Presentation is loading. Please wait.

Presentation is loading. Please wait.

Automatic Locating of Anthropometric Landmarks on 3D Human Models

Similar presentations


Presentation on theme: "Automatic Locating of Anthropometric Landmarks on 3D Human Models"— Presentation transcript:

1 Automatic Locating of Anthropometric Landmarks on 3D Human Models
Good afternoon everybody, in this presentation I will talk about In this talk I’ll present an approach we developed to automatiquely……….. Zouhour Ben Azouz and Chang Shu National Research Council of Canada Institute of Information Technology Visual Information Technology Group

2 Anthropometry Traditional Anthropometry 3D surface Anthropometry
As was mentioned in the previous presentations, During the last decade the field of anthropometry has known a remarkable change due to the use of whole body scanners to measure the body shape rather than using simple tools such as calipers and measuring tapes to generate 1D measurements. The body scanners has the advantage of capturing the details of the body shape in a few seconds Poor description of the body shape Captures the details of the body Shape within a few seconds

3 3D Anthropometric Data No correspondence 172 267 surface points
However Processing the 3D anthropometric data in order to make it useful for the design of several products Is still a challenging problem. The main difficulty comes from the fact that if we scan two differenrent subjects their 3D models will have different number of surface point and there is no anatomical correspondance between the surface points of these models. Without correspondance we can not compare 3D models between each other and hense we can not analyse the shape variability within a given population. The correspondance between different models is required to compare modeels between each other in order in order to analyse the shape varaibility. The main difficulty is that the human models have diffetent number of surface points and there’s no anatomical correspondance beween them. Without correspondance we can not compare compare human scans between them and study the variability of their shape. Establishing a correspondance between the human scans is crucial for characterizing the human shape and analysing its variability within a given population. =========================================================================================== Processing the 3D anthropometric data in order to analyse the body shape variablility is a challenging problem. The main difficulty is that the human models have diffetent number of points and there’s no anatomical correspondance beween them Thus In order to make human scans useful for the design of products, medical applications and generating realistic human characters it is important to establish a correspondance between different human scans The use of body scnners in the field of anthropoemtry has generated a new research field that aims to use the human scans for characterizing the human shape and comparing the shape of different individuals. The main difficulty is to The use of whole body scanners raised a new research area that aims to answer the question how to characterize the human body shape using 3D…………. The main difficulty is that body scans have different number of surface points and there’s no correspondance between these points. Establishing a correspondance between different models is a key issue to compare different indivisuals of a given population. To make use of 3d /…. How to charcterize the human body shape and compare the shpae of different individuals. Les systèmes de numérisation génèrent des modèles 3D de corps humain qui sont constitués de centaines de milliers de points de surface. Le nombre de points varie d’un individu à un autre et il n’y a pas de correspondance anatomique entre les points des différents modèles. La question qui se pose est Comment traiter les données 3D pour en extraire une quantification de la forme. La question qui se pose est Comment traiter les données 3D pour les rendre utile dans la phase de conception. La question qui se pose est comment traiter ces données pour en extraire une quantification de la forme. Pour cela ces données doivent être traité afin d’en extraire une quantification de la forme La question qui se pose est comment ces données Il est alors essentiel La question qui se pose est comment quantifier et comparer les formes des … Pour cela ces données doivent être traité pour en extraire de l’information utile pour la phase de conceptions de divers produits afin de quantifier la forme est d’analyser ’en extraire de l’information sur la forme du corps humain qui aide à concevoir des produits qui conviennent à un grand pourcentage de a population ciblée. Pour cela afin de rendre ces données utile il est nécessaire de les traiter afin de les convertir en une représentation qui aide le concepteur à concevoir des produits qui conviennent à un grand pourcentage de a population ciblée. Ces données ne peuvent être utilisé directement par un concepteur c’est pour cela qu’il est nécessaire de les traiter pour en extraire une quantification de la forme qui permet La question qui se pose est comment quantifier surface points surface points

