SP-SIFT: enhancing SIFT discrimination via super-pixel-based foreground-background segregation

Existing point-of-interest (POI) descriptions are biased by the information surrounding the point. Whereas in self-contained images this information is useful for enhancing the repeatability of the description, its use is inadequate for the description of objects that might be surrounded by variable backgrounds. To tackle these situations, a new POI descriptor - super-pixel-based isolation of the scale invariant feature transform (SP-SIFT) - is proposed. The classical SIFT descriptor is modified by isolating the information of the flat areas that compose it. It is proposed to include superpixel information in the description stage of the SIFT. The obtained results suggest that a so-built descriptor increases the repeatability of SIFT points in these scenarios while keeping its robustness to global transformations of the image: blurring, changes in viewpoint, scale and lighting. The method is presented here as an extension of the SIFT. However, the idea behind it may be easily exported to most of the existing POI-descriptors in the state-of-the-art.