1. No category

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