Scale-Less Feature-Spatial Matching

In this paper, we improve the discriminability of the Scale-Less SIFT (SLS) descriptor, which is constructed without requiring scale estimation of interest points. We thereby avoid to find stable scales which are difficult to obtain in many cases. Scale-Less SIFT descriptors of interest points are represented as sets of SIFT descriptors at multiple scales. We construct the linear subspace as the geometric representation for sets of SIFT descriptors. Then an embedding representation is learned that combines the descriptor similarity across scales and the spatial arrangement in a unified Euclidean embedding space. The learned subspace are highly capable of capturing the scale-varying values of SIFT descriptors. Experiment results demonstrate significant improvements by our constructed descriptors over existing methods on standard benchmark datasets.