"Deep learning architectures for 3D shape analysis and synthesis"
Evangelos Kalogerakis, University of Massachusetts Amherst
Wednesday, October 11, 2017 at 12:00 Noon
Room 368 (CIT 3rd Floor)
The emergence of low-cost 3D acquisition devices, such as the Kinect, and the appearance of large-scale shape repositories, such as the 3D Warehouse, are revolutionizing computer graphics, making 3D content ubiquitous. The need for algorithms that understand and intelligently process 3D shapes is thus greater than ever. In this talk, I will present my latest research on deep learning architectures for 3D shape analysis and synthesis. Specifically I will first describe architectures that combine image-based convolutional networks and surface-based probabilistic graphical models. The image-based networks process shape projections across multiple views, and the probabilistic models aggregate the resulting multi-view outputs on the shape surface. In contrast to other deep learning approaches for 3D shape processing, these projective architectures allow fast shape processing at high resolutions, are robust to input geometric representation artifacts (non-manifold geometry, polygon soups, arbitrary or no interior), combine both image and shape datasets for training, and focus their representation power on the shape surface. As a result, they offer state-of-the-art performance in several shape processing tasks, such as shape segmentation, retrieval, correspondences, and sketch-based 3D shape synthesis. Finally, I will discuss deep learning architectures that combine view-based and volumetric shape representations to achieve 3D reconstruction from partial shape data.
Evangelos Kalogerakis obtained his PhD from the department of Computer Science, University of Toronto in 2010. His PhD introduced machine learning algorithms for geometry processing. He was a postdoctoral researcher at the Computer Graphics lab of Stanford University from 2010 to 2012. He joined the College of Information and Computer Sciences at the University of Massachusetts Amherst as an assistant professor in 2012. His research deals with the development of computer vision, graphics and machine learning techniques that analyze and synthesize visual content. He is particularly interested in designing learning algorithms that help people to easily create and process 3D models of objects.
Host: Professor Daniel Ritchie