Incremental Graph Representation for Feature-based Object Learning and Detection

Abstract

Visual recognition is one of the most important and challenging topics in computer vision. In the past decades, many approaches have been proposed to develop more accurate and robust recognition systems. Among existing approaches, recently more attention is paid to feature-based methods due to their remarkable performance and speed. However, most of them still have some restrictive assumptions such as descriptive object database and accurate feature correspondences. To tackle such limitations, this thesis particularly deals with the problems of feature correspondence and object learning and representation.

The main objective of this thesis is to develop a robust and easily applicable object learning and recognition algorithm. For this purpose, new algorithms that are used as important components of entire recognition system are proposed. First, we develop a feature matching algorithm that utilizes both geometric and photometric properties of local features to obtain better feature correspondences. Second, the proposed matching algorithm is extended to an unied framework for feature matching and image retrieval. Third, we propose an incremental graph representation method which is employed to simultaneously detect target objects in query images and build object database based on the detection results. Fourth, a robust stereo matching algorithm is developed to improve the performance of 3D object reconstruction. By combining these methods, we develop a new object learning and detection system which incrementally builds target object database from a single seed image. Each algorithm performance is experimentally evaluated using well-known benchmark datasets. The experimental results demonstrate that the proposed methods are effective and produce better performance compared to the conventional algorithms. Although this thesis focuses on feature-based object recognition, all algorithms proposed here can be applied to other related fields.