SIFT on manifold: An intrinsic description

Scale Invariant Feature Transform is a widely used image descriptor, which is distinctive and robust in real-world applications. However, the high dimensionality of this descriptor causes computational inefficiency when there are a large number of points to be processed. This problem has led to several attempts at developing more compact SIFT-like descriptors, which are suitable for faster matching while still retaining their outstanding performance. This paper focuses on the SIFT descriptor and explore a dimensionality reduction for its local representation. By using the manifold learning algorithm of Locality Preserving Projections, a more effective and efficient descriptor LPP-SIFT can be obtained. A large number of experiments have been carried out to demonstrate the effectiveness of LPP-SIFT. Besides, the practicability of LPP-SIFT is also shown in another set of experiments for image similarity measurement.