telepresence (immersive virtual reality) systems (H. Fuchs, A. State, and J.-C. Bazin. Immersive 3d telepresence. Computer, 47(71:46-52, 2014). The combination of model-based and video-based approaches to creating avatars may be a possible hybrid solution for real-time performance demands.

While, for facial avatar creation, there has been extensive research and there are many commercial systems, the full body has received less attention. Current approaches require either skilled avatar creators (artistic approach), anthropometric measurements from trained personnel (anthropometric approach), photographs (image-based approach), or specialized hardware and software (technological approach). In all cases, current avatar creation approaches involve a time-consuming pipeline and/or require expertise. However, many applications require not only real-time creation of avatars, but also real-time performance in terms of pose and animation to realize their full potential.

There are many pros and cons to consider when designing an avatar creation tool. One major issue, which has been disregarded by most methods, is the protection of the privacy and identity of the individual. To assess the need for privacy-preserving methods, the inventors conducted an online survey (52 subjects, 27 female) that asked what data people would be willing to share for different purposes such as buying clothes online, medical applications, etc.: 80.8% preferred to share metric body measurements, followed by photographs in clothing (17.3%). People least preferred (94.2%) to share videos of themselves in underwear, which is required today for the methods that generate the most metrically accurate avatars from 3D scanning systems. Currently no tool exists with which novice users can create reliably precise avatars while protecting privacy and identity.

Additional factors need to be considered by any practical solution for full-body avatar creation. One cannot assume user expertise in 3D modeling or artistic talent. The cost and/or availability to the public is key to adoption. The process of creating an avatar should be fun, easy, and quick. Finally, it is important that the avatar creation process is computationally lightweight and allows creators access to the virtual avatar for other applications, not just as an image or embedded in an application. The summary in Figure 2 shows the requirements to consider when choosing an avatar creation tool. PRIOR ART

Previous research has, quite naturally, focused on standard tailoring measurements. Several methods have shown that relating a statistical model of 3D body shape to such measurements works well (N. Hasler, C. Stoll, M. Sunkel, B. Rosenhahn, and H.-P. Seidel. A statistical model of human pose and body shape. In Computer Graphics Forum, vol. 28, pp. 337-346. Wiley Online Library, 2009 H. Seo and N. Magnenat-Thalmann. An example-based approach to human body manipulation. Graphical Models, 66(1) 11-23, 2004 S. Wuhrer and C. Shu. Estimating 3d human shapes from measurements. Machine vision and applications, 24(6):li33-ii47, 2013). There are, however, several problems with tailoring measurements that have not previously been elucidated. First, such measurements require expertise to be taken precisely. Even trained experts vary significantly in measuring the same person. Second, tailoring measurements must be taken by another person; it is not possible to accurately measure one’s own body.

Artistic approaches allow a user to enter values for the body dimensions of the avatar. Many 3D models of avatars and avatar creation tools are available: AXYZDesign (AXYZdesign. https://secure.axyz-design.com/ metropoly-3d-people, 2018), Daz3D (Daz3D. https://www.daz3d.com/genesis8, 2018), Mixamo (] Mixa o. https://www.mixamo.com/, 201), Poser (Poser 3D Character Art and Animation Software. http://my. smithmicro.com/poser-3d-animation-software.html, 2018), Rocketbox (Rocketbox. http://www.cgsociety.org/index.php/CGSFeatures/CGSFeatureSpecial/rocketbox_character s, 2018) and MakeHuman. Artists use their training and skilled judgment of human shape to judge body proportions. But novices are not very good at producing bodies based on their judgement of dimensions. One possibility for the artist is to simply measure bodies, however measuring bodies precisely is challenging. Artists will often instead use ratios, or average values as an initial guideline.

The pioneering work of Blanz et al. (V. Blanz and T. Vetter. A morphable model for the synthesis of 3d faces. In Proceedings of the 26th annual conference on Computer graphics and interactive techniques, pp. 187-194. ACM Press/Addison-Wesley Publishing Co., 1999) proposes a method to map facial attributes, such as femininity, to the shape space of their 3D face model. Since then, many methods have approached the problem of the body shape space reparametrization. The idea is to learn a mapping between semantic labels and the body model space. By modifying these semantic features the user can explore the human body shape space. For instance, in Body Talk (S. Streuber, M. A. Quiros-Ramirez, M. Q. Hill, C. A. Hahn, S. Zuffi, A. O’Toole, and M. J. Black. Body talk: Crowdshaping realistic 3d avatars with words. ACM Transactions on Graphics (TOG), 35(4):54, 2016) a plausible 3D body is created with words that describe body shape. The focus of these approaches is usually visual plausibility and ease of use. Although height and weight are usually used, the approaches mostly use semantic values that are difficult to quantify. Thus the methods may not achieve metric precision. Other research efforts have specifically focused on the regression of avatars from body measurements by linking these measurements with a representation of the shape space. These approaches show that regressing bodies from accurate measurements is possible and works well. Seo and Magnenat-Thalmann (H. Seo and N. Magnenat-Thalmann. An example- based approach to human body manipulation. Graphical Models, 66(i):i— 23, 2004) approach the creation of new bodies as a data sample interpolation problem, and their interpolants are accurate for circumference predictions. Hasler et al. train semantically meaningful regressors to generate novel meshes according to several constraints and report metric accuracy for different linear and non-linear regression techniques. Wuhrer et al. learn a mapping between measurements and bodies from the CAESAR dataset (K. M. Robinette, S. Blackwell, H. Daanen, M. Boehmer, and S. Fleming. Civilian american and european surface anthropometry resource (caesar), final report volume 1. summary. Technical report, SYTRONICS INC DAYTON OH, 2002). As the measured bodies may lie outside their body shape space, they propose a refinement step that involves optimization, which is not well suited for real-time applications.

Interesting related works on anatomical models have been proposed. For instance Saito et al. (S. Saito, Z.-Y. Zhou, and L. Kavan. Computational bodybuilding: Anatomically-based modeling of human bodies. ACM Trans. Graph., 34(4):4i:l~4i:i2, July 2015. doi: 10.1145/2766957) create a personalized anatomical avatar by measuring some dimensions, such as bone lengths. The created models are targeted towards physics simulation.

