Oregon State University



Event Details

MS Final Examination – Xu Hu

Monday, March 9, 2015 10:00 AM - 1:00 PM

Reconfigurable Chain Models for Part-Based Object Recognition
This paper presents a new model for parsing and localizing chain structured objects, where object parts may be invisible due to partial occlusions or phenotypic variations. We model the locations of parts and the dependencies between all possible pairs of parts using a graphical model. For a testing image, the goals are finding the locations of parts and the best chain structured parse graph. We show that an exact inference using dynamic programming can be performed under the assumption that all the parse graphs are substrings of a chain graph. The parameters are learned jointly using a structured SVM. We evaluate our model for two real problems: human pose estimation with partial occlusions and part-based biological character recognitions. The results show that our model achieves significant improvements over baseline models.

Major Advisor: Sinisa Todorovic
Committee: Thomas Dietterich
Committee: Prasad Tadepalli

Kelley Engineering Center (campus map)
Nicole Thompson
1 541 737 3617
Nicole.Thompson at oregonstate.edu
Sch Elect Engr/Comp Sci
This event appears on the following calendars: