Incremental object matching with probabilistic methods

Thumbnail Image
Journal Title
Journal ISSN
Volume Title
Aalto-yliopiston teknillinen korkeakoulu | Doctoral thesis (monograph)
Checking the digitized thesis and permission for publishing
Instructions for the author
Degree programme
Verkkokirja (13736 KB, 115 s.)
Department of Biomedical Engineering and Computational Science publications. A, Report, 21
This thesis deals with object matching, that is, the problem of locating the corresponding points of an object in an image. Conventional approaches to object matching are batch methods, meaning that the methods first learn the object model from a training set of example images that contain instances of the object, and then use the learned object model to match instances of the same object (or object class) in unseen test images. In matching the object in a test image, the visual correspondence of the points as well as their spatial layout is usually considered simultaneously. Typically the corresponding points of the object are manually pre-annotated in the training images which facilitates the learning process. In this thesis, a novel approach is taken: The object matching is incremental. This means that the system is given images one at the time, and the images are matched by updating the object model after each processed image. In addition, the methods presented in this thesis use images as such, without utilizing any pre-annotation or pre-segmentation information, so the task can be considered of being extremely difficult. Although an incremental learning procedure has been presented before that learns an object model incrementally and detects whether the object appears in a test image or not, the proposed methods are the first ones that try to locate the corresponding points of the object by handling images one by one. Like in the traditional object matching methods, the object model of the proposed methods also consists of spatially distributed local features. The adopted approach to the incremental matching is Bayesian; the likelihood corresponds to the appearance of the features and the prior distribution to the spatial layout. Recursive Bayesian formulas are utilized in updating the object model after each processed image and particle Monte Carlo methods are used to sample the posterior distribution. Results show that the methods are able to locate the corresponding points with similar accuracy as the batch matching methods. The learned object model can also be used in detection tasks, and according to the results, the detection capabilities of the presented methods are (almost) on a par with the batch detection methods, and superior to the reference incremental detection method.
Supervising professor
Lampinen, Jouko, Prof.
Thesis advisor
Lampinen, Jouko, Prof.
object matching, incremental learning, Bayesian inference, Monte Carlo methods
Other note