The objective of this thesis is to explore the use of graph matching in object recognition systems. In the continuity of the previously described articles, rather than using descriptors invariant to misalignment, this work directly tries to find explicit correspondences between prototypes and test images, in order to build a robust similarity measure and infer the class of the test images. In chapter 2, we will present a method that given interest points in two images tries to find correspondences between them. It extends previous graph matching approaches [Leordeanu and Hebert, 2005a] to handle interactions between more than two feature correspondences. This allows us to build a more discriminative and/or more invariant matching method. The main contributions of this chapter are: The introduction of an high-order objective function for hyper-graph matching (Section 2.3.1). The application of the tensor power iteration method to the high-order matching task, combined with a relaxation based on constraints on the row norms of assignment matrices, which is tighter than previous methods (Section 2.3.1). An l1-norm instead of the classical l2-norm relaxation, that provides solutions that are more interpretable but still allows an efficient power iteration algorithm (Section 2.3.5). The design of appropriate similarity measures that can be chosen either to improve the invariance of matching, or to improve the expressivity of the model (Section 2.3.6). The proposed approach has been implemented, and it is compared to stateof-the-art algorithms on both synthetic and real data. As shown by our experiments (Section 2.5), our implementation is, overall, as fast as these methods in spite of the higher complexity of the model, with better accuracy on standard databases. In chapter 3, we build a graph-matching method for object categorization. The main contributions of this chapter are: Generalizing [Caputo and Jie, 2009; Wallraven et al., 2003], we propose in Section 3.3 to use the optimum value of the graph-matching problem associated with two images as a (non positive definite) kernel, suitable for SVM classification. We propose in Section 3.4 a novel extension of Ishikawa's method [Ishikawa, 2003] for optimizing MRFs which is orders of magnitude faster than competing algorithms (e.g., [Kim and Grauman, 2010; Kolmogorov and Zabih, 2004; Leordeanu and Hebert, 2005a]) for the grids with a few hundred nodes considered in this article). In turn, this allows us to combine our kernel with SVMs in image classification tasks. We demonstrate in Section 3.5 through experiments with standard benchmarks (Caltech 101, Caltech 256, and Scenes datasets) that our method matches and in some cases exceeds the state of the art for methods using a single type of features. In chapter 4, we introduce our work about object detection that perform fast image alignment. The main contributions of this chapter are: We propose a novel image similarity measure that allows for arbitrary deformations of the image pattern within some given disparity range and can be evaluated very efficiently [Lemire, 2006], with a cost equal to a small constant times that of correlation in a sliding-window mode. Our similarity measure relies on a hierarchical notion of parts based on simple rectangular image primitives and HOG cells [Dalal and Triggs, 2005a], and does not require manual part specification [Felzenszwalb and Huttenlocher, 2005b; Bourdev and Malik, 2009; Felzenszwalb et al., 2010] or automated discovery [Lazebnik et al., 2005; Kushal et al., 2007].
Identifer | oai:union.ndltd.org:CCSD/oai:tel.archives-ouvertes.fr:tel-01063352 |
Date | 29 November 2012 |
Creators | Duchenne, Olivier |
Publisher | Ecole Normale Supérieure de Paris - ENS Paris |
Source Sets | CCSD theses-EN-ligne, France |
Language | English |
Detected Language | English |
Type | PhD thesis |
Page generated in 0.0016 seconds