MFD: Mutual feature description for image matching

In general, SIFT has been proven to be the most robust local feature descriptor. However, it was designed mainly for gray-level images. Many Objects cannot be distinguished if their color contents are ignored. In addition, SIFT greatly depend on the main orientation assignment, and about one third of orientations are assigned with a big error. Thus many corresponding points are mismatched. This paper addresses these two problems and proposes a local feature descriptor with color and geometric invariance. In this method, the image is transformed into the Gaussian color space to construct the color invariant variable for detecting key points. And then the computation of canonical orientation of key point in the feature descriptor is substituted by the relative gradient direction around the candidate key points which reduces the accumulated orientation error. The pixel gradient and orientation around the extracted key points are computed to build the local gradient orientation histogram to describe the features. The results on the Amsterdam Library of Object Images dataset and 3D objects have shown that the proposed descriptor is better than SIFT and CSIFT.