Improved Locality affine-invariant feature matching

Abstract

It is a fundamental and important task in computer vision to find a reliable correspondence between two feature sets. However, due to the limitation of descriptors, the raw matches are often polluted by most outliers, and the matching results given by the algorithm may be cross-exchanged in the images with similar features. Based on the idea of local affine invariance, this paper divides the image into several local sub-feature sets, and the sub-feature sets often satisfy affine transformation. Then, the local barycentric coordinates of the seed points of the subset were checked to eliminate the mismatching of similar features. In order to prove the versatility of our strategy for dealing with image matching problems, we conducted extensive experiments on various actual image pairs, especially for images with similar features, we established an additional test set. The test results show that our method has similar performance on the common test set compared with other methods in terms of efficiency and effectiveness, and is more competitive in similar feature images.