Image matching using feature set transformations

The paper presents a novel idea for filtering nearest neighbor feature point pairs in order to yield greater accuracy in the image matching problem. Filtering is based on the alignment of feature points which is achieved by the application of affine transformations on the complete feature set. Using affine transformations, nearest neighbor feature point pairs from different images, which are geometrically dissimilar in the inter feature point geometrical structure in their neighboring regions, are filtered from the nearest neighbor calculations. The feature alignment resulting from the affine transformations is followed by the filtering step which removes outlier nearest neighbor feature pairs. The algorithm makes the assumption that the feature points, for a given object type, maintain their general inter feature point geometrical structure from one image to another. The algorithm can be combined with existing image matching techniques to yield greater accuracy. We show that the algorithm gives good results on known image datasets over 1-NN nearest neighbor based image matching. Furthermore we discuss the extent of increase in the inter image distance due to the filtering of outlier feature pairs.