Most of the methods that regress body shape from measurements rely on anthropometric measurements, as these remain the gold standard for body measurement. Without internal scanning of bone lengths, all dimensions are only estimates. Unavoidably, there is variability between measurers and by the same measurer over time. Typical certified anthropomorphic programs, which teach how to measure the body, are costly, and time consuming. A common measurement protocol involves taking the same measurement twice. If the relative error between both measurements lies inside a certain threshold, then the mean of the measurements is taken. In other protocols, i.e. ANSUR (C. C. Gordon, T. Churchill, C. E. Clauser, B. Bradtmiller, J. T. McConville, I. Tebbetts, and R. A. Walker. Anthropometric survey of us army personnel: Summary statistics, interim report for 1988. Technical report, ANTHROPOLOGY RESEARCH PROJECT INC YELLOW SPRINGS OH, 1989) and ISAK (International Society of the Advancement of Kinanthropometry: ISAK. http://www.isak.global/, 2018), a third measurement is taken and the median of the 3 values is used as the final value. The inventors adopted the ISAK method for their measurements. Several approaches use video cameras, either image based (RGB) or image and depth sensor (RGB-D), to estimate the human body shape and create avatars. Some methods rely on single or multiple images to compute the pose and shape of a body. Guan et al. (P. Guan, A. Weiss, A. O. Balan, and M. J. Black. Estimating human shape and pose from a single image. In Computer Vision, 2009 IEEE 12th International Conference on, pp. 1381-1388. IEEE, 2009) estimate the shape and pose of a body from a single image. To account for the ambiguity between the height and distance from the camera, the user inputs the height of the person to estimate a height-constrained body shape. Boisvert et al. (] J. Boisvert, C. Shu, S. Wuhrer, and P. Xi. Three-dimensional human shape inference from silhouettes: reconstruction and validation. Machine vision and applications, 24(1) :14s- 157, 2013) compute a metrically accurate reconstruction of the body shape from two silhouettes, assuming a perfect pose of the body. However the precision is highly influenced if the pose at acquisition time is not accurate. Recent methods use convolutional neural networks in order to infer the pose and shape of the human body from a single image. The methods using a depth sensor in addition to the images usually combine silhouette information, together with the depth and color data. The method by Bogo et al. (F. Bogo, M. J. Black, M. Loper, and J. Romero. Detailed full-body reconstructions of moving people from monocular RGB-D sequences. In International Conference on Computer Vision (ICCV), pp. 2300-2308, Dec. 2015) achieves accuracy on the order of 3mm, but requires the user to be in minimal clothing, and is computationally expensive. Li et al. (H. Li, E. Vouga, A. Gudym, L. Luo, J. T. Barron, and G. Gusev. 3D self-portraits. ACM Trans. Graph., 32(6):I87:I-I87:9, Nov. 2013. doi: 10. 1145/2508363.2508407) provide an original solution to create 3D self- portraits with a low cost depth sensor. While the method achieves impressive accuracy for static objects, the accuracy drops when scanning a real human. Bauer et al.( A. Bauer, A.-H. Dicko, F. Faure, O. Palombi, and J. Troccaz. Anatomical mirroring: Real-time user-specific anatomy in motion using a commodity depth camera. In Proceedings of the 9th International Conference on Motion in Games, MIG T6, pp. 113-122. ACM, New York, NY, USA, 2016. doi: 10.1145/2994258.2994259) propose an anatomic mirror. The length of the bones are estimated from an RGB-D sequence and used to register an anatomical model. Tong et al. (J. Tong, J. Zhou, L. Liu, Z. Pan, and H. Yan. Scanning 3d full human bodies using kinects. IEEE Transactions on Visualization and Computer Graphics, 18(4): 643-650, April 2012. doi: 10.1109/TVCG.2012.56) use a fixed setting of 3 Kinect sensors to rapidly create an avatar in about 6 minutes. The obtained results are visually appealing. The reported errors comparing the results to actual anthropometric measurements on the bodies are 3cm in arm length and 2cm in leg length. While 3D body scanning technologies are becoming less expensive, they still involve specialized equipment, trained technicians and heavy computations on powerful computers. Several studies compare the performance of commercial 3D scanning systems relative to hand measurements. As of today, scanning technologies achieve the best results in terms of accuracy. From early setups to more recent ones, the technology has evolved, allowing the creation of avatars, including the face, in one single shot.

Several approaches describe a pipeline to rapidly produce an avatar from 3D body and face scan data. While these approaches are promising, they still take more than 10 minutes and are only semi-automatic. Furthermore, they do not metrically or perceptually assess the accuracy of the avatars and, like all such approaches, suffer from privacy issues (i.e. scans in minimal clothing).

Moreover, users prefer to be scanned in clothing rather than in tight-fitting clothes. Several works have addressed the challenging problem of computing the shape of a person from only scans of the person wearing clothes. Existing methods tackle this problem either from images, a single scan, by averaging the results on each frame, or by considering a full sequence of scans.

Although not a scientific publication, a tutorial (Calibrating Height and Arm Length in VR with Unity3d. https : / / www. youtube . com / watch?time_continue= 2&v= -5GvMk4zRWs , 2017) shows how to use basic scaling to manually adjust the height and the arm length of the avatar using the HTC Vive wands. The body is uniformly scaled in the three dimensions to match the heights of the avatar and user. The arms are individually scaled (the rest of the body is not modified) by visually matching the 3D locations of the HTC wands and the virtual avatar; no 3D measurements are taken. While the height and arm span measurements are matched, the rest of the body is not captured and the correlations between different measurements are not modeled. It is therefore an object of the invention to provide a method and a system for determining body dimensions of a user and a method and a system for generating full-body avatars that require only a short amount of time, protect privacy, employ widely available means and are easily accessible by novice users.

