SIFT vs. SOFT — A comparison of feature and correlation based rotation estimation for panoramic images

Orientation estimation based on image data is a key technique in many applications. Robust estimates are possible in case of omnidirectional images due to the large field of view of the camera. Traditionally, techniques based on local image features have been applied to this kind of problem. Another very efficient technique is to formulate the problem in terms of correlation on the sphere and to solve it in Fourier space. While both methods claim to provide accurate and robust estimates, a quantitative comparison has not been reported yet. In this paper we evaluate the two approaches in terms of accuracy, image resolution and robustness to noise by comparing the estimated rotations of virtual as well as real images to ground-truth data.