Date of Award

2010

Degree Type

Thesis

Degree Name

Master of Science

Program

Computer Science

Supervisor

Dr. Olga Veksler

Abstract

ALTHOUGH multiple object recognition is an effortless job for humans, it still remains a challenging task in computer vision. The problem is challenging because even for rigid objects, there is a large variation in appearance due to difference in viewpoint, orientation, and occlusion. Most object recognition approaches are built to recognize an object in a rectangular image window. Therefore at application time, an image is broken into multiple sub-windows, and a detector is applied to each subwindow, in turn. This is called the sliding window scanning. However there are three main problems associated with the sliding window scanning. First of all, it is mostly appropriate for recognizing objects that are well-approximated by a rectangle, such as faces and pedestrians, but not for objects that are not well approximated by a box, such as giraffes. Another problem is that this approach does not provide a precise pixel-level segmentation of the detected object. Last but not least, another problem is that sliding window approach is very expensive computationally, since a large amount of sub-windows needs to be processed in a single image. To avoid all of the above mentioned problems, a recent popular approach to object detection is to include image segmentation in the detection process as a pre-processing step. Segmentation is used to subdivide an image into a set of regions. These regions do not have predefined shapes like rectangular patches and their boundaries are more likely to align with the boundaries of objects in the scene. In addition, the shape and boundary properties can be used for feature extraction in the object detection system. Image segmentation is notoriously brittle, so a solution is to integrate the results over multiple different segmentations of the same scene. This thesis develops a novel object detection algorithm that relies on multiple segmentations as a pre-processing step. There is previous work that uses multiple segmentations, however they usually make the assumption that all the regions generated by multiple segmentations are useful for the task of recognition. This assumption 111 is often violated since lots of these regions do not have informative features to distinguish an object or an object part. The overall idea of the approach in this thesis is to find reliable regions and discard the ones that are not reliable for object detection. This will result in more certain and stable decisions at the time of object recognition.

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.