Some New Directions in Energy Minimization with Graph Cuts Ramin Zabih Cornell University Algorithms based on graph cuts have had a major impact on an important class of vision problems. In these problems, which arise in applications such as stereo, every pixel must be assigned a label from some predefined set. There is a known cost to assign a given label to any pixel, as well as a known cost for assigning different labels to adjacent pixels. I will describe some recent work that addresses a broader class of problems, where the labels are not specified in advance, and where the cost of assigning a given label to any pixel must be determined. This broader class of problems arises from applications such as multi-modal imaging, stereo imaging with unknown camera gain/bias, texture segmentation and layered motion segmentation. I will present some preliminary results that suggest that these problems can be solved by combining graph cuts with ideas from Expectation-Maximization and mutual information. This is joint work with Junhwan Kim and Vladimir Kolmogorov. Bio: Ramin Zabih is an Associate Professor of Computer Science at Cornell University. He did his undergraduate work at MIT, and received his PhD from Stanford in 1994. He is best known for his work on the use of graph cut algorithms in computer vision; two of his papers on this topic received a Best Paper award at the European Conference on Computer Vision (ECCV) in 2002. Since 2001 he has held a joint appointment with the Radiology Department at Cornell Medical School. He has also consulted extensively for a number of groups at Microsoft, including product groups as well as MSR, and has recently been serving as an expert witness for Microsoft.