CENS Technical Seminar Series

Object Categorization in Computer Vision

Invited Speaker: Brian Fulkerson, CS, UCLA
Date: February 1, 2008
Time: 1:00 PM - 2:00 PM
Venue: CENS Conference Room

Abstract

In computer vision, object category recognition refers to the process of automatically deciding what kind of object is present in an image. The most successful approaches so far have been variations on an algorithm known as bag of features. We will provide an overview of the major pieces of the bag of features pipeline: feature extraction, dictionary construction, signature construction, learning, and classification.

We will then focus the talk on our experiences with and modifications to this pipeline. Specifically, we have been working on two key areas: the incorporation of region boundary information into the features, and the construction of the dictionary. First, we will detail our experiences using segmentation to improve the quality of the features. Next, we will describe an algorithm which minimizes the size of the dictionary while maintaining its usefulness for the task of category recognition. This will lead us to an approach which leverages this compressed visual dictionary to provide an efficient and accurate algorithm for solving a related problem: determining which pixels in an image contain the object category we are searching for.

Biography

Brian Fulkerson is a PhD student in the Vision Lab at UCLA under Stefano Soatto. His interests center on computer vision as it applies to a variety of domains including object category recognition, robot navigation, image databases, and embedded imaging. Brian obtained his Masters from UCLA in 2006 and his B.S. from UCSD in 2004.