4 Establishing Correspondence between Human Scans
Canonical sampling (Tahan, Buxton, Ruiz, 2005) regularly sample each model requires segmentation Volumetric method (Ben Azouz, Shu, Lepage, Rioux, 2005) volumetric representation signed distance landmark free Template fitting (Allen, Curless, and Popovic, 2003) fit a template human model to instances of human scans In the litterature there’s different approaches to solve the correspondance problem between human scans. The most efficient one is based on fitting a template models to instances on human scans The deformation is guided by the positions of a set of anthropometric landmarks. After fitting the human models have the same number of points and there’s an anatomical correspondance between the points of different models the different models Locating Anthropometric landmarks is useful for establishing correspondonce between human scans for processing 3D anthropometric Data. In the litterature there’s 3 different approaches to solve the correspondance problem. The first one is based on regularly sampling the human scans. The limitations of this approach is that it applies to convex objects and thus require the segmentation of the body wich is a challenging task. The second approach was developped in our previous work. It consists in representing It compares differents models using a volumetric representation that is based on signed distance. Even tough this approach allows the extraction of the principal modes of shape variation in a given population, it has the limitation that it provides only an approximation of the anatomical correspondence between different models. The most efficient way to establish a correspondance between body scans is based on fitting a template model to instances of human scans. After fitting the human models have the same number of points and there’s an anatomical correspondance between the different models.

5 Establishing Correspondence between Human Scans (cont)
Template fitting (Allen, Curless, and Popovic, 2003) Scan instance Template model It basically consist on deforming a template model in order to fit to a scan instance. The deformation is guided by a set of antgropomrtric landmarkd so that after deformation The landmarks of the template model fit those of the scan instance. By applying this procedure to different human scans we obtain human models that have the same number of surface point and there’s correspondance between the points of these models. This procedure of establishing a correspondance between human models requires the identification of anthropometris landmarks. It consist in using a template human model and deform in order to fit it to instances of human scans. The deformation is guided by a set of antgropomrtric landmarkd so that after deformation The landmarks of the template model fit those of the instance scan. By applying this procedure to different human scans we obtain human models that have the same number of surface point and there’s correspondance between the points of different point. In the litterature there’s different approaches to solve the correspondance problem between human scans. The most efficient one is based on fitting a template models to instances on human scans The deformation is guided by the positions of a set of anthropometric landmarks. After fitting the human models have the same number of points and there’s an anatomical correspondance between the points of different models the different models Locating Anthropometric landmarks is useful for establishing correspondonce between human scans for processing 3D anthropometric Data. In the litterature there’s 3 different approaches to solve the correspondance problem. The first one is based on regularly sampling the human scans. The limitations of this approach is that it applies to convex objects and thus require the segmentation of the body wich is a challenging task. The second approach was developped in our previous work. It consists in representing It compares differents models using a volumetric representation that is based on signed distance. Even tough this approach allows the extraction of the principal modes of shape variation in a given population, it has the limitation that it provides only an approximation of the anatomical correspondence between different models. The most efficient way to establish a correspondance between body scans is based on fitting a template model to instances of human scans. After fitting the human models have the same number of points and there’s an anatomical correspondance between the different models. Requires anthropometric landmarks

6 Anthropometric Landmarks
Anthropometric landmarks are points on the surface that correspond usually to skeletal features. The definition of these landmarks is well established in traditional anthropometry and they are used to define trditional meaurements. Locating these points is important for processing 3D anthropometric data. The key point of our approach is that the positions of landmarks are correlated and hence we can think of the landmark locating We want a global localisation. To identify landmarks we can learn their surface properties and learn their spatial relationship. Knowing this information we can locate landmarks by identifying the vertices that correspond the best to these two information. An adequate mathemathecal framework to formulate this problem is the markov network.

7 Previous work on Anthropometric Landmark locating
Landmark locating based on prior marking G.R. Geisen,C.P. Mason, Automatic Detection, Identification, and Registration of Anatomical Landmarks, 1995. D. Burnsides, M. Boehmer and K.M. Robinette, 3-D Landmark Detection and Identification in the CAESAR Project, 2001. Prevous works on landmarks locating can be classified into two main categories The first one witch the most common is based on placing markers on the human body before scanning. The lomitation of this categorie is that placing the markers is a time consumin pocess that increases the cost of 3D anthropometric surveys. Eliminating this procedure will simplify considerably 3D anthropometric surveys and will help collecting 3D anthropometric data more frequently. and will mostlikely be eliminated from future Project of 3D anthropometric Data collection.

