S. Saito, Building 3D Deformable Body Model and Torso Shape Estimation System *Corresponding author. Email: ssaito@aoki-medialab.org 1 Building 3D Deformable Body Model and Torso Shape Estimation System S. SHUNTA*†, M. KOUCHI‡, M. MOCHIMARU‡ and Y. AOKI† † Keio University, Graduate School of Integrated Design Engineering, Yokohama, Kanagawa, Japan ‡ Digital Human Research Center, National Institute of Advanced Industrial Science and Technology, Koto-ku, Tokyo, Japan Abstract Human body shape represented as a 3D mesh model can be achieved by various methods, for example, laser scan, projector camera system and volume intersection, etc. However, a more easier method is required in apparel markets. Because a virtual fitting technique or virtual cloth evaluation system is increasingly demanded on the back of online shopping and those techniques commonly requires the 3D body shape of users. Therefore, more simple and accessible method to obtain the 3D shape is necessary. Then, we focus the estimation of upper body shape in 3D using ordinary cues, silhouette images. In this paper, to obtain 3D body trunk shape approximately but with sufficient degree of accuracy to represent individual differences, we propose a body model. The model has deformable body type and pose and is synthesized from a human shape database provided by DHRC. We use the model to estimate the 3D torso shape by optimizing the model shape and pose parameters. It is achieved by comparing input silhouettes and model silhouettes. The resulting estimation errors are evaluated by measuring the average errors in the nearest vertex distance between model and answer shape. Keywords: 3D Body Shape Estimation, Human Body Model, Silhouette 1. Introduction Human body shape represented as a 3D mesh model can be achieved by various methods, for example, laser or range scan, projector camera system and volume intersection, etc. However, a easier method is required in apparel markets. Because a virtual fitting technique or virtual cloth evaluation system is increasingly demanded on the back of online shopping and those techniques commonly requires the 3D body shape of users. Therefore, more simple and accessible method to obtain the 3D shape is necessary. Some methods using range scan (M. A. Brunsman et al. 1997) or projectors and cameras (L. Zhang, et al. 2004; J. Siebert and S. Marshall 2000) can reconstruct 3D shape with high accuracy. However, they cannot be said easy for personal use. Because those works require some dedicated hardware or specialistic preparation, for example, projector and camera calibration. On the other hand, some researches use previous knowledge of target objects for 3D shape estimation. One of the model based body shape estimation system is proposed by Suzuki et al. The system is for human face (S. Suzuki and H. Saito 2009). It uses 4 freehand captured images as input, and fit a model based on Active Shape Model (ASM) (T. F. Cootes et al. 1995) to inputs. ASM is a parameterized model method for constructing shape deformable model with principal component analysis (PCA). To use ASM for 3D shape estimation, many 3D mesh data is required and each data should have same topology and same number of points. ASM is appropriate to deal with the shapes without a topological difference among individual shape like human body part. However, it is difficult to represent pose-varied objects. In the 3D face reconstruction system, it is assumed that face has no intrinsic pose variation and only has shape variation and difference of configuration. In the case of face, this is a rational hypothesis because it is easy to align facial bone when the mouth is closed. By contrast, in the case of torso shape, the pose variation is a critical factor for surface shape when a user captured by camera. Therefore, to reconstruct torso shape, we must construct a model which consider not only shape variation but also postural changes. In this paper, we address the problem by combining ASM and pose changing method (J. P. Lewis, et al. 2000) with joint implanted shape data. And we apply the model to estimation problem of human upper body shape estimation from only 2 images, front and side silhouettes. The input 2 silhouettes are preprocessed