Machine Learning Algorithms For 3 D Shape Understanding And Modeling
Abstract: The world we see around us consists of three-dimensional shapes. Having algorithms that are able to automatically interpret and understand three-dimensional geometry is very useful for assisting us to create shapes for virtual worlds, architecture, computer games, 3D printing and information visualization applications.
In this talk, I will discuss machine learning techniques that map raw three-dimensional geometry to higher-level concepts, such as shape parts, tags and attributes. The techniques are based on discriminative classifiers, such as CRFs learned by human-labeled data. After inferring these higher-level concepts, we can use them to form a meaningful space of shapes that users can quickly explore to create new shapes. For example, if a user wants to create a strong and agile character for a computer game, she can browse the shape space according to attributes that are related to strength and agility of character parts. In addition, I will discuss how a probability distribution can be defined over this shape space, so that novice users are guided into creating more plausible shapes. The probability distribution is based on a generative probabilistic model with latent variables that is able to capture the relationships between high-level concepts and three-dimensional geometry.
Short bio: Evangelos (Vangelis) Kalogerakis joined the Department of Computer Science at the University of Massachusetts in September 2012. He obtained his PhD from the department of Computer Science, University of Toronto in September 2010. He was co-supervised by Aaron Hertzmann and Karan Singh. He was a postdoctoral researcher in the Computer Graphics lab of Stanford University from September 2010 to August 2012.