These objects are achieved by the methods and the server according to the independent claims. Advantageous embodiments are defined in the dependent claims. According to a first aspect, the invention provides a computer-implemented method for measuring body dimensions of a human user, comprising the steps of: Obtaining sensor data using one or more electronic sensor devices; Determining one or more body measurements of the user, based on the sensor data; and Outputting the determined body measurements. The body measurements may comprise height and distance measurements of the user’s body. The height measurements may include height measurements of at least one of an overall height, a nipple height, a navel height or an inseam height of the user. The distance measurements may include at least one distance measurement of the torso region, including shoulder width, chest width (at nipple height), waist width (at navel height), torso depth (at navel height) or a hip width (at inseam height). The distance measurements may also include at least one distance measurements of arm lengths, including one of an arm span fingers (finger to finger), an arm span wrist (wrist to wrist), an arm length (shoulder to wrist) or a forearm length (inner elbow to wrist). The body measurements may consist of an overall height, an arm span fingers, an inseam height, a hip width and an arm length. The body measurements may further include a weight of the user. The electronic sensor data may comprise data indicating one or more locations on the floor. The floor sensor data may be used for calibration. The method may further comprise the step of outputting audio and / or visual information in order to guide the user in performing the measurements. The electronic sensor device is a hand-held device, e.g. a mobile phone equipped with sensors. The electronic sensor device may be a 3D position-tracking device, e.g. an off-the-shelf game controller or other.

The sensor data may comprise spatial coordinates, e.g. 3D coordinates.

The step of determining may include determining a difference between two or more sensor data, based on a chosen measure. According to a second aspect, the invention provides a computer-implemented method for generating full body avatars of a user, comprising the steps: Obtaining body measurements of the user; Determining / deriving shape parameters of a virtual body model, based on the measurement data; Generating an avatar, based on the shape parameters / the virtual body model; and Outputting the avatar. The body measurements may be set manually by the user. Determining the shape parameters may comprise applying one or more regressor functions (R.2, R4, R5, R6) to the body measurements. One or more regressor functions may be linear. A first regressor function (R2) may be applied to an overall height and a weight measurement as inputs. A second regressor function (R4) may be applied to an overall height, a weight, an arm span Fingers and an inseam height measurement. A third regressor function (R5) may be applied to an overall height, a weight, an arm span fingers, an inseam height and a hips width measurement. A fourth regressor function (R6) may be applied to an overall height, a weight, an arm span fingers, an inseam height, a hips width and an arm length measurement. Parameters of one or more regressor functions (R2, R4, R5, R6) may have been learned from empirical body shapes. The virtual body model may be a virtual SMPL model. Outputting the avatar may comprise the step of exporting the generated full-body avatar in a standard format, e.g. in an FBX file format.

According to a third aspect, the invention provides a server, adapted to execute a method for generating full body avatars of a user as described under the second aspect of the invention. The server may make the method available via an application-programming interface (API). The application-programming interface (API) may be accessible via a web browser.

In summary, the invention provides precise - and yet simpler - regressors, while focussing on the measurements that can be best performed by novice users.

In the following various embodiments of the invention are discussed in connection with the attached drawing, in which Fig. 1 shows an overview of a method according to an exemplary embodiment of the invention;

Fig. 2 shows various 3D Measurements of a real and a virtual body studied by the inventors. Fig. 3 shows Left: Repeatability of the measurements. Middle: Repeatability values for the measurer condition (self - other). Right: Accuracy of the measurements.

Fig. 4 shows a relation between the volume and the weight for the CAESAR dataset subjects.

Fig. 5 shows Relation between the 3D measurements on the SMPL mesh alignment to the subjects scans and the 3D measurement performed by the other measurer. Left: Hips Width values. Right: Arm Length values. The dashed line indicates the identity line for scale clarity.

Fig. 6 shows an exemplary use of a method according to an embodiment of the invention.

Fig. 7 shows results obtained with Unite the People. First triplet in clothed condition (UPC). Second triplet in minimal condition (UPM). A: input image. B: detected body parts. C: obtained body shape and pose.

Fig. 8 shows unnatural body shapes generated with the R6 regressor and the self measurements.

Fig. 9 shows a visualization of the surface errors displayed on the average female body for the proposed and previously available methods. Dark blue indicates zero error, red indicates an error of 5 cm or more. The proposed methods outperform the previous methods in terms of error when compared to the scan. In R5L, hip width correction is applied resulting in a reduced error when compared to R5 both in self and other condition. Notice that R6sL resulting in unnatural body shapes has a high surface error.

Fig. 10 shows a perceptual evaluation. Participants performed two tasks on the desktop and ranked printed images.

Fig. 11 shows overall similarity ratings of the 15 different body images. The methods are ranked with respect to their mean rating. The error bars represent standard errors of the mean. Fig. 12 shows rankings of the 15 body images (1 - least similar, 15 - most similar).

Fig. 13 shows signed percent error of the adjusted body in the Method of Adjustment task on the Desktop compared to the Fit (by gender condition). Negative values indicate underestimation of body dimensions as compared to the Fit.

DETAILED DESCRIPTION

Figure l shows an overview of a method 100 according to an exemplary embodiment of the invention. 3D measurements are performed on real (step 110) and virtual (step 130) bodies. After selecting reliable measurements in step 120, relations between the real and virtual body measurements are determined in step 140. Based on the identified reliable measurements, regressors Ri to R6 are trained in step 150. A user makes sparse 3D self measurements in step 160, e.g. hip width and arm length. One of the trained regressor Ri to R6 is applied to these measurements for creating a metrically accurate personalized avatar in step 170.

Figure 2 shows various 3D Measurements of a real and a virtual body studied by the inventors. Shown, in particular, are four heights, five in the torso and four in the arms, each for a real and for a virtual body.

In order to identify real, easy to instruct measurements on the human body that laypersons can perform on their own in a repeatable and accurate way, the present embodiment takes advantage of new proliferating technologies, such as the HTC Vive. In order to make rapid and precise measurements of the human body, a software tool in the Unity game engine employs the SteamVR tracking technology of the HTC Vive system. It allows the user to make position and distance measurements on the human body using the two HTC Vive wand controllers. The measurement setup uses two SteamVR Lighthouse base stations. An initial floor calibration routine ensures accurate height measurements by guiding the participant to measure 5 different locations on the floor around the center of the tracking space. To take each measurement, the user places the tip of the wand controller handle on the floor surface and presses the trigger button. In order to reduce the number of false samples, the trigger button should be pressed at least for one second. Successful data capture of the wand controller tip location is confirmed with visual and auditory signals. After the floor calibration, the floor offset is calculated as the median height of the five samples and applied to the wand controller pose to measure the correct height above ground. The tool then uses instructional videos to guide the user through the measurement steps. In total, contemplates thirteen 3D measurements that can be easily measured on real as well as on virtual bodies. They are distributed in four height measurements and nine distance measurements. The measured heights are: overall height, nipple height, navel height and inseam height. Five other distances focus on the torso region: shoulder width, chest width (at nipple height), waist width (at navel height), torso depth (at navel height) and hip width (at inseam height). Four arm lengths were measured by holding the two controllers in a special way: arm span fingers (finger to finger), arm span wrist (wrist to wrist), arm length (shoulder to wrist) and forearm length (inner elbow to wrist) (see Fig. 2). The virtual 3D measurements are defined on the SMPL mesh template with simple vertex to vertex Euclidian distances (see also Fig. 2). During setup, the vertex indices are identified manually at the desired locations.

