Category-level object detection and image classification

Abstract

Automatic recognition of object categories from complex real-world images is an exciting problem in computer vision. While humans can perform recognition tasks effortlessly and proficiently, replicating this recognition ability of humans in machines is still an incredibly difficult problem. On the other hand, successful automatic recognition technology will have significant and mostly positive impact in a plethora of important application domains like image retrieval, visual surveillance and automotive safety systems. This dissertation addresses two main goals of recognition: image classification and object detection. Image classification seeks to separate images which contain an object category from other images, where the focus is on identifying the presence or absence of an object category in an image. Object detection concerns the identification and localization of object instances of a category in an image, where the goal is to localize all instances of that category from the image.

Our main contributions toward this end are threefold. Firstly, we develop a novel and powerful method to detect ellipses from edge images. Our method specifically addresses the structural issues related to broken edge maps, background clutter and partial occlusion. Additionally, we incorporate a self-correcting mechanism into the ellipse detector which empowers it with an ability to identify weak ellipses and to regenerate new ellipses that better represent edge information. Experimental evaluation on complex synthetic and real images shows our ellipse detection method to have systematic and substantial improvements over previous methods. We are unaware of any other works that can detect ellipses from such difficult images. Consequently, the proposed method advances the state-of-the-art in ellipse detection.