Multi-image matching using invariant features

This thesis concerns the problems of automatic image stitching and 3D modelling from multiple views. These are basic problems of computer vision, with applications in robotics, architecture, industrial inspection, surveillance, computer graphics and film. Recent work has brought increasing automation to these tasks, but despite a large amount of progress, state-of-the-art algorithms still require some form of user input or assumptions about the image sequence. For example, the best image stitchers currently require an ordered set of input images, or user input to identify the matching images, before automatic registration can proceed. In this work we show how such tasks can be performed automatically and without any user input at all. We formulate the multi-image matching problem as one of finding all matching images, subject to the constraint that they are consistent views from a perspective camera. We use invariant features as a mechanism for finding correspondences, and indexing techniques to efficiently find matches between multiple views. We then find all sets of geometrically consistent feature matches, using a probabilistic model for verification. This allows us to identify each object or scene in the dataset using only the structure already present in the data. The major contributions of this thesis are the development of a system that can automatically recognise and stitch 2D panoramas in unordered image datasets, and a new class of invariant features for this purpose.