An efficient and robust rotation invariant descriptor

A novel scheme to calculate feature descriptors with rotation invariance was proposed in this paper. Unlike traditional methods of computing feature orientations and descriptors using a single sampling pattern, we calculate the orientation of a feature point with one sampling pattern and generate its descriptor with another one. Based on this idea, we present a method to generate a rotation invariant descriptor using the sampling patterns of FREAK and BRIEF. The experimental results on image matching show that our method is not only invariant to image rotation but also robust to many other situations such as image blur, viewpoint change, illumination change and JEPG compression. At the same time, the proposed method performs competitively with BRISK and FREAK in computational time.