Robust Estimation Methods for Image Matching

This dissertation, "Robust Estimation Methods for Image Matching" by Chunlin, Feng, 馮淳林, was obtained from The University of Hong Kong (Pokfulam, Hong Kong) and is being sold pursuant to Creative Commons: Attribution 3.0 Hong Kong License. The content of this dissertation has not been altered in any way. We have altered the formatting in order to facilitate the ease of printing and reading of the dissertation. All rights not granted by the above license are retained by the author. Abstract: Abstract of Thesis Entitled Robust Estimation Methods for Image Matching Submitted by FENG Chun Lin for the degree of Master of Philosophy at The University of Hong Kong in August 2004 This study proposes new image matching methods for matching feature points across a pair or triple of views through robust recovery of epipolar geometry or trifocal geometry. The aim of this study was to enable 3D scenes to be automatically reconstructed using projective geometry, assuming corresponding points are matched robustly using the methods proposed in the thesis. Its findings have applications to 3D reconstruction, robust estimation and object recognition. Image matching, i.e. depicting the process of recovering feature correspondences, is a challenging problem in computer vision. The crux of this problem is that putative matches are often poorly or incorrectly extracted by intensity-based cross-correlation methods. Geometry-based methods, inspired by geometric relationships governing point correspondences, are therefore used to examine the correctness of correspondence of putative matches. However, retrieving geometric relationships can be difficult in the presence of a fair proportion of mismatches. For this reason, the fundamental matrix for epipolar geometry or trifocal tensor in the case of trifocal geometry may well be incorrectly estimated due to misclassification of matches and mismatches. In order to overcome this problem, two methods are proposed in this study for matching two-view and three-view images, both involving intensity-based and geometry-base matching. Intensity-based matching forms putative matches based on image intensity, which involves corner detection and cross-correlation measurement. Geometry-based matching examines the correctness of correspondences of putative matches in a geometric perspective, thus enabling correct matches to be distinguished from mismatches. This method first determines the fundamental matrix or trifocal tensor from random samples, and then evaluates each solution through reprojection error measurement, parameter estimation, dataset classification and scoring, and finally yields the solution with the best score together with associated matches. The matching methods proposed in this study include: (a) incorporation of a maximum likelihood estimator for unknown parameters in the image error model, thus removing complexity arising from manual configuration of these parameters; (b) formulation of an effective cost function to score each solution by considering its consistency with estimated matches and the shape of its residual error distribution, thus enabling a fair measurement of solution error; and (c) determination and evaluation of solutions (the fundamental matrix for two views and the trifocal tensor for three views) by means of the same measure viz. the geometric error. The novelty of the proposed methods mainly lies in the study in part (a) and (b), and in the integration of the characteristics (a)-(c) into a single algorithm. Extensive evaluations are performed for both synthetic and real image sequences to validate the proposed methods. This study also includes a novel investigation of random sampling strategy to determine the optimal size for random sampling in the fundamental matrix estimation, thereby improving the computational efficiency of linear estimators. DOI: 10.5353/th_b2975269