AG 1, AG 2, AG 3, AG 4, AG 5, SWS, RG1, MMCI   Info

Public Audience

English

-- Not specified --

Saarbrücken

This thesis proposes several methods to statistically analyze static and
dynamic 3D face data. First, we present a fully-automatic method to
robustly register entire facial motion sequences. The representation of
the 3D facial motion sequences obtained by the registration allows us to
perform statistical analysis of 3D face shapes in motion. We then
introduce a new localized multilinear model that is able to capture
fine-scale details while being robust to noise and partial occlusions.
To obtain a suitable registration for multilinearly distributed data, we
introduce a groupwise correspondence optimization method that jointly
optimizes a multilinear model and the registration of the 3D scans used
for training. To robustly learn a multilinear model from 3D face
databases with missing data, corrupt data, wrong semantic
correspondence, and inaccurate vertex correspondence, we propose a
robust model learning framework that jointly learns a multilinear model
and fixes the data. Finally, we present one application of our
registration methods, namely to obtain a sizing system that incorporates
the shape of an identity along with its motion. We introduce a general
framework to generate a sizing system for dynamic 3D motion data

Ellen Fries

9325-4003

--email hidden

passcode not visible

logged in users only