Hierarchical graphical model-based methods for object identification and categorization with visual context

Abstract

The goal of object recognition is to label objects from images and to estimate the poses of the labeled objects. The field of object recognition has seen tremendous progress with successful applications in some specific domains such as face recognition. However, the current state-of-the-art methods show unsatisfactory results for more general object domains in complex natural environments with visual ambiguities. In this dissertation, we aim to enhance the object identification and categorization with the guide of visual context and graphical model. In this dissertation, we propose a general framework for the cooperative object identification and object categorization. Examplars used in identification provide useful information of similarity in categorization. Conversely, novel objects are rejected in identification but the proposed object categorization can label the novel objects and segment them out for database update in identification. In the first part of the work, we propose a hierarchical graphical model (HGM) for the disambiguation of blurred objects. We define three types of visual context such as spatial, hierarchical, and temporal context, which provide powerful disambiguation. To handle both the visual relation and uncertainty, we model them by the HGM. It consists of part layer, object layer, and a place node. Pose information in part and object layer is inserted into nodes for the utilization of part-object context. Due to the complexity of graphical model, we apply the piecewise learning which gives practical learning of the HGM, and propose a context-guided sample generation and pruning for the variable graph estimation and distribution estimation. The bidirectional interaction in the HGM can discriminate ambiguous objects and places simultaneously in real environment. Large scale experiments for building guidance validate the robustness. As a direct extension, the HGM is adapted for the video interpretation by incorporating additiona...