1 CADANS seminar 2015 HUMAN BODY MODELING Femke Danckaers, Toon Huysmans, Jan Sijbers iMinds Vision Lab, Dept. of Physics, University of Antwerp, Belgium 02/06/2015 Overview 2 Introduction Correspondence preserving elastic surface registration Building a shape model Shape prediction from features Conclusions and future work 3 / 55 Introduction Introduction 4 Capturing shape variability Predict a shape based on a simple set of features Applications 5 Early validation of near-body product design, such as chairs, backpacks, clothing,... Virtual validation for wide range of body shapes by linking CAD design on mannequin Improved fit leads to increased comfort Development of sizing systems for retail For e.g. furniture Based on body type Correspondence Preserving Elastic Surface Registration Surface Registration Framework 7 Source mesh Target mesh Decrease β Increase α Corresponding points + similarity criteria Affine transformation source F Convergence? Model regularized deformation Elasticity regularized deformation (β) Non-rigid deformation (α) T Result F. Danckaers, T. Huysmans, D. Lacko, A. Ledda, S. Verwulgen, S. Van Dongen,and J. Sijbers: Correspondence Preserving Elastic Surface Registration with ShapeModel Prior. ICPR 2014. Find Corresponding Points 8 corresponding points affine alignment elastic deformation source target Find Corresponding Points 9 Cast a ray in the direction of the normal from source points to target surface corresponding points Constraints Maximum angle between source and target normal maximum distance from source no intersections allowed between source and target Good correspondence affine alignment Bad correspondence source target elastic deformation Affine Alignment 10 Surface alignment based on corresponding points affine transformation least squares solution corresponding points affine alignment elastic deformation source aligned source target Elastic Deformation 11 displace each vertex separately translation vector per vertex stiffness neighboring vertices forced to move along decreases as iterations progress corresponding points affine alignment Target surface Deformed source surface Source surface * B. Amberg, S. Romdhani, T. Vetter. Optimal Step Nonrigid ICP Algorithms for Surface Registration. In Proc. Conf. Computer Vision and Pattern Recognition, 2007. elastic deformation Elastic Deformation 12 displace each vertex separately translation vector per vertex corresponding points stiffness neighboring vertices forced to move along decreases as iterations progress affine alignment elastic deformation source deformed source target * B. Amberg, S. Romdhani, T. Vetter. Optimal Step Nonrigid ICP Algorithms for Surface Registration. In Proc. Conf. Computer Vision and Pattern Recognition, 2007. Surface Registration Framework 13 affine alignment r + elastic deformation e surface at iteration i: 𝒔𝑖 = (𝛼 − 1)𝐫 𝒔𝑖−1 + 𝛼𝐞(𝒔𝑖−1 , 𝛽) stiffness factor 𝛽 to control elasticity influence of elastic deformation 𝛼 no surface parameterization involved corresponding points affine alignment applicable to complex topologies not sensitive to topological noise elastic deformation Correspondences 14 Average surface Registered population Building a Shape Model Data 16 CAESAR database Standing and seated pose Antropometrical measures + meta-data Markers (annotated + 3D coordinates) In our current model: 147 men and 265 women Population Correspondence 17 target corresponded aligned points model source ... ... ... ... Population Correspondence 18 statistical shape model mean surface 𝑿 modes of variation Instance 𝒀 = 𝑿 + 𝚽𝒃 Eigenvector matrix: 𝚽 Parameters of instance: 𝒃 model parameters size and shape position normalization via Procrustes alignment iterative affine alignment to population mean Shape Modes all subjects 19 Principal Component Analysis -3s Mode 1 Mode 2 mean +3s -3s mean +3s Shape Modes all subjects 20 -3s Mode 3 Mode 4 mean +3s -3s mean +3s Model Performance compactness 21 The compactness of the model measured by the cummulative variance Detail of 20 first modes 1 0,9 Cumulative variance Cumulative variance 1 0,8 0,7 0,6 0,5 0,4 0,3 0,2 0,1 0,95 0,9 0,85 0,8 0,75 0,7 0,65 0,6 0 1 21 41 61 81 101 121 141 161 181 201 221 241 261 281 301 Number of shape modes 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Number of shape modes Other Shape Models 22 Shape Prediction from Features Input feature parameters 24 height Gender Height Weight Chest circ. Age Waist circ. Waist circumference Hip circ. Thigh circ. Chest circumference Hip circumference Thigh circumference Other interesting features? Predict Shape from Features 25 Build mapping matrix 𝑴 = 𝑩𝑭+ , with 𝑩 the PC weights of each surface and 𝑭 the feature matrix of each surface from the population 𝑴 Calculate new weights 𝒃 = 𝑴𝒇, with 𝒃 the individual PC weights and 𝒇 the desired feature vector Calculate new surface 𝒙′ 𝒙 ′ = 𝒙 + 𝑷 ⋅ 𝒃 , with 𝒙 the mean surface and 𝑷 the PC vectors Allen, B., Curless, B., Popovic, Z., The Space of Human Body Shapes: Reconstruction and Parameterization from Range Scans. ACM SIGGRAPH, 2003, pp,1-8 Predicted shape from features 26 Distance original - predicted (mm) Predicted shape from features 27 Distance original - predicted (mm) BMI Variation 28 Male Female BMI 15 BMI 20 BMI 25 BMI 30 BMI 35 Gender Variation 29 female average male Conclusion and Next Steps Conclusions 31 Surface registration leads to accurate correspondences Statistical shape model can be used to describe the population New shapes can be generated from a given set of features Next Steps 32 More complex modeling of feature modification E.g. Via subpopulations or non-linear models Shape clustering Distinguish different body types Pose normalization More accurate shape clusters More accuracy on feature influence Deducing asymmetry original

© Copyright 2021