To identify which measurements can be repeatably and accurately performed on the self, the inventors conducted a user study and collected data from 20 participants who used the new 3D measuring device. In addition to the 3D measurements, the participants’ bodies were 3D scanned, anthropometric measurements were performed, and photographs were taken in different clothing conditions (see Fig. 7), enabling the quantitative evaluation of the proposed approach. Twenty participants (gender-balanced, mean age = 29.15 (SD= 6.92), mean BMI = 21.08 (SD = 2.44), two left handed and all with normal or corrected-to-normal vision) took part in the experiment. The experimental protocol was approved by the local ethics committee and was performed in accordance with the Declaration of Helsinki. All participants gave written informed consent for their participation.

In order to study repeatability, each of the 13 body measurements was repeated three times. To study accuracy, in addition to the participant measures (self-measures), a coordinator also measured each participant (other-measures) three times. In the following, the measures of the user taken by himself are referred to as self. The measures of the user taken by the coordinator are referred to as other. A total of 78 measurements per participant were taken (3 repetitions x two measurers x 13 3D measurements). The self-measures took « 11 sec per measurement, while the other-measures took « 18 sec per measurement.

Figure 3A shows statistics on a repeatability of the measurements. To assess repeatability, for each subject and 3D measurement, the range of the 3 repetitions was determined by taking the difference between the maximum and the minimum. To obtain comparable values across measurements and subjects the range was normalized, dividing by the mean of the 3 measurements. Then, for each 3D measurement, the mean and standard deviations of all values were determined (see Fig. 3A). Figure 3B shows statistics for a repeatability values for the measurer condition (self - other). The 3D measurements were sorted from the most repeatable to the least repeatable according to the other condition as this shows potentially how repeatable the measurements are with the proper instruction.

Figure 3C shows statistics of an accuracy of the measurements. To assess accuracy, the 3D measurement value was determined using the ISAK protocol from the 3 repetitions of each 3D measurement. If the two first measurements are below 1,5% difference, their mean is computed. If they are over 1,5% difference, a third measurement is taken and the median of the three is used. In order to study the deviation between the self and other measurements, their relative error was determined by considering the other measure as the reference.

Overall Height and Arm Span Fingers are the most repeatable and accurate. For the torso height, it was decided to retain Inseam Height. Although Navel Height and Nipple Height are slightly more accurate, Inseam Height was selected, as it is more relevant for the clothing industry. Arm Span Wrist was not selected as a representative measure, as Arm Span Fingers is already accounting for the Arm Span. The next best measurements are Arm Length and Hip Width. It was decided to include them both. The rest of the measurements (Chest Width, Shoulder Width, Torso Depth, Forearm Length and Waist Width) were not retained as they were too inaccurately performed. The errors in waist and chest regions were expected, as even personnel trained in anthropometric measurements identify the waist and chest as the most variable dimensions. The five selected 3D measurements are Overall Height, Arm Span Fingers, Inseam Height, Hip Width, and Arm Length. Additionally, the weight of the subject may also be used, as it is a fairly easy measure to obtain in any household or laboratory.

In order to transfer real measurements to a virtual body, corresponding measurements must be made on the virtual body. While computing 3D distances between two vertices is easy, measuring the weight of a virtual body is not straightforward.

To obtain an approximation of the weight of a virtual body, a relation between the weight and the volume is. The SMPL body model is registered to all female and male subjects of the CAESAR dataset, which contains the weight measurement of all the subjects, and compute the volume of the obtained meshes (registrations). Figure 4 shows a scatter plot of the weight of the female (left) and male (right) CAESAR subjects as a function of the volume of the SMPL fit. The volume was determined by registering the SMPL model to the CAESAR scans. The weight of each subject is provided in the CAESAR measurements dataset. Volume and weight are heavily linearly correlated. The linearity seems to break down for the very heavy subjects. A clear linear relation appears, and two simple linear regressors may be learned using iterative reweighted least squares (IRLS), one for the female and one for the male subjects, as the correlations are slightly different for each gender. The learned regression values for females are (coef. 1001.44, intercept -2.36) and for males (coef. 1056.44, intercept -5.28). The estimated density values match existing clinical estimations (] H. J. Kr2ywicki and K. S. Chinn. Human body density and fat of an adult male population as measured by water displacement. The American journal of clinical nutrition, 20(41:305-310, 196).

In order to link the real and the virtual 3D measurements, the SMPL body model was aligned to the scans of the 20 participants and the SMPL meshes were virtually measured. For two measurements, significant differences were observed: Hip Width and Arm Length (see Fig. 3). Indeed, when performing these measurements the pose of the bodies in the real world is not the same as in the virtual world. For instance, the Hips Width is measured in the real world by instructing the participant to stand with both feet together. However, the virtual meshes are measured with the feet at T-pose (see Fig. 2).

Figure 5 shows a relation between the 3D measurements on the SMPL mesh alignment to the subjects’ scans and the 3D measurement performed by the other measurer. Left: Hips Width values. Right: Arm Length values. The dashed line indicates the identity line for scale clarity.

A mapping between the 3D measures obtained with the real humans and the virtual ones is determined. Because of the relatively low number of samples, a K-i-fold (leave one out) training technique is to learn two linear mappings, one for Hips Width and one for Arm Length.

The last step is to find the relation between the 3D measurements and the body shape space. First, an overview of the SMPL body model is given, then an analysis of the relation between the measurements and its shape space, and finally a determination of regressors.

