Regression Methods in Semantic Segmentation, Video Object Discovery and Tracking
KEC 1003 Mon, 11/16/2015 - 4:00pm Fuxin Li Assistant Professor, School of EECS, Oregon State University Abstract: Many of the structural prediction problems in computer vision can be converted into simple regression problems on object proposals followed by specific higher-order inference schemes. The benefit of this approach is that learning can be simple without the need of doing structural prediction, while inference carries the bulk of the optimization load. This approach started from object proposals in semantic segmentation, while carrying over to many other problems such as video object discovery and multi-target tracking, where the efficiency of the least squares regression helps learning thousands of models with ease. A new type of higher order inference called composite statistical inference is proposed to utilize those regression estimates. This inference breaks and recombines object proposals and is flexible, higher-order, and statistically consistent. It is shown to work well in complex problems with significant interactions among objects. Bio: URL: http://eecs.oregonstate.edu/colloquium/regression-methods-semantic-segmentation-video-object-discovery-and-tracking _______________________________________________ Colloquium mailing list [email protected] https://secure.engr.oregonstate.edu/mailman/listinfo/colloquium
