| Human shape modeling plays an important role in many applications ranging from digital entertainment and garment manufacturing, to simulation and training. However, creating human shape models remains a tedious and labor intensive task. There are difficulties in acquisition, merging, and editing of human shapes. So our paper mainly aims to address these problems.Since the current datasets all have their defects, we start by bringing a large data set of body scans into correspondence. We develop a non-rigid registration algorithm with regularization, to register the raw range scans with a template mesh.For human model reshaping, we extend the existing method with a novel regres-sion model, which we refer to as Local Mapping, to explore the space of detailed semantic attributes. For each triangular face, a linear mapping between semantic at-tribute parameters and the corresponding shape variations is learned, and a mapping constraint is introduced to avoid the over-fitting problem. Our method is more robust together with our established dataset, and can achieve local deformation accurately.To address the human shape estimation problem, we present a real-time system of which the key idea is to search a best match for the depth image from our dataset. To insure the accuracy, we divide the input depth map into several parts and search for the best matching shape per region. We begin with depth image smoothing by using a Weighted Least Squares (WLS) method. After that, a Gaussian Mixture Model method is used to establish correspondences between point cloud and human model. Invalid correspondences are eliminated and laplacian deformation is performed to provide a reasonable estimation of the eliminated points, which maintains correspondences be-tween them and at the same time preserves geometry structure of the human model.Once the deformed model is obtained, we search for a best match for each body part and merge them to get a complete human model. |
PDF Full Text Request