Extracting multi-size local descriptors by GPU computing

This paper presents fast computational techniques for extracting local descriptors from multiple local regions associated with an image feature such as a feature point. Multiple local regions with different sizes are detected by multiplying multiple scale factors to the characteristic scale of the image feature. The descriptors obtained from multiple local regions are called multi-size local descriptors. Multi-size local descriptors enable us to use various types of feature representation and matching schemes based on many different spatial sizes, which is a promising way to control the balance among the robustness against for occlusions, the invariance, and the distinctiveness of the descriptors to the contents of scenes. Because multi-size local descriptors increases the computational costs of feature extraction, we introduce parallel computational techniques for extracting the multi-size local descriptors consisting of the histograms of gradient orientations through the use of a graphics processing unit (GPU). In particular, we demonstrate that orientation maps are useful for efficient extraction of the multi-size local descriptors. Using orientation maps, we can calculate the descriptors by a table look-up manner. We show implementation details and then conclude with the experimental results that demonstrate the usefulness of GPU computing with orientation maps.