Improved CenSurE feature and a new rapid descriptor of GSIP

Abstract

Image matching is a fundamental aspect of many problems in computer vision applications. Many of these applications have real-time constraints. Recently there has been much interest in object and view matching using local invariant features and descriptor. Center surround extrema (CenSurE) features have aroused wide interest because of the high compute efficiency. Due to its linear scales, the filter response signal is very sparse, can't acquire high repeatability. So, we modified the center-surround detectors (CenSurE) using logarithmic scale sampling, and to detect the features of image interpolation, obtained the subpixel level of accuracy. The results of image matching experiments and theory analyzing demonstrate that improved center-surround detectors are more stable and repeatable than CenSurE. At the same time, we proposed a new rapid descriptor based on gradient of the summed image patch, called GSIP. The results of image region matching experiments demonstrate that the GSIP descriptor has better distinctiveness than the state-of-the-art SURF descriptor, and achieves a 2 fold speed increase. The local feature and descriptor present here can be used widely in real-time view matching fields, including motion tracking, solving for 3D structure from multiple images, and robots visual navigation.