University of California, Riverside

Department of Electrical and Computer Engineering



Multi‐class Object Detection and Layered Segmentation


Multi‐class Object Detection and Layered Segmentation
 

Abstract:

There is a large body of research on both the problems of object detection and image segmentation. I will describe our recent work on extending these methods to the problem of parsing a scene which contains multiple objects of different classes. We have introduced a unified model for multi‐class object detection that casts the problem of outputting multiple detections as a structured prediction task. Rather than predicting a binary label for each candidate image sub‐window independently, our model simultaneously predicts a structured labeling of the entire image. Our model learns statistics that capture the spatial arrangements of various object classes in real images, both in terms of which arrangements to suppress and which arrangements to favor.

Turning such detections into a detailed parse of the scene requires estimating segmentations for each detected object. We formulate a probabilistic layered model for object detection and multi‐class segmentation. Our system uses the output of a bank of object detectors in order to define shape priors and then estimates appearance, depth ordering and labeling of all pixels in the image. We train our system on the PASCAL segmentation challenge dataset and show good test results in several categories including segmenting humans.

Biography:

Charless Fowlkes, Ph.D.Charless Fowlkes received his BS from Caltech in 2000 and a PhD in Computer Science from UC Berkeley in 2005 where his research was supported by an NSF graduate research fellowship. He was a post‐doc at Lawrence Berkeley Labs where he worked on applications of computer vision techniques to understanding gene regulatory networks in animal development. Since 2007 he has been an assistant professor in Computer Science at UC Irvine. His research is in computational vision, in particular how to combine bottom‐up processing, such as image segmentation with top‐down information, such as recognition of familiar shapes. He is interested in measuring the predictive power of different visual cues to derive general informationtheoretic constraints on human visual processing. He also works on developing computational tools for the analysis of biological shape and image data.

 

Date: November 29th
Location: A265 Bourns Hall
Time: 2:00-3:00pm

More in Colloquia

More Information 

General Campus Information

University of California, Riverside
900 University Ave.
Riverside, CA 92521
Tel: (951) 827-1012

Department Information

Electrical and Computer Engineering
Suite 343 Winston Chung Hall
University of California, Riverside
Riverside, CA 92521-0429

Tel: (951) 827-2484
Fax: (951) 827-2425
E-mail: E-mail/Questions

Footer