Visual understanding by salient proto-objects

Download files
Access & Terms of Use
open access
Copyright: Li, Zhidong
Altmetric
Abstract
In this thesis, we propose biologically inspired approaches for visual understanding tasks including object detection, image categorization, and visual tracking, based on the concept of salient proto-objects. A salient proto-object refers to a visual unit that can form a coherent and stable object on the basis of saliency. It is been demonstrated in this work that the salient proto-object is useful in machine based visual understanding. First, we propose a novel computational framework for salient proto-object detection, which integrates the saliency map computation and proto-object detection. The proto-objects are detected based on the saliency map, and the detected proto-objects are then utilized to improve the saliency map computation in an iterative manner. Second, we propose an approach of image categorization by integrating a latent topic model with proto-object detection. After discriminating salient proto-objects from background parts in images, a hierarchical latent topic model is proposed to discover image topics for scene and event categorization. In addition, a joint topic model is proposed for image based object categorization. Third, we propose an approach of visual tracking based on salient proto-objects. Given an image sequence, proto-objects are first detected by combining the saliency map and topic model. Then the target is tracked based on spatial and saliency information of the proto-objects. In the proposed Bayesian approach, states of the target and proto-objects are jointly estimated over time. Gibbs sampling has been used to optimize the estimation during the tracking process. The proposed method robustly handles occlusion, distraction, and illumination change in the experiments. The algorithms have been tested using public datasets for visual understanding. The experimental results show that the proposed approaches outperform the state-of-art methods in various challenging tasks, including salient proto-object detection, image categorization, and visual tracking.
Persistent link to this record
Link to Publisher Version
Link to Open Access Version
Additional Link
Author(s)
Li, Zhidong
Supervisor(s)
Yang, Wang
Fang, Chen
Alan, Blair
Creator(s)
Editor(s)
Translator(s)
Curator(s)
Designer(s)
Arranger(s)
Composer(s)
Recordist(s)
Conference Proceedings Editor(s)
Other Contributor(s)
Corporate/Industry Contributor(s)
Publication Year
2013
Resource Type
Thesis
Degree Type
PhD Doctorate
UNSW Faculty
Files
download public version.pdf 4.72 MB Adobe Portable Document Format
Related dataset(s)