Visual Object Modeling And Detection With And-Or Graph Model And Discriminative Learning

Object detection refers to locate and name all the object instances in an input imagewhich belong to one or more than one object categories (i.e., what objects are where).Object detection is one of most important and challenging problems in computer vision andpattern recognition, which has a wide range of applications, including intelligent videosurveillance, driver assistant system, driverless car, content-based image or video retrieval,etc.. This dissertation focuses on the core issue in object detection, that is, the hierarchicaland compositional modeling and discriminative learning, by developing two types ofmethods, one is based on implicit and flat structure and the other explicitly hierarchical andcompositional representation.This dissertation proposes a sparse image patch model for object detection. Itconstructs an over-complete set of candidate image patches, and then develops a learningalgorithm, like Adaboost, to mine stable and discriminative image patches for objectdetection with their weights being learned simultaneously. The experimental results showthat the proposed method is very effective.This dissertation presents a generic quantization And-Or graph and a dynamicprogramming algorithm for learning the globally optimal latent object structure in objectdetection, which is one of the most challenging problem in object modeling. To tackle thewell-known large variations of both structure and appearance of objects, thedivide-and-conquer methodology is adopted to learn hierarchical and compositional objectmodels. However, in the training dataset, only the bounding boxes of objects themselves aregiven, so model structures are treated as hidden variables. By unfolding the space of latentobject structures, the proposed quantization And-Or graph can explore all possible partconfigurations using a dynamic programming algorithm, so the globally optimal structurecan be sought efficiently.This dissertation presents a method of discriminatively training And-Or tree model forobject detection in PASCAL VOC which is the most widely-used and challengingbenchmark for object detection. Based on the quantization And-Or graph, it proposes amethod of learning the subcategories of an object category in an unsupervised manner (e.g.,different viewpoints which are unknown in traning data), and a semi-supervised method oflearning the globally optimal And-Or tree model for each subcategory. The model parameters are trained under the latent structural SVM framework. Thanks to the treestructure, a dynamic programming algorithm is utilized in object detection. Theexperimental results on PASCAL VOC show that the proposed And-Or tree modeloutperforms state-of-the-art comparable methods.