Object appearance modeling and salient object detection for image understanding

Abstract

Object analysis is an indispensable part of image understanding. Although challenging, discovering object properties remains a main task of various computer vision applications, such as object recognition, figure-ground segmentation, object localization, etc.. In this thesis, we explore object properties from two aspects, modeling object appearance changes in a sequence of images, and finding objects of interest in an image automatically. Although these two tasks extract different image information, they could be combined to form an automatic visual tracking system. This is because tracking object appearance changes requires input of the initial target objects, while bottom-up salient object detection locates objects of interest automatically, which may serve as tracking targets. First, we present a novel locality sensitive histogram (LSH) algorithm to represent object appearances for visual tracking. Based on the proposed efficient histogram, a robust tracking framework is then presented, which consists of two main components: a new feature for tracking which is robust to illumination changes, and a novel multiregion tracking algorithm that runs in real-time even with hundreds of regions. Evaluations using our dataset and the latest benchmark show that the proposed tracking algorithm is the top performer. Second, we propose two salient object detection algorithms to locate objects of interest. The first salient object detection algorithm that we propose uses a pair of flash and no-flash images. It is independent of color information, enabling robust detection even in the scenes with similar foreground and background colors. The second salient object detection algorithm that we propose learns the hierarchical contrast features describing salient objects. A novel superpixelwise convolutional neural network approach, called SuperCNN, is proposed to learn the internal representations of saliency in an efficient manner. Both of these two algorithms show superior performances over the state-of-the-art algorithms.

Date of Award	2 Oct 2015
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Hao YUAN (Supervisor) & Rynson W H LAU (Supervisor)