8 Previous work on Anthropometric Landmarks locating (cont)
Landmark locating without prior marking A. Certain and W. Stueltzle, Automatic Body Measurement for Mass Customization of Garments, 1999.  L. Dekker, I. Douros, B.F. Buxton and P. Treleaven, Building Symbolic Information for 3D Human Body Modeling from Range Data, 1999. y z Variation of z < v1 Variation of z > v2 Place acromion The second categorie attempts to locate automatiquely landmark without prior marking. Most of the proposed methods are limited to locating branching points such the armpits. Few methods extended the identification for a larger subset of landmarks. They rely on building specific functions to locate each landmark. For instance Dekker proposes to identify the acromion wich is the outer end of the shoulder using the following function: Variation of y < v1 The acrominion is difined as the first point in this line trversing down from the nape where there is a gradient less than a value g1 and than a value grater than a threshold g2. Defining such detectors is subjective and not straitforward for most of the anthropometric landmarks It’s the first point on the torso traversing down from the nape on the ridge line of maxim

9 Objectifs Locating automatically landmarks based on learning techniques. Use a general framework to identify all the landmarks. The objectif of our work is aligned with the motivation of the second categorie in the sense that we aim to locate the anthropometric landmarks without increasing the time of the human body measurement. We propose to place markers on a reduced number of individuals rather than placing markers on all the indivisuals measured during 3D anthropometric surveys This training set will be used to learn the propoerties of these landmarks For the rest of mesured individuals the landmarks will be located ased on the learned information We propose to locate all the landmarks based on the same framework and not build specific functions for each landmark. The method we propose uses the fact that there’s two types of information that can be learned from identified landmarks The first one is and the second one is the interdependancu between the position of the landmarks. In fact we would like to place markers on a few number of indivisuals and learn La solution souhaitable pour concevoir des produits qui conviennent à un grand pourcentage d’une population ciblée est un système qui pour une application donnée permet d’extraire à partir d’une base de données un ensemble de modèles qui sont représentatifs de la population. Nous présentation ici notre vision de la stucture d’un tel système. Afin d’aider le concepteur dans sa tàche nos proposons de convertir les données anthropmétrique 3D en une représdentation qui soit compacte tout en permettatnt la reconstruction de la forme originale ce qui prouve qu’elle preserve l’information sur la form. Cette représentation sera untilie poour le concepteur de deux facons. D’une part elle sera la base de la selection de modèles représentatifs de la population étudiéa D’autres part elle permettera la visualisatiob des principaux modes de variations de la forme du corps humain. Cette visualisation est une information tangible au concepteur pour comprendre comment la forme varie au sein de la population donéée. Le c D’abord il nous paraît plus utile d’inclure le concepteur dans la boucle plutot que de chercher à automatiser son expertise. Une cdcomposante essentielle du sytème est un module qui permet de convertinr le grand nombre de points numérisés en une description qui soit compacte Tout en preservant l’information sur la forme su corps humain. Cette description servira d’une part à la selection des modèles représentatifs de la population étudiée. D’autre part elle permet d’extraire et de visualiser les principaux modes de variation de la forme du corps humain. La visualisation est une information tangible qui permet au concepteur qu’es ce qui varient le plus dans la population étudiée et d’identifier les variables de la description qui influence plus la conception du produit en question. L’expertise du concepteur lui permet de combiner la visualisation ainsi que les caractéristiques du produit concue et les contraintes budgétaires pour déterminer le nombre de mod`les représentatifs a extraire ainsi que les variables de la description a utiliser pour le faire. Le processus globlae est validé par un test de satisfaction des clients. Puisque ce travail n’est pas lié a une application particulière on se concentre sur le développement des deux premiers modules. Il faut fournir au concepteur une représentation sur laquelle il peut se baser pour comparer les différents individus Le système commence par convertir les données anthropométriques en une représentation compacte. L’utilité de cette représentation et de réduire les données tout en préservant l’Information sur la forme. La reprsentation sera utilisé pour l’extraction et la visualisation des principaux modes de variation de la forme du corps humain. Pour le concepteur La visualisation est une information tangible sur la forme. L’expertise du concepteur lui permet de combiner cette information avec les caractéristiques du produits a concevoir pour choisir le nombre de modèles representatif. Ce dernier peut être déterminé par regroupement des individus en se basant sur la description compacte. Dans ce travail de recherche on se concentre sur le développement des modules de description compacte et À regrouper les individus d’une population afin d’en extraire des indivi Une composante essentielle de ce système est un module qui permet d’extraire Ce système inclut le concepteur plutôt que de chercher à automatiser son expertise. Afin d’orienter nos recherches on propose d’abord de présenter notre vision de la structure globale d’un tel système. Il est difficile d’imaginer dans le futur prochain un système qui est complètement automatique. Il est difficile d’imaginer un système qui pourra remplacer l’expertise des concepteurs. Il est plutot intéressant de chercher des solutions qui font appel a l’expertise des concepteurs. Nous proposons la structure suivante Perhaps the most widely believed design myth is the idea that people can be grouped into sizes based upon their anthropometry, regardless of the design and performance requirements of a system. For example, it is a common misbelief that people naturally fall into three sizes: small, medium and large. An engineer once complained about spending money on sizing because he believed "everyone knows what a size medium is." A Navy scientist, Sirvart Mellian, encountering this same response many times, developed a quick way to refute this statement when speaking to a group. She asks everyone in the room who believes they are a size medium to stand up. A wide variety of people stand, demonstrating that all size mediums are not alike. The point is that anthropometry is not the only factor to consider when developing a sizing system. Design features and performance requirements of a system affect the amount of body size variability the system can accommodate. For example, a helmet used for a virtual reality (VR) game will not have to fit as tightly as an aircrew helmet that has to stay in place at 7 Gs. The VR helmet could probably fit more people in a single size as a result. The fit requirements are more loosely defined. Also, the VR helmet may need less adjustability than the aircrew helmet because precise placement of the display is less critical. Quality of fit is the degree to which the system can accommodate any individual in a population. The quality of fit for a population can be maximized in four ways: 1) good proportioning and shaping of a single size, 2) design features which broaden the accommodation range in a single size, such as adjustable straps or liners, 3) adding sizes, and 4) adding the ability to completely custom-make the system for individuals. Adding sizes can affect cost, adding adjustable design features can affect cost, and custom making equipment can affect cost. Therefore, methods for achieving a good quality fit for a population are both performance and cost factors. Anthropometry and sizing are design trade-offs with other performance criteria, and it is usually cheaper to include them in the development process to optimize the design. This can perhaps be best understood by reviewing past sizing practices.

