Unsupervised alignment of objects in images

With the advent of computer vision, various applications become interested to apply it to interpret the 3D and 2D scenes. The main core of computer vision is visual object detection which deals with detecting and representing objects in the image. Visual object detection requires to learn a model of each class type (e.g. car, cat) to be capable to detect objects belonging to the same class. Class learning benefits from a method which automatically aligns class examples making learning more straightforward. The objective of this thesis is to further develop the sate-of-the-art feature-based alignment method which rigidly and automatically aligns object class images to a manually selected seed image. We try to compensate the weakness by providing a method to automatically select the best seed from dataset. Our method first extracts features by utilizing dense sampling method and then scale invariant feature transform (SIFT) descriptor is used to find best matches as initial local feature matches. The final alignment is based on spatial scoring procedure where the initial matches are refined to a set of spatially verified matches. The spatial score is used next to calculate similarity scores. We propose an algorithm which operates on spatial and similarity scores and finally selects the best seed. We also investigate the performance of step-wise alignment using minimum spanning tree (MST) and Dijkstra shortest path instead of direct alignment utilizing a single seed. We conduct our experiments using classes of Caltech-101 for which our unsupervised seed selection and step-wise alignment achieve state-of-the-art performance