Effective Corner Matching P. Smith , D. Sinclair , R. Cipolla and K. Wood Department of Engineering, University of Cambridge, Cambridge, UK Olivetti and Oracle Research Laboratory, Cambridge, UK pas1001@eng.cam.ac.uk Abstract This paper tackles the problem of obtaining a good initial set of corner matches between two images without resorting to any constraints from motion or structure models. Several different matching metrics, both traditional and statistical, are evaluated and the effect of matching using sub-pixel informa- tion is studied. It is found that, in most cases, the commonly-used cross- correlation does not perform as well as some other measures, such as the test or the sum of squared differences, and that it is essential to use sub-pixel accuracy if mismatches are to be avoided. Further, a new technique, the Median Flow Filter, is introduced. This detects outliers by assuming that the image motion is locally similar. Any matches which are in gross disagreement with the local “median flow” are discarded. Experiments show this technique to be particularly effective, typ- ically lowering the percentage of outliers from around 35% to less than 5%, permitting direct model fitting rather than random sampling techniques for any further analysis. 1 Introduction Feature matching is a key component in many computer vision applications, for exam- ple stereo vision, motion tracking, and identification. Feature matches may be sufficient alone, but are also an ideal platform for “bootstrapping” denser and more complex anal- ysis of images. Of all possible features, “corners” are the most widely used; their two- dimensional structure providing the most information about image motion. Feature matching is commonly referred to as the correspondence problem. The prob- lem is how to automatically match corresponding features from two images, while at the same time not assigning matches incorrectly. The common approach for corners, and the one followed in this paper, is to take a small region of pixels (referred to as a correlation window) from around the detected corner and compare this with a similar region from around each of the candidate corners in the other image. Each comparison yields a score, a measure of similarity. The match is assigned to the corner with the highest matching score. The most popular measure of similarity is the cross-correlation (see, for example, [3]). Matching algorithms typically assume that the correlation windows from each image are related by a simple translation, an assumption which is valid in a number of typical ap- plications. In the cases where this is not a valid assumption there are several approaches, the most common of which is to estimate the relationship between the two images and BMVC 1998 doi:10.5244/C.12.55