10 Landmark Locating problem
Learning step: Local surface properties of landmarks. The spatial relationship between landmarks. Matching step: for an instance of a human model assign to the anthropometric landmarks the position that is the most compatible with the learned information. The objectif of our work is aligned with the motivation of the second categorie in the sense that we aim to locate the anthropometric landmarks without increasing the time of the human body measurement. We propose to place markers on a reduced number of individuals rather than placing markers on all the individuals measured during 3D anthropometric surveys This training set will be used to learn the propoerties of these landmarks For the rest of mesured individuals the landmarks will be located ased on the learned information. We propose to The method we propose uses the fact that there’s two types of information that can be learned from identified landmarks The first one is and the second one is the interdependancu between the position of the landmarks. La solution souhaitable pour concevoir des produits qui conviennent à un grand pourcentage d’une population ciblée est un système qui pour une application donnée permet d’extraire à partir d’une base de données un ensemble de modèles qui sont représentatifs de la population. Nous présentation ici notre vision de la stucture d’un tel système. Afin d’aider le concepteur dans sa tàche nos proposons de convertir les données anthropmétrique 3D en une représdentation qui soit compacte tout en permettatnt la reconstruction de la forme originale ce qui prouve qu’elle preserve l’information sur la form. Cette représentation sera untilie poour le concepteur de deux facons. D’une part elle sera la base de la selection de modèles représentatifs de la population étudiéa D’autres part elle permettera la visualisatiob des principaux modes de variations de la forme du corps humain. Cette visualisation est une information tangible au concepteur pour comprendre comment la forme varie au sein de la population donéée. Le c D’abord il nous paraît plus utile d’inclure le concepteur dans la boucle plutot que de chercher à automatiser son expertise. Une cdcomposante essentielle du sytème est un module qui permet de convertinr le grand nombre de points numérisés en une description qui soit compacte Tout en preservant l’information sur la forme su corps humain. Cette description servira d’une part à la selection des modèles représentatifs de la population étudiée. D’autre part elle permet d’extraire et de visualiser les principaux modes de variation de la forme du corps humain. La visualisation est une information tangible qui permet au concepteur qu’es ce qui varient le plus dans la population étudiée et d’identifier les variables de la description qui influence plus la conception du produit en question. L’expertise du concepteur lui permet de combiner la visualisation ainsi que les caractéristiques du produit concue et les contraintes budgétaires pour déterminer le nombre de mod`les représentatifs a extraire ainsi que les variables de la description a utiliser pour le faire. Le processus globlae est validé par un test de satisfaction des clients. Puisque ce travail n’est pas lié a une application particulière on se concentre sur le développement des deux premiers modules. Il faut fournir au concepteur une représentation sur laquelle il peut se baser pour comparer les différents individus Le système commence par convertir les données anthropométriques en une représentation compacte. L’utilité de cette représentation et de réduire les données tout en préservant l’Information sur la forme. La reprsentation sera utilisé pour l’extraction et la visualisation des principaux modes de variation de la forme du corps humain. Pour le concepteur La visualisation est une information tangible sur la forme. L’expertise du concepteur lui permet de combiner cette information avec les caractéristiques du produits a concevoir pour choisir le nombre de modèles representatif. Ce dernier peut être déterminé par regroupement des individus en se basant sur la description compacte. Dans ce travail de recherche on se concentre sur le développement des modules de description compacte et À regrouper les individus d’une population afin d’en extraire des indivi Une composante essentielle de ce système est un module qui permet d’extraire Ce système inclut le concepteur plutôt que de chercher à automatiser son expertise. Afin d’orienter nos recherches on propose d’abord de présenter notre vision de la structure globale d’un tel système. Il est difficile d’imaginer dans le futur prochain un système qui est complètement automatique. Il est difficile d’imaginer un système qui pourra remplacer l’expertise des concepteurs. Il est plutot intéressant de chercher des solutions qui font appel a l’expertise des concepteurs. Nous proposons la structure suivante Perhaps the most widely believed design myth is the idea that people can be grouped into sizes based upon their anthropometry, regardless of the design and performance requirements of a system. For example, it is a common misbelief that people naturally fall into three sizes: small, medium and large. An engineer once complained about spending money on sizing because he believed "everyone knows what a size medium is." A Navy scientist, Sirvart Mellian, encountering this same response many times, developed a quick way to refute this statement when speaking to a group. She asks everyone in the room who believes they are a size medium to stand up. A wide variety of people stand, demonstrating that all size mediums are not alike. The point is that anthropometry is not the only factor to consider when developing a sizing system. Design features and performance requirements of a system affect the amount of body size variability the system can accommodate. For example, a helmet used for a virtual reality (VR) game will not have to fit as tightly as an aircrew helmet that has to stay in place at 7 Gs. The VR helmet could probably fit more people in a single size as a result. The fit requirements are more loosely defined. Also, the VR helmet may need less adjustability than the aircrew helmet because precise placement of the display is less critical. Quality of fit is the degree to which the system can accommodate any individual in a population. The quality of fit for a population can be maximized in four ways: 1) good proportioning and shaping of a single size, 2) design features which broaden the accommodation range in a single size, such as adjustable straps or liners, 3) adding sizes, and 4) adding the ability to completely custom-make the system for individuals. Adding sizes can affect cost, adding adjustable design features can affect cost, and custom making equipment can affect cost. Therefore, methods for achieving a good quality fit for a population are both performance and cost factors. Anthropometry and sizing are design trade-offs with other performance criteria, and it is usually cheaper to include them in the development process to optimize the design. This can perhaps be best understood by reviewing past sizing practices.

