The effects of anisotropic Gaussian diffusion in scale invariant feature detection

Many feature detection algorithms use Gaussian scale space in order to locate scale-invariant and rotationally invariant keypoints in an image, including the Scale-Invariant Feature Transform (SIFT) algorithm. During the creation of this scale space, edge information and fine details in an image are often degraded or lost as a result of the Gaussian smoothing operation. In this paper, we study the effects of using edge preserving anisotropic diffusion during the creation of a scale space for use in the SIFT algorithm. We find that preserving edge information and fine details during the creation of a scale space allows SIFT to gather a much larger set of keypoints from images, and these keypoints tend to be far more robust towards scaling and rotation.