Data driven models of human shape, pose and garment deformation

• Main Author: Neophytou, Alexandros
• Other Authors: Hilton, Adrian
• Published: University of Surrey 2015
• Subjects: 006.6
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.667618

This thesis addresses the problem of modeling human shape in three dimensions. Specifically, this thesis is focused on modeling body shape variation across multiple individuals, pose induced shape deformations and garment deformations that are influenced both by body shape and pose. A methodology for constructing data driven models of human body and garment deformation is provided. Additionally, an application for online fashion retailing is presented. Abstract Firstly, a quantitative and qualitative evaluation, is introduced, of surface representations used in recent statistical models of human shape and pose. It is shown that the Euclidean representation generates a more compact human shape model compared to other representations. A small number of model parameters indicates better convergence in a human body estimation framework. In contrast, a high number of model parameters increases the risk of the optimization getting trapped in a local optimum. Based on these insights a system for human body shape estimation and classification for on-line fashion applications is presented. Given a single image of a subject and the subject's height and weight the proposed framework is able to estimate the 3D human body shape using a learnt statistical model. Results demonstrate that a single image holds sufficient information for accurate shape classification. This technology has been exploited as part of a collaborative project with fashion designers to develop a mobile app to classify body shape for clothing recommendation in online fashion retail. Abstract Next, Shape and Pose Space Deformation (SPSD) is presented, a technique for modeling subject specific pose induced deformations. By exploiting examples of different people in multiple poses, plausible animations of novel subjects can be synthesized by interpolating and extrapolating in a joint shape and pose parameter space. The results show that greater detail is achieved by incorporating subject specific pose deformations as opposed to a subject independent pose model. Finally, SPSD is extended to a three layered data-driven model of human shape, pose and garment deformation. Each layer represents the deformation of a template mesh and can be controlled independently and intuitively. The garment deformation layer is trained on sequences of dressed actors and relies on a novel technique for human shape and posture estimation under clothing.