The Development of 3D Statistical Shape Models for Diverse Applications

Date: 18 February 2019

Venue: Campus Groenenborger, U0.25 - Groenenborgerlaan 171 - 2020 Antwerpen (route: UAntwerpen, Campus Groenenborger)

Time: 4:30 PM

Organization / co-organization: Faculty of Science

PhD candidate: Femke Danckaers

Principal investigator: Jan Sijbers & Toon Huysmans

Short description: PhD defence Femke Danckaers - Faculty of Science


The human body appears in many shapes and sizes. For product developers, it is useful to have a virtual 3D mannequin available to generate and validate their designs. Such anthropometric tools are widely available, but often provide only a simplified representation of the body, based on 1D measurements. Modification of the shape is often done in an univariate way, so 3D shape variation is not incorporated in such models. A statistical shape model (SSM) can be used as digital mannequin, because it describes the main variations of shape inside the population.

To perform shape analysis, the shapes within the population are brought into correspondence with each other by applying elastic surface registration. From a population of corresponded shapes, an SSM can be built. This model consists of the average 3D shape of the object class and the main shape variations that occur inside the shape population. The shape of an SSM can be changed by adapting the shape parameters. Those parameters are typically not linked with specific shape characteristics. Therefore, body shape modeling based on intuitive parameters is discussed in this work. Shape variation captured by an SSM is often polluted by variations in posture, which may incorrectly correlate with features and negatively affect the compactness of those models. Therefore, a framework that has low computational complexity to build a posture-invariant SSM, by capturing and correcting the posture of an instance, is shown. SSMs are typically a static representation of a population. A movement acquired by a motion capturing system is integrated in the SSM, allowing to modify its pose in a realistic way and to add pre-recorded motion to different body shapes in a realistic way.

This PhD thesis presents methods for an improved shape analysis. The methodology is not restricted to body shapes and is applicable to almost every object class that contains natural variance.