[Colloquium] TTIC Colloquium: Pawan Mudigonda, Stanford

[Julia MacGlashan]

When:             *Monday, June 28 @ 1:00pm*

Where:           * TTIC Conference Room #526*, 6045 S Kenwood Ave, 5th Floor


Who:              * **Pawan Mudigonda*, Stanford University


Title:          *      **LP Relaxation and Move Making Algorithms for Energy
Minimization***


The linear programming (LP) relaxation for minimizing the energy of discrete
random field models is known to provide the best multiplicative bounds for a
large class of problems. However, despite the recent message passing
algorithms for the LP, it remains a computationally expensive method. In
contrast, the iterative move-making algorithms are typically faster (as they
solve an st-MINCUT problem), but provide poor guarantees.

The first part of the talk will describe some new move-making algorithms
that combine the best of both worlds. Specifically, (i) they provide the
same guarantees as the LP relaxation; and (ii) they retain the computational
benefit of move-making by relying only on st-MINCUT.

The second part of the talk will focus on the problem of selecting regions
for high-level scene understanding. Typically, such regions are selected
beforehand using a bottom-up over-segmentation approach and then an energy
function is defined over them. We argue that the regions themselves must be
selected using the energy function. To facilitate the selection of such
regions from a large dictionary, we extend the LP relaxation and design an
efficient algorithm for solving it. We demonstrate the efficacy of our
method on semantic scene segmentation using the Stanford background dataset.

Host:              Tamir Hazan, tamir at ttic.edu <tamir at ttic.edu>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20100622/4d9dce9d/attachment.htm