PhD Defense: "Localized Feature Representations for Classification and Visual Search"

Niloufar Pourian

May 21st (Thursday), 11:00am
Harold Frank Hall, Room 4164 (ECE Conference RM)

Searching for images with a specific visual content has been a topic of intense research in recent years. However, much of the recent work is focused on global image representations. Searching for regions of interest in a given image, or querying the image database with a given configuration of objects are still challenging problems. In such scenarios, having an effective visual feature representation is crucial. In this talk I will present localized feature representations for classification and visual search. A weakly supervised approach to semantic segmentation is developed. Starting with an initial coarse segmentation, a spectral clustering approach groups related image parts into communities. A community-driven graph is then constructed that captures spatial and feature relationships between communities while a label graph captures correlations between image labels. Finally, mapping the image level labels to appropriate communities is formulated as a convex optimization problem. The proposed methods are computationally efficient, can scale to large image databases, and experimental results compare favorably with the state-of-the-art methods.

About Niloufar Pourian:

photo of niloufar pourian Niloufar Pourian was born in Tehran, Iran. She received her Bachelors and Masters degrees in Electrical and Computer Engineering from the University of California at Santa Barbara in 2010 and 2012, respectively. She is currently pursuing her PhD degree at Vision Research Lab in UCSB. Her research interests include large scale image classification, object recognition, semantic segmentation, and feature extraction. She is the recipient of 3-year UC Regents Fellowship and 4-year Dean’s Doctoral Scholar Award in 2007 and 2010, respectively.

Hosted by: Professor B.S. Manjunath