Lei Wang Joint work with Tamal K Dey Huamin Wang and Bo Fu Problem Statement Lowend scanning devices are becoming popular But quality of their output Reconstruct from problematic human scans by deforming a prior high quality template mesh ID: 674826
Download Presentation The PPT/PDF document "Automatic Posing of a Meshed Human Model..." is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
Automatic Posing of a Meshed Human Model Using Point Clouds
Lei Wang
Joint work with Tamal K. Dey,
Huamin Wang and Bo FuSlide2
Problem Statement
Low-end scanning devices are becoming popular. But quality of their output…
Reconstruct from problematic human scans by deforming a prior high quality template meshSlide3
Related Work
Model-based Registration
Parametric human body models
Registration without a Model
Manual intervention
Non-rigid registration by nonlinear optimization
Isometric deformations: keep geodesic distance unchangedSlide4
Our Method: Outline
computing correspondences
posing
outputSlide5
Computing CorrespondencesSlide6
Computing Correspondences
The Global Point Signature (GPS) framework by [Rustamov 2007]:
Apply Gaussian-weighted Graph Laplacian on the adjacency graph of input point cloud
Graph Laplacian
GPS in dimension 1, 2 and 3Slide7
Computing Correspondences
GPS is invariant under isometric deformation
We choose only the first three eigenfunctionsSlide8
Computing Correspondences
Correspondences are expanded from 5 extremums of aligned GPS embeddingsSlide9
Posing Template Mesh
Formulated as an energy minimization problem:
is the internal energy of template mesh
measures difference between template and the point cloud
Solved by
I
nvertible FEM [Irving et al. 2004]Slide10
It has two stages
Point cloud alignment: deformed by
input point cloud
Posing Template Mesh
Initial alignment:
deformed by
correspondenceSlide11
Posing Template Mesh
For efficiency, posing is performed on a simplified template mesh, called
control mesh
Deformed template is recovered by
M
ean
V
alue
C
oordinates [Ju et al.
2005], i.e., weighted sum of the control meshSlide12
ResultsSlide13
ResultsSlide14
ResultsSlide15
Results
(a) input scan (b) our method (c) SCAPESlide16
Results
can be used to control the body sizeSlide17
Supplementary VideoSlide18
Discussion
An automatic approach to align a detailed template mesh with human point clouds in different poses
Robust to noise and occlusions
Sensitive to topological change
Can not handle details like fingers
Does not run in real-timeSlide19
Q & A
Thank you!