| Object segmentation, a fundamental problem in computer vision, remains a challenging task after decades of research efforts. This task is made difficult by the intrinsic variability of the object's shape, appearance, and its surrounding. It is compounded by the uncertainties arising from mapping the 3D world to the image plane and the noise in the acquisition systems. However, the human visual system often effectively entails the segmentation of the object from its background by fusing the bottom-up image cues with the top-down context. In this thesis we propose a novel probabilistic graphical modeling framework for object segmentation that effectively and flexibly fuses different sources of information, top and bottom, to produce highly accurate segmentation of objects in a computationally efficient manner. The main contributions of our work are: (1) We present a graphical model representing the relationship of the observed image features, the true region labels, and the underlying object contour based on the integration of Markov Random Fields (MRF) and deformable models. We propose two different solutions to this otherwise intractable joint region-contour inference and learning problem in the graphical model. (2) We introduce a Profile Hidden Markov Model (PHMM) built on the shape curvature sequence descriptor to improve the segmentation of specific objects. The special states and structure of PHMMs allow considerable shape contour perturbations and provide efficient inference and learning algorithms for shape modeling. Further embedding of the PHMM parameters captures the long term spatial dependencies on a shape profile, hence the global characteristics of a shape class. (3) We incorporate the proposed methods in a spatio-temporal MRF model to solve the video-based object segmentation problem. Our new model is a simultaneous object segmentation, background modeling, and pose estimation framework, which combines the top-down high-level object shape constraints with the bottom-up low-level image cues, and features a flexible graph structure induced by the motion information for more reliable temporal smoothness.;We demonstrate the effectiveness and robustness of all our methods in a wide variety of thorough experiments. |
