The doctoral dissertations of the former Helsinki University of Technology (TKK) and Aalto University Schools of Technology (CHEM, ELEC, ENG, SCI) published in electronic format are available in the electronic publications archive of Aalto University - Aaltodoc.
|
|
|
Probabilistic Methods for Pose-Invariant Recognition in Computer Vision
Ilkka Kalliomäki
Dissertation for the degree of Doctor of Science in Technology to be presented with due permission of the Department of Electrical and Communications Engineering for public examination and debate in Auditorium C at Helsinki University of Technology (Espoo, Finland) on the 2nd of November, 2007, at 12 noon.
Dissertation in PDF format (ISBN 978-951-22-8996-7) [10782 KB]
Dissertation is also available in print (ISBN 978-951-22-8995-0)
Abstract
This thesis is concerned with two central themes in computer vision, the properties of oriented quadrature filters, and methods for implementing rotation invariance in an object matching and recognition system. Objects are modeled as combinations of local features, and human faces are used as the reference object class. The topics covered include optimal design of filter banks for feature detection and object recognition, modeling of pose effects in filter responses and the construction of probability-based pose-invariant object matching and recognition systems employing oriented filters.
Gabor filters have been derived as information-theoretically optimal bandpass filters, simultaneously maximizing the localization capability in space and spatial-frequency domains. Steerable oriented filters have been developed as a tool for reducing the amount of computation required in rotation invariant systems. In this work, the framework of steerable filters is applied to Gabor-type filters and novel analytical derivations for the required steering equations for them are presented. Gabor filters and some related filters are experimentally shown to be approximately steerable with low steering error, given suitable filter shape parameters. The effects of filter shape parameters in feature localization and object recognition are also studied using a complete feature matching system.
A novel approach for modeling the pose variation of features due to depth rotations is introduced. Instead of manifold learning methods, the use synthetic data makes it possible to apply simpler regression modeling methods. The use of synthetic data in learning the pose models for local features is a central contribution of the work.
The object matching methods considered in the work are based on probabilistic reasoning. The required object likelihood functions are constructed using feature similarity measures, and random sampling methods are applied for finding the modes of high probability in the likelihood probability distribution functions. The Population Monte Carlo algorithm is shown to solve successfully pose estimation problems in which simple Metropolis and Gibbs sampling methods give unsatisfactory performance.
Keywords: pattern recognition, face recognition, Gabor filter, steerable filter
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
© 2007 Helsinki University of Technology
Last update 2011-05-26