Approximated Scale Space for Efficient and Accurate SIFT Key-Point Detection

Abstract

The SIFT (scale invariant feature transform) key-point serves as an indispensable role in many computer vision applications. This paper presents an approximation of the SIFT scale space for key-point detection with high efficiency while preserving the accuracy. We build the scale space by repeated averaging filters to approximate the Gaussian filters used in SIFT algorithm. The accuracy of the proposed method is guaranteed by that an image undergoes repeated smoothing with an averaging filter is approximately equivalent to the smoothing with a specified Gaussian filter, which can be proved by the center limit theorem. The efficiency is improved by using integral image to fast compute the averaging filtering. In addition, we also present a method to filter out unstable key-points on the edges. Experimental results demonstrate the proposed method can generate high repeatable key-points quite close to the SIFT with only about one tenth of computational complexity of SIFT, and concurrently the proposed method does outperform many other methods.