A new modal matching algorithm based on dynamically located sub-regions for point feature correspondence

The correspondence problem is often encountered in computer vision. In particular, point feature matching of two images is a critical aspect in: stereo vision, structure-from-motion problems, and moving object tracking in image sequences. Besides, it is one possible approach to define similarity metrics among shapes in content-based retrieval for pictorial databases. This paper proposes an algorithm for reliably matching a large number of point features between two images by means of modal matching, and introduces a new method for the point-matching phase that is the most critical step when dealing with large deformations. The proposed algorithm performs a reliable and exhaustive point-feature association between shapes. The robustness is achieved by limiting risky situations using the principle of locality of associations. The approach does not use any distance based control, therefore it avoids any ad hoc threshold and copes effectively even with very large deformations. Practical experiments have been done for content based image retrieval, and for tracking clouds movements in remote sensing image sequences. In both these applications very promising results have been obtained.