Automatic Locating of Anthropometric Landmarks on 3D Human Models
description
Transcript of Automatic Locating of Anthropometric Landmarks on 3D Human Models
1
Automatic Locating of AnthropometricAutomatic Locating of Anthropometric LandmarksLandmarks on 3Don 3D Human ModelsHuman Models
Zouhour Ben AzouzZouhour Ben Azouz and Chang Shu and Chang Shu
National Research Council of CanadaNational Research Council of Canada
Institute of Information TechnologyInstitute of Information Technology
Visual Information Technology GroupVisual Information Technology Group
2
AnthropometryAnthropometry
Traditional Anthropometry Traditional Anthropometry 3D surface Anthropometry 3D surface Anthropometry
Poor description ofPoor description of the body shapethe body shape
Captures the details of the bodyCaptures the details of the bodyShape within a few secondsShape within a few seconds
3
3D Anthropometric Data3D Anthropometric Data
191 412 surface points191 412 surface points172 267 surface points 172 267 surface points
No No correspondencecorrespondence
4
Establishing Correspondence Establishing Correspondence between Human Scansbetween 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 Template fitting ((Allen, Curless, and Popovic, 2003)Allen, Curless, and Popovic, 2003)
fit a template human model to fit a template human model to instances of human scansinstances of human scans
5
Establishing Correspondence Establishing Correspondence between Human Scans (cont)between Human Scans (cont)
Scan instanceTemplate model
Template fitting Template fitting (Allen, Curless, and Popovic, 2003)(Allen, Curless, and Popovic, 2003)
Requires anthropometric landmarksRequires anthropometric landmarks
6
Anthropometric Anthropometric LandmarksLandmarks
7
Previous work on Previous work on Anthropometric Anthropometric Landmark Landmark locatinglocating
G.R. Geisen,C.P. Mason, G.R. Geisen,C.P. Mason, Automatic Detection, Identification, Automatic Detection, Identification, and Registration of Anatomical and Registration of Anatomical Landmarks,Landmarks, 1995. 1995.
D. Burnsides, M. Boehmer and D. Burnsides, M. Boehmer and K.M. Robinette, K.M. Robinette, 3-D Landmark 3-D Landmark Detection and Identification in the Detection and Identification in the CAESAR Project,CAESAR Project, 2001. 2001.
Landmark locating Landmark locating based on prior markingbased on prior marking
8
Previous work on Previous work on Anthropometric Anthropometric Landmarks Landmarks locatinglocating (cont) (cont)
A. Certain and W. Stueltzle, A. Certain and W. Stueltzle, Automatic Body Measurement for Automatic Body Measurement for Mass Customization of Garments,Mass Customization of Garments, 1999. 1999.
L. Dekker, I. Douros, B.F. Buxton L. Dekker, I. Douros, B.F. Buxton and P. Treleaven, and P. Treleaven, Building Symbolic Building Symbolic Information for 3D Human Body Information for 3D Human Body Modeling from Range Data,Modeling from Range Data, 1999. 1999.
Landmark locating Landmark locating without prior markingwithout prior marking
yy
zzVariation of z < v1
Variation of z > v2Place acromion
9
ObjectifsObjectifs
Locating automatically landmarks based Locating automatically landmarks based on learning techniques.on learning techniques.
Use a general framework to identify all Use a general framework to identify all the landmarks.the landmarks.
10
Landmark Locating problemLandmark Locating problem
Learning step:Learning step: Local surface properties of landmarks.Local surface properties of landmarks. The spatial relationship between landmarks.The spatial relationship between landmarks.
Matching step: for an instance of a human Matching step: for an instance of a human model model
assignassign to the to the anthropometricanthropometric landmarks the landmarks the position that is the most compatible with the position that is the most compatible with the learned learned informationinformation..
11
Pairwise Markov Random Pairwise Markov Random Field (MRF)Field (MRF)
)l( ii
)l,l( jiijll22
ll66
ll33
ll44 ll55
ll11
The joint probability represented The joint probability represented by a pairwise MRF:by a pairwise MRF:
i j,ijiii )l,l()l(
Z1
)L(P
Likelihood that landmark li correspond to a given
position on the surface
Compatibility between the Positions of landmark pairs
}l,...l,..,l,l{L ni21
il is the position of the landmark i
12
Landmark locating using a Landmark locating using a pairwise MRFpairwise MRF
Learning step: define the parameters of the Learning step: define the parameters of the probabilities associated to the MRF.probabilities associated to the MRF.
Probabilistic inference step: Assign to each Probabilistic inference step: Assign to each landmark the position that maximizes the joint landmark the position that maximizes the joint probability defined by the MRF.probability defined by the MRF.
13
Learning stepLearning step
Local surface properties
is a gaussian distribution of is a gaussian distribution of Spin Spin ImagesImages
)l( ii
[Johnson 97]
14
Learning stepLearning step
Structure of the landmark graph Structure of the landmark graph (73 landmarks)(73 landmarks)
Spatial relationship between landmarksSpatial relationship between landmarks
)l,l( jiij is a gaussian distribution of is a gaussian distribution of the the relative position of relative position of landmark landmark lljj with respect to landmark with respect to landmark llii..
15
Probabilistic InferenceProbabilistic Inference
j,i
jiiji
ii )l,l()l(Z1
)L(P
}j{)i(Nk
itki
ljiijiij
1tij )l(m)l,l()l()l(m
i
)i(Nj
ijiiiii )l(m)l(k)l(b
Maximizing the probability function:Maximizing the probability function:
through message passing:through message passing:
Attribute to each landmark the position that has the highest beliefAttribute to each landmark the position that has the highest belief
Loopy Belief PropagationLoopy Belief Propagation
16
Experimental ResultsExperimental ResultsTraining set of 200 human models from the CAESAR Training set of 200 human models from the CAESAR data basedata base
17
Experimental ResultsExperimental Results
18
Experimental ResultsExperimental Results
19
Experimental ResultsExperimental ResultsError of landmark locating computed over 30 test Error of landmark locating computed over 30 test human models.human models.
20
Conclusion and Future Conclusion and Future workwork
Our Approach locates a large number of Our Approach locates a large number of anthropometric landmarks without placing anthropometric landmarks without placing markers on all the measured individuals witch is markers on all the measured individuals witch is useful for future 3D anthropometric data useful for future 3D anthropometric data collection.collection.
The current results can be improved by:The current results can be improved by: Identifying automatically the most correlated pairs of Identifying automatically the most correlated pairs of
landmarks.landmarks. Developing surface descriptors that are posture and Developing surface descriptors that are posture and
resolution invariant .resolution invariant . Use of geodesic distance to characterize the spatial Use of geodesic distance to characterize the spatial
relationship between pairs of landmarks.relationship between pairs of landmarks.