In summary, the SMPL model is a function M(b,q ) that takes pose (q) and shape [/?] parameters and produces a watertight triangulated mesh M with N = 6890 vertices and F = 13,776 triangles. The SMPL model function outputs the vertex locations of the triangulated surface. The shape parameters b are the coefficients of a low-dimensional shape space, learned from thousands of registered scans. In this embodiment, 10 coefficients are used: b€ R¹⁰. The pose of the body is determined by angular rotations in a kinematic structure containing K = 23 joints. Every relative rotation between parts is parameterized using the axis-angle representation. Hence, the full pose, q R ⁷² , consists of 23 x 3 + 3 parameters, 3 parameters per joint plus 3 for the global orientation. The global translation t adds 3 additional parameters. SMPL relies on a linear blend skinning (LBS) function, that takes the unposed vertices in the rest pose (or zero

pose), , joint locations, J, a pose, q , and the blend weights, W, and returns the posed vertices. SMPL effectively parametrizes the skinning function with pose and shape by

where are vectors of vertices representing offsets from the

mean shape . These are referred to as shape and pose blend shapes respectively. The joint locations are inferred using a learned sparse regressor matrix, from the

unposed shape, that is . For more details refer to Loper et al. (. Loper,

N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. Smpl: A skinned multi-person linear model. ACM Transactions on Graphics (TOG), 34(6): 1248: 1-248: 16, Oct. 2015. doi: 10.1145/2816795.2818013).

One important property of SMPL is that the body vertices have a linear relation with the underlying shape space. Because the vertex shape space was learned using principal component analysis (PCA), the relation between a modification of a shape space parameter and the subsequent modification of the Euclidean distance between two vertices, is purely linear. 3D measurements relate linearly to the SMPL shape space. The volume of a mesh is computed by adding the signed volumes described by every triangle of the mesh and an arbitrary point. This computation has the property that the relationship between a shape space parameter displacement and the change in volume is cubic; i.e. each vertex undergoes a linear displacement with a modification of a shape space parameter. Thus, the cubic root of the volume of a mesh also has a linear relationship with the underlying shape space parameter. MEASUREMENTS TO BODY SHAPE REGRESSORS

The linear relation between the measurements and the body shape space is one of the inventor’s key observations that allows to train very simple, yet accurate, linear regressors. In different embodiments of the invention, the following regressors may be built, taking, 2, 4, 5 and 6 measurements respectively as input: R2 (from Overall Height and Weight), R4 (from Overall Height, Weight, Arm Span Fingers and Inseam Height), R5 (from Overall Height, Weight, Arm Span Fingers, Inseam Height and Hips width) and R6 (from Overall Height, Weight, Arm Span Fingers, Inseam Height, Hips width and Arm Length).

One may use the first 10 PCA components of the SMPL body shape space, and generate training subjects by sampling all the corners of the 10-dimensional hyper space at the {-2, +2} standard deviations locations. Specifically, one may generate and measure 2¹⁰ = 1024 bodies. A 1024x6 matrix may be constructed containing the 6 measurements for all the 1024 bodies. From this matrix, the five linear regressors may be learned with a least squares computation.

Figure 6 shows an exemplary use of a method according to an embodiment of the invention. A novice user takes 3D measurements of her body using the wand controllers of the HTC Vive in just under 5 minutes by following simple instructional videos. Additionally, the user is asked to enter her weight and gender. The measurements are fed to the regressors, and the SMPL model creates an accurate avatar from the resulting shape parameters. The rigged 3D model can be used straight away in Unity or exported for use in other 3D modeling packages. Figure 6 (last image) illustrates a virtual mirror application where the user can look down and experience her body in VR.

EXPERIMENTAL EVALUATION

Figure 7 shows results obtained with Unite the People. First triplet in clothed condition (UPC). Second triplet in minimal condition (UPM). A: input image. B: detected body parts. C: obtained body shape and pose.

The goal of the experimental evaluation is to determine the performance of a regressor in terms of metric accuracy, which provides a visually plausible body. The perceptual validation is important for two reasons. First, some generated bodies with reasonable metric accuracy may not look human at all. Second, it enables us to investigate to what extent the generation of metrically accurate avatars can rely on human perception. To compare metrically, bodies were generated using the open-source methods MakeHuman and Unite the People. In addition, a perceptual comparison to BodyVisualizer was made. Make Human (MHi, MH2). MakeHuman is an open-source toolkit that can be used to create a personalized avatar. Specifically, one can set the height (in meters) and the weight (as a percentage of the average) of the avatar. Under the“measures” tab one can additionally set the dimensions of many local body parts in terms of dimensions and circumferences. Importantly, while adjusting height and weight influences the whole body, the changes to specific body part dimensions do not influence the measurements of neighboring body parts. For each participant two MakeHuman avatars were created. MHi is a height and weight- matched avatar, where 100% in MakeHuman was set to be the world average in BMI (female = 25, male = 26), and a participant’s BMI was calculated as a percentage of this world average. MH2 was matched in height and weight, and additionally the arm length (upper arm, and lower arm length) as well as inseam height (upper leg, and lower leg length) were adjusted to match the participant’s measurements.

Unite the People (UPC, UPM). two pictures of the participants were used, in minimal and cloth conditions and obtained one body shape per picture using Lassner et al. (C. Lassner, J. Romero, M. Kiefel, F. Bogo, M. J. Black, and P. V. Gehler. Unite the people: Closing the loop between 3d and 2d human representations. In IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), July 2017) (see Fig. 8). The cloth condition result is referred to as UPC, and the minimal condition is referred to as UPM. It is important to note that Unite the People uses a gender-neutral body model. The authors informed us that the method could not easily be adapted to gender as the CNN was trained using a gender-neutral model

While this allows the method to be fully automatic, the resulting body shapes for both males and females look rather male. BodyVisualizer (BV). BodyVisualizer is an web-based tool where one enters anthropometric measurements of a person and visualizes the predicted body. As the web application does not export the result, it was not possible to compare to BodyVisualizer metrically. However, for the perceptual evaluation, for each participant, BodyVisualizer was used to create a body using body measurements performed by an ISAKcertified measurer and a screen-shot of the resulting body was taken. This body shape is referred to as BV. Body Scan (Fit, Scan). The 3D scan of the participants was used to compute two body shape meshes. The public SMPL model approximates the human body using 10 principal components. Consequently it is evaluated how well the predicted avatars match the SMPL approximation of a subject. To that end, the inventors optimized for the first ten components of the shape space, b, that best fit the scan data in terms of point-to-surface distance. The obtained mesh is referred to as the Fit. In addition, the SMPL mesh surface is allowed to freely deform in order to best fit the scan data in terms of point-to-surface distance. The obtained mesh is considered to be the reference and referred to as Scan. Certified Anthropometric Measurements. To compare the 3D measurements to a reference, an ISAK certified technician took 3D measurements using measuring tape and mechanical calipers.

