Speeding up SURF

Abstract

SURF has emerged as one of the more popular feature descriptors and detectors in recent years. While considerably faster than SIFT, it is still considered too computationally expensive for many applications. In this paper, several algorithmic changes are proposed to create two new SURF like descriptors and a SURF like feature detector. The proposed changes have comparable stability to the reference implementation, yet a byte code implementation is able run several times faster than the native reference implementation and faster than all other open source implementations tested. © 2013 Springer-Verlag.