A Performance Evaluation of Local Descriptors

A descriptor is computed on a local region around a feature point and is used to characterize and compare the features. Various descriptors have been proposed in the literature which have different properties and performance in different image data. Evaluation of the local feature descriptor is important to identify the strengths and weaknesses of each algorithm in different applications. In this paper a performance evaluation of the state of the art in local descriptors is performed on a set of satellite images under varying imaging conditions. Ten descriptors are included, which are spin image (SI), shape context (SC), SIFT, PIIFD, SURF, DAISY, LSS, LBP, LIOP and BRISK. 80 satellite image pairs in three groups including simulated images, multi-temporal, and multi