Seminar: High-Level Representations for Shape Understanding

Dr. Oliver van Kaick (Tel Aviv University)

Monday Feb 10, 2014, 10:30AM-12PM, EV 3.309

Abstract

During the last decade, the focus of research in computer graphics has shifted from generating images (rendering) to modeling and creation of 3D content. Modeling is a laborious task where artists need to be highly skilled to use the existing modeling tools, which typically involve working on low-level shape representations, e.g., triangle meshes. A recent trend in computer graphics research is the development of techniques that facilitate the creation of 3D models by manipulating shapes at a higher-level, relieving the users from considerable manual work. In this framework, shapes are represented as a collection of primitives defined at a more semantic level, e.g., shape parts (such as the legs, seat and back of a chair) or structural features (such as the feature contours of a chair). These higher-level representations then allow manipulating shapes at a more abstract level, independently of the underlying low-level representation. However, to create such higher-level representations, we first need to analyze the shapes, learn their semantic parts and the geometric relations among them.

In this talk, I will present our developments towards this goal. First, we introduce an unsupervised co-segmentation technique where we consistently segment a set of shapes coming from the same family. We achieve that by clustering shape parts in a descriptor space, which makes use of diffusion maps to explore the presence of third-party connections between parts. Next, we extend the unsupervised co-segmentation to efficiently incorporate direct user input, to arrive at a semi-supervised co-segmentation approach that allows obtaining a consistent segmentation that is close to error-free. Here, we make use of a spring system to obtain a part clustering that is constrained by the user input. Moreover, we are extending such representations to incorporate more semantics about the shapes. We learn the typical geometric configurations of parts throughout the set of shapes with a series of probability distributions, which can be used in applications such as repository exploration and guided shape editing. I will conclude the talk by giving a perspective on future directions for using such shape representations in content creation.

Bio

Oliver van Kaick received the B.Sc. and M.Sc. degrees in computing science from Universidade Federal do Parana (UFPR), Brazil (2003 and 2005), and a Ph.D. from the School of Computing Science at Simon Fraser University (SFU), Canada (2011). In 2012, Oliver was a postdoctoral researcher at SFU as a MITACS Elevate Fellow, collaborating with his industrial partner PDFTron Inc. on a project on document layout analysis. Currently, he is a postdoctoral researcher at Tel Aviv University as an Azrieli Fellow. Oliver's research interests are concentrated in the area of computer graphics, including topics such as shape analysis, shape matching, and geometric modeling, while his general interests also include computer vision and machine learning. In his work, Oliver has collaborated in the development of techniques for shape matching and correspondence, as well as techniques for higher-level analysis of 3D shapes.