Boston University | Center for Computational Science
HomeNews and EventsResearchEducationPeopleSeminarsFacilitiesContact Us

 

CCS Seminar
Benjamin Chandler
Boston University

Center for Computational Neuroscience and Neural
Technology (CompNet) - CNS
Friday - April 20, 2012
12:00 noon
Physics Research Building - Room 595
3 Cummington Street

"Real-time Disocclusion in Natural Environments"

Ben Chandler [1, 2], Greg Snider [1], and Ennio Mingolla [2]
[1] HP Labs, 1501 Page Mill Road, Palo Alto, CA 94304
[2] CELEST and CompNet, Boston University, 677 Beacon Street, Boston, MA 02215
{ben.chandler, snider.greg}@hp.com, ennio@bu.edu

Many problems that are unavoidable in real-world environments pose significant difficulty for the current state-of-the-art in computer vision. Unknown lighting, shadows, and occlusion are pervasive in such environments. Insight in to the mechanisms by which biological organisms behave robustly under these challenging conditions offers a potential solution. Biological vision systems can discount the influence of the illuminant to build surface representations invariant to lighting. Modal and amodal boundary completion are powerful tools for building similar kinds of invariances in the boundary representation. Junctions in the boundary representation are local hints for depth and occlusion. Diffusive propagation of information fuses boundary, depth, and feature information into a single, consistent percept. These principles offer a significant scientific base from which to build, but are not sufficient to build an implementation suitable for use in technology. Real-time operation on high-resolution video using commodity GPU hardware requires technical tools like the Cog ex Machina development platform and mathematical tools like steerable filters, Fourier-domain convolution, and geometric algebra. In this talk, I’ll discuss progress towards leveraging those insights to build an artificial visual system capable of monocular, single-frame disocclusion in real time in natural environments.

Acknowledgments
This work was supported in part by CELEST, a National Science Foundation Science of Learning Center (SBE-0354378)





























Home


Page last updated on April 10, 2012. Please send comments to Cheryl Endicott

copyright © 2006, Center for Computational Science | Boston University , MA, 02215