Object Segmentation using Spatial and Spatio-Temporal Features.
Omid Aghazadeh · 2011
This thesis investigates a variational algorithm for object segmentation using multiple spatial and spatio-temporal cues. The aim is to segment images into two regions: foreground region representing an object of interest or a conspicuous object and background region representing everything else. A variational segmentation framework(region based active contours) is utilized with a functional combining statistics of different cues for segmentation such as color, motion and texture. Moreover, a classification approach to object boundary (occlusion boundary) detection, which combines appearance and several motion cues is presented. A score function was developed for the purpose of assessing the quality of any edge detection algorithm, which is more reliable than the accuracy of the detections. The concept of localized classifiers is presented and discussed, which leads to significant speedups in the training / testing time of classifiers with the cost of additional bias towards the training data. A localized classifier set and a radial basis kernel SVM were trained and using the mentioned score function, they were compared to each other. Comparable results of the mentioned methods verifies the efficiency of the concept of the localized classifier sets in terms of computational costs and accuracy. Furthermore, the use of Geodesic Active Contours to encourage the level sets to converge to some sparsely defined intermediate boundaries (which in this case would be the detected object boundaries) is investigated. Additionally, a data-set of 2 frame sequences with the ground truth information is prepared and presented and the results of both algorithms are presented on the mentioned data-set. The data-set consists of 25 sequences, 17 of which were taken carefully by a camera undergoing a translational movement and contain indoor / outdoor sequences with different lighting conditions and different degrees of complexity of the scene as well as 8 sequences chosen from an earlier published data-set. Various concepts of the variational segmentation approach such as sensitivity to the parameters, parameter tuning, performance of the algorithm in different cases were investigated and the usefulness of the proposed methods were investigated using qualitative and quantitative results.