Computer Science Colloquia
Friday, February 27, 2015
Vladimir Kim, Stanford University
10:45 AM, Rice Hall, Rm. 242
HOST: Connelly Barnes
Structure and Function in Large Collections of 3D Models
As large repositories of 3D shape collections grow, understanding the geometric data, especially encoding the inter-model similarity, their variations, semantics and functionality, is of central importance. My research addresses the challenge of deriving probabilistic models that capture common structure in large, unorganized, and diverse collections of 3D polygonal shapes. In my talk, I will present two such structural models and algorithms for inferring them, as well as their applications to exploration, organization, analysis, and synthesis of geometric data. First, I will describe part-based templates for encoding structural variations in collections of man-made shapes. In this work, we propose an automatic algorithm that starts with an initial template model and then jointly optimizes for part segmentation, point-to-point surface correspondence, and a compact deformation model for the templates to best explain the input shape collection. As output, the algorithm produces a set of probabilistic part-based templates that groups the original collection into clusters of objects capturing their styles and variations. The second structural model is motivated by the observation that a majority of man-made shapes are designed to be used by people. Thus, in order to fully understand their semantics, one needs to answer a fundamental question: “how do people interact with these objects?” This work demonstrates that variations in some man-made shapes can be encoded as variations in poses that a person would need to adopt in order to use an object. More specifically, we observe that when a person is interacting with an object, contact points usually share consistent local geometric features related to the anthropometric properties of corresponding parts, and that human body is subject to kinematic constraints and priors. We learn these priors from example data and represent them with local region classifiers and global kinematic constraints. Given an input 3D shape, we use our structural model to predict a corresponding human pose, including contact points and kinematic parameters. Finally, I will demonstrate that part-based and human-centric models enable effective user interfaces for exploring geometric datasets and synthesizing novel data.
Bio: Vladimir G. Kim is a postdoctoral scholar at Stanford University working in Geometric Computation group. Vladimir obtained his Ph.D. from Princeton University in 2013. His main research interest is geometric analysis with applications in computer graphics and 3D vision. Vladimir's recent work focuses on data-driven algorithms that leverage large geometric repositories to analyze functional, structural, and semantic attributes of objects and environments. He also works on interactive tools to facilitate content creation, exploration, and manipulation of geometric and other visual data. Vladimir has served/will serve on program committees for SIGGRAPH 2015, SGP 2013, SGP 2014, SGP 2015, Eurographics (short papers) 2014, Eurographics (short papers) 2015, and CAD/Graphics 2015. Vladimir is a recipient of the Siebel Scholarship and the Alexander Graham Bell Canada Graduate Scholarship.