SISTA Colloquium Series
|Date||Monday, October 3, 2011|
|Details||Light refreshments in Gould-Simpson 9th floor atrium before talk.|
|Speaker||Kobus Barnard, Associate Professor|
|School/Dept.||School of Information: Science, Technology, and Arts (SISTA)|
|Affiliation||University of Arizona|
Sampling plants, fungus, and bedrooms.
Inferring and quantifying geometric structure from images has great potential to impact many disciplines. I will present a general approach for doing so being developed by the SISTA computer vision group. We are particularly interested in semantic representations of objects and organisms, where the attributes of the parts and their geometric relationships can be linked to other data, such as molecular data, survivability, yield, and environmental factors. For example, our representation of plants involves familiar organs such as stems and leaves and their associated geometry, and thus our overall geometric model encodes things like the distribution over branch points, average leave size, and many other similar quantities. It is thus distinctive from image processing methods that focus quantifying a particular aspect of the plant.
We connect these models to image data using generative statistical models, and infer the parameters using MCMC sampling methods (hence the title). Our approach can integrate a variety of imaging methodologies, and I will present examples using three of them: 1) multiple 2D views in the case of Arabidopsis, 2) brightfield image stacks in the case of filamentous fungus from the genus Alternaria; and 3) single perspective images in the case of indoor scenes.
This is joint work with Barry Pryor, Ravi Palanivelu, Joseph Schlecht, Kyle Simek, and Luca del Pero. Contributions have also been made by Ernesto Brau, Jinyan Guan, Emily Hartley, Sam Martin, and Ekatarina Taralova.
Kobus Barnard is an associate professor with the School of Information: Science, Technology, and Arts (SISTA) at the University of Arizona. He received his Ph.D. in computer science from Simon Fraser University in the area of computational color constancy. He then spent two years at the University of California at Berkeley as a post doctoral researcher working on image understanding and digital libraries. His current interests include statistical modeling of 3D object structure in everyday, robotic, and biological domains, extracting semantics from multimodal data, and applying new algorithmic approaches to scientific data.