11 Pairwise Markov Random Field (MRF)
Likelihood that landmark li correspond to a given position on the surface l2 l6 l3 l4 l5 l1 is the position of the landmark i A good mathmethecal framework to formulate the identification of landmarks is Pairwise Markov random Field. For eacvh node we attribute a potential representing the likeliFro hood that a landmark be in a given point on the body surface. For each edge we associate a potential that constraint the landmarks to be assigned to positions that are consistent with The probability that the positions attributed to a pair of landmark is consistent with the spatial relationship between these landmarks. Compatibility between the Positions of landmark pairs The joint probability represented by a pairwise MRF:

12 Landmark locating using a pairwise MRF
Learning step: define the parameters of the probabilities associated to the MRF. Probabilistic inference step: Assign to each landmark the position that maximizes the joint probability defined by the MRF.

13 Local surface properties
Learning step Local surface properties is a gaussian distribution of Spin Images [Johnson 97]

14 Spatial relationship between landmarks
Learning step Spatial relationship between landmarks is a gaussian distribution of the relative position of landmark lj with respect to landmark li. Structure of the landmark graph (73 landmarks)

15 Loopy Belief Propagation
Probabilistic Inference Loopy Belief Propagation Maximizing the probability function: through message passing: Attribute to each landmark the position that has the highest belief

16 Experimental Results Training set of 200 human models from the CAESAR data base

17 Experimental Results

18 Experimental Results

19 Experimental Results Error of landmark locating computed over 30 test human models.

20 Conclusion and Future work
Our Approach locates a large number of anthropometric landmarks without placing markers on all the measured individuals witch is useful for future 3D anthropometric data collection. The current results can be improved by: Identifying automatically the most correlated pairs of landmarks. Developing surface descriptors that are posture and resolution invariant . Use of geodesic distance to characterize the spatial relationship between pairs of landmarks.


Download ppt "Automatic Locating of Anthropometric Landmarks on 3D Human Models"

Similar presentations


Ads by Google