Four aspects of the inventive method are quantitatively assessed: 1) the regressors’ precision by comparing the expected and obtained values; 2) the measurements obtained with The Virtual Caliper by comparing participants’ 3D measurements to anthropometric measurements; 3) the difference between the obtained mesh surfaces to the body surfaces (Fit and Scan) obtained with the body scanner; and 4) the robustness of The Virtual Caliper under different clothing conditions.

Regressor Precision. A regressor takes a set of measurements as input, and creates a body that should ideally fulfill these measurements. For each regressor (R₂, R₄, R₅, and R6) used the acquired measurements (self and other) and computed the mean absolute error between the input measurements and the measurements of the produced avatar.

Table 1 reports the mean absolute error and max error averaged across self and other measurements for all subjects. The values not used in the regressor are shown in bold.

Table 1. Metric differences between the input measurements obtained with the The Virtual Caliper and the resulting avatar for the different regressors, reported as the mean absolute error of all generated body shapes (self and other measurements). Units are in cm for all values but weight (W), in kg. Values in bold were not used as input in the respective regressor. Legend: H: Overall Height, W: Weight, AS: Arm Span Fingers, I: Inseam Height, HW: Hip Width, AL: Arm Length.

As expected, due to the linear relationship between the measurements and the shape space parameters, the errors between the input and output body dimensions are consistently zero or very small; i.e. on the order of millimeters for the Hip Width and Arm Length. Minor errors for the weight occur for R.2, R4, and R5, (averaging 200 grams). These errors are likely to arise due to rounding errors in the computations. R6 showed a significantly higher maximum error of 5 kg.

Figure 8 shows unnatural body shapes generated with the R6 regressor and the self measurements.

When visually examining the bodies, R6 was found to produce unnatural body shapes (see figure. 8). Therefore, the error was also computed using only the measurement values of the other measurer, resulting in a significantly reduced error. The versions using self and other measurements are referred to as R6s and R60 respectively and this notation is used in the remainder of the application.

HTC-Vive self and other to anthropometric measurements. To investigate the precision of the measurements using The Virtual Caliper (for self and other measurers), the measurements were compared to the anthropometric hand measurements performed by the ISAK certified measurer. The Mean Absolute Errors (MAE) and Relative Errors (RE) between the measurements are shown in Table 2.

Table 2. Mean Absolute Errors (MAE) and Relative Error (RE) between the measurements obtained with the The Virtual Caliper for the self and other measurer, compared to ISAK certified measurements. Most errors are below 2% both for self and other measurements. The biggest error for self measurements was observed for the Hip Width and Arm Length indicating that those are either more difficult to measure, or instructions about the measurement location should be improved. Body surface distance. The surface-to-surface distance between the generated bodies and the Scan was computed. This is especially important as the measurements of the generated bodies could be precise even when the generated body has a non-human-like body shape, as observed for R6s. As the surface distance between two meshes with different heights depends on their rigid registration relative to each other, all avatars were set on a common floor and centered horizontally and in depth using their bounding boxes.

For this evaluation an additional variant of the R5 regressor was considered: R50 and R5S which use the raw values of the HTC Vive, and R5sL and R50L which use the learned mapping for the Hips Width measurement. The goal is to quantitatively evaluate the impact of the learned mapping on the surface distances. With a consistent notation R6sL and R60L refer to the R6 regressor using respectively the self and other measurements, and with the learned mappings.

Figure 9 shows heat maps of the surface errors displayed on the average female body for the proposed and previously available methods. Dark blue indicates zero error, red indicates an error of 5 cm or more. The proposed methods outperform the previous methods in terms of error when compared to the scan. In R5L, hip width correction is applied resulting in a reduced error when compared to R5 both in self and other condition. Notice that R6sL resulting in unnatural body shapes has a high surface error.

Table 4 reports the numeric surface distance values.

Table 4. Surface to surface results of the created avatars. Mean and Max reported values are in centimeters.

The results of the computed surface-to-surface distances show that the proposed regressors metrically outperform MakeHuman and Unite the People in terms of their deviation from participants’ body scans. R20, R40 and R50L obtain the best metric results with an accuracy of 1.11- 1.18 cm. The results also show that for the proposed methods, the other measurements consistently have less error than self measurements. It is interesting to note that because the avatars were computed with the R5 regressors without the Hips Width correction, the lower torso part is excessively protruded. RssL and R50L, using the Hips Width mapping, properly recover this error and better explain the hip area. Robustness to clothing conditions. To evaluate the robustness of the method with respect to different clothing conditions, 8 participants were invited back (gender balanced), and asked them to perform the measurements in three types of clothing: minimal, loose (wearing long sports pants, and a loose t-shirt), and own clothes (the street clothes they were naturally wearing).

The results are presented in Table 3:

Table 3. Relative Error (RE) between measurements by the same subjects in different clothes. Clothing minimal is used as reference. Relative Range (RR) between measurements obtained with different clothes.

The relative errors in the measurements are consistent with the ones observed in Table 2 and the relative ranges consistent with the observed ranges in Figure 5. The Virtual Caliper is robust to different clothing conditions and does not require minimal clothing to obtain faithful measurements.

PERCEPTUAL EVALUATION

Figure 10 shows a perceptual evaluation. Participants performed two tasks on the desktop and ranked printed images.

To examine how the bodies generated with the different techniques are perceived and evaluated, all participants from the first study were invited back for a perceptual evaluation session comprising three short experiments (see Fig. 10). Out of the 20 participants who were scanned and measured, 18 (nine males, nine females) took part in the perceptual study. The perceptual evaluation session took place five weeks after the scan session and took approximately 60 minutes. In two of the three experiments, participants evaluated images of personalized bodies in terms of similarity to their own body. For each participant, 15 different body stimuli were generated and images of the body stimuli were rendered. As in the quantitative evaluation, the stimuli set comprised the Fit, the bodies generated based on the self and other measurements using R2, R4, and R5, as well as R50L, R60L, MakeHuman (MHi, MH2), and United the People (UPC, UPM). Additionally, a body generated with BodyVisualizer was added, as well as the Scan. It was predicted that the scan would receive the highest similarity ratings as it contains most identity-specific features. Further, as the Fit is a good approximation of the body shapes, it should also be among the highest rated bodies. In the third experiment, participants adjusted an avatar to their perceived own body dimensions (method of adjustment) using three of the developed regressors. Participants completed the experiments in the same order as presented below.

EXPERIMENTS

Overall Similarity Ratings. The experiment was created in Open- Sesame (v.3.1.9) and presented on a 21” monitor. In each trial, participants were presented with one of the 15 personalized images and were asked to rate how similar the body was to their own body on a 7-point Likert scale (1 - not at all, 7 - very). Participants could view each image as closely and as long as they wanted. The order of the image presentation was randomized across participants. The task took approximately 5 minutes.

Ranking. For each participant, the 15 personalized bodies were printed on A4 paper sheets. Participants stood at a large table and were instructed to carefully view the bodies and arrange them from least (left) to most (right) similar to their own body in terms of body dimensions and shape. Once completed they wrote the ranking on the images from 1 to 15, resulting in a forced-ranking of all bodies. The task took approximately 5-10 minutes.

Desktop Method of Adjustment. For the method of adjustment task, 3 different regressors were used: R2, R4, R6 where the body dimensions could be individually adjusted. The desktop application was programmed in Unity game engine and was displayed on a 2i”monitor. Since many on-line tools for clothing try-on allow the user to enter the measurement values directly, participants were asked to adjust a gender-matched avatar to their own body dimensions with the adjusted metric values being visible. They were asked to perform this task hierarchically for three regressors, by first adjusting height and weight, then inseam height and arm span, and finally arm length and hip width. Participants were specifically instructed to avoid creating statistically implausible body shapes as indicated by the shape space values shown in the application turning red. Once the participants were done adjusting the body, the six values were recorded. The task took approximately 10 minutes. Figure n shows overall similarity ratings of the 15 different body images. The methods are ranked with respect to their mean rating. The error bars represent standard errors of the mean.

Overall Similarity Ratings. Figure 11 shows the results of the overall similarity ratings. As expected, the Scan was rated as most similar to the participants’ bodies. To examine which of the methods received significantly lower similarity ratings as compared to the Scan, planned comparisons using paired t-tests with p-value correction for multiple comparisons were conducted. The results show that MH2, UPC, and UPM received lower similarity ratings than the Scan (all p-values < .05). There was no significant difference between the similarity ratings of the Scan and BV, R2S, Fit, R50L, R4S, R20, R40, R50, R5S, R60L, and MHi, indicating that the bodies generated with these methods were statistically similarly rated.

Figure 12 shows rankings of the 15 body images (1 - least similar, 15 - most similar). Ranking. Figure 12 shows the results of the image ranking task. To examine which of the methods received significantly lower ranking values as compared to the Scan, planned comparisons using paired ttests with p-value correction for multiple comparisons were conducted. The results show that R50, R5S, R60L, MHi, MH2, UPC, and UPM received significantly lower ranking values (all p-values < .05). There was no significant difference between the ranking of the scan and the ranking of BV, R2s, Fit, R50L, R4S, R20, and R40, indicating that they are statistically visually similar.

For each participant, the resulting body from the Method of Adjustment task was compared to the body dimensions of the Fit. Figure 13 shows signed percent error of the adjusted body in the Method of Adjustment task on the Desktop compared to the Fit (by gender condition). Negative values indicate underestimation of body dimensions as compared to the Fit. Since the metric values were visible, participants were accurate when adjusting their body heights. However, in terms of weight, arm span, inseam, arm lengths and hip width the adjusted values were not so accurate, with e.g. errors over 5% for the arm span and arm length. This finding is in line with previous studies on visual body perception showing that one can not rely on human perception to create metrically accurate bodies.

DISCUSSION OF EXPERIMENTAL RESULTS

As stated in the beginning of the Experimental Evaluation, the main goal is to find the best regressor providing metric accuracy as well as a perceptually appealing body. Table 1 indicates that regressors with the most measurements should be preferred, as they faithfully generate bodies with the accurate measurements. However, the limits of the present approach with R6. The self measurements produced unrealistic avatars (Fig. 8), and even the bodies created with R60L were poorly ranked. The regressor R5L, which uses Overall Height, Weight, Arm Span Fingers, Inseam Height and Hips Width as input, generates accurate metric results both in the dimensions (Table 1) and in surface error (Table 4). It also obtains as good overall similarity ratings as the Scan (Fig. 11) and is among the best ranked bodies (Fig. 12). While R4 obtains slightly better metric surface errors than R5L (Table 4) and good perceptual ratings (Fig. 11) and rankings (Fig. 12), it has errors of almost 7 cm in the Hip Width dimension (Table 1) as it is not used as input. For these reasons, it is concluded that the best overall performing regressor is R5L.

In a second level, the hypothesis was confirmed that human perception is not suited to create metrically accurate avatars. The present perceptual study provides evidence in this direction, showing that even bodies with significant differences in the Hips Width obtained high overall rating, like for instance R20 and R2s having mean errors of 2cm. This further argues for the need of The Virtual Caliper. Users can not rely on their perception and need a tool to precisely measure themselves.

The results of the metric evaluation showed that the inventive methods systematically outperform the open-source state-of-the-art methods Make- Human and Unite the People. However, the achieved metric accuracy when comparing to the Scan is on the order of 1:1cm, and so The Virtual Caliper is less accurate than the RGB-D methods, reporting errors on the order of 2:4mm, 3mm and 2:4mm respectively. If the protection of privacy and identity are not required, these methods should be preferred in terms of accuracy.

In a further embodiment, body measurements may also include capturing the face. For some applications, such as VR social media interactions, the face is important for identifying an individual. For this case, already available hardware such as the iPhoneX could be used; the AR-Kit extracts a personalized mesh of the face, which could be combined with the inventive mesh. RAPID AVATAR CREATION TOOLS

Two pieces of software are made available for non-commercial purposes at www.hiddenforreview.org. The first tool is The Virtual Caliper; i.e. the HTC Vive measurement tool. The user is guided with simple videos to perform the 3D measurements needed for the creation of the avatars. In order to account for the offset of the HTC Vive initial calibration, the user is instructed to measure the floor at five locations. Then the user is guided to perform the 3D measurements. In the submitted version, the measured values are exported into a text file. The second tool is an interactive Desktop application that integrates the four regressors, R2, R4, R5 and R6. It allows users to manually set measurement values and export the created avatars in the FBX file format for content creation tools and game engines. The values from the first tool can be used in the second one to generate a rigged avatar model that matches the measured user body dimensions. Because the avatars are built upon the openly available SMPL model (M. Loper, N. Mahmood, J. Romero, G. Pons-Moll, and M. J. Black. Smpl: A skinned multi-person linear model. ACM Transactions on Graphics (TOG), 34(6):248:I-248:I6, Oct. 2015. doi: 10.1145/2816795.2818013), they can be easily animated and posed in real-time. The desktop application is available on OSX and Windows.

CONCLUSIONS

The invention empowers novice users to create a metrically precise avatar in real-time with their own hands. It can be used to precisely measure one’s own or another person’s body and automatically creates an animatable avatar. The invention exploits an inexpensive and easy to use VR system (HTC Vive). It also provides a desktop application tool for creating bodies and editing body shape. The inventive approach overcomes potential privacy issues with previous methods in that no pictures of the user are required and the protocol can be performed in clothing.

While the produced software tools use the HTC Vive, any other device that can measure precise 3D distances, such as Oculus Rift controllers or two simple reflective markers of a motion capture system, may also be used. Measurements obtained using these methods may then be entered in the Desktop application to produce the metrically accurate 3D model or animatable avatar. The obtained results show that the precision of the measurements play an important role, especially as the number of measurements used in the regressor increases. The user performance has two tangled factors: the landmark location and the human pose. The first is the capability of the user to identify the body landmark where the measurement needs to be taken. The pose adopted by the user while performing the measurement may be taken into account by equipping the users with motion capture markers or IMU sensors in order to obtain the pose of the subject while performing the measurements.

Further, measurement of the waist and chest regions may be difficult, due to the combination of soft tissue as well as the breathing cycle. These body regions are also known to be challenging even for anthropometric measurers. According to a further embodiment of the invention, the user may start with a basic embodiment of the invention to generate an initial body. Then, the user may locally refine the body parts that do not match by pointing with the wand controller to the virtual and physical body while experiencing a first-person avatar in either augmented or virtual reality. This approach is similar to the pipeline proposed by Wuhrer et al.

Finally, further embodiments of the invention may not only include punctual measurements, where the user holds the trigger of the wand controller for one second and the measurement is saved, but continuous measurements relevant to further refine the body shape. For example, dynamic measurements of the torso during the breathing cycle may be taken into account.

Claims

Claims

1. Computer-implemented method for measuring body dimensions of a (human) user, comprising the steps of:

Obtaining sensor data using one or more electronic sensor devices;

Determining one or more body measurements of the user, based on the sensor data; Outputting the determined body measurements.

2. The method of claim l, wherein the body measurements comprise height and distance measurements of the user’s body.

3. The method of claim 2, wherein the height measurements include height

measurements of at least one of an overall height, a nipple height, a navel height or an inseam height of the user.

4. The method of claims 2 or 3, wherein the distance measurements include at least one distance measurement of the torso region, including shoulder width, chest width (at nipple height), waist width (at navel height), torso depth (at navel height) or a hip width (at inseam height).

5. The method of claims 3 or 4, wherein the distance measurements include at least one distance measurements of arm lengths, including one of an arm span fingers (finger to finger), an arm span wrist (wrist to wrist), an arm length (shoulder to wrist) or a forearm length (inner elbow to wrist).

6. The method of claim 1, wherein the body measurements consist of an overall height, an arm span fingers, an inseam height, a hip width and an arm length.

7. The method of claim 3, wherein the body measurements further include a weight of the user.

8. The method of claim 1, wherein the electronic sensor data comprises data indicating one or more locations on the floor.

9; The method of claim 8, wherein the floor sensor data is used for calibration.

10. The method of claim l, further comprising the step of outputting audio and / or visual information in order to guide the user in performing the measurements.

11. The method of claim l, wherein the electronic sensor device is a hand-held device, e.g. a mobile phone equipped with sensors.

12. The method of claim l, wherein the electronic sensor device is a 3D position-tracking device.

13. The method of claim 12, wherein the 3D position-tracking device is an off-the-shelf game controller.

14. The method of claim 1, wherein the sensor data comprises spatial coordinates, e.g. 3D coordinates.

15. The method of claim 1, wherein determining includes determining a difference

between two or more sensor data, based on a chosen measure.

16. Computer-implemented method for generating full body avatars of a user, comprising the steps:

' - Obtaining body measurements of the user, determined by a method according to one of claims 1 to 15;

Determining / deriving shape parameters of a virtual body model, based on the measurement data;

Generating an avatar, based on the shape parameters / the virtual body model; and Outputting the avatar.

17. The method of claim 16, wherein the body measurements are set manually by the

user.

18. The method of claim 16, wherein determining the shape parameters comprises

applying one or more regressor functions (R.2, R4, R5, R6) to the body measurements.

19. The method of claim 18, wherein one or more regressor functions are linear.

20. The method of claim 18, wherein a first regressor function (R2) is applied to an overall height and a weight measurement as inputs.

21. The method of claim 18, wherein a second regressor function (R4) is applied to an overall height, a weight, an arm span Fingers and an inseam height measurement.

22. The method of claim 18, wherein a third regressor function (R5) is applied to an

overall height, a weight, an arm span fingers, an inseam height and a hips width measurement.

23. The method of claim 18, wherein a fourth regressor function (R6) is applied to an overall height, a weight, an arm span fingers, an inseam height, a hips width and an arm length measurement.

24. The method of claim 18, wherein parameters of one or more regressor functions (R2, R4, R5, R6) have been learned from empirical body shapes.

25. The method of claim 16, wherein the virtual body model is a virtual SMPL model.

26. The method of claim 16, wherein outputting the avatar comprises the step of

exporting the generated full-body avatar in a standard format, e.g. in an FBX file format.

27. Server, adapted to execute a method according to claim 16.

28. The server of claim 27, making the method available via an application-programming interface (API).

29. The server according to claim 28, wherein the application-programming interface (API) can be accessed via a web browser.

30. Computer-program product, comprising instructions that, when executed on a computer, implement a method according to one of claims 1 to 26.

3₁. Non-volatile, computer-readable medium, storing an avatar generated by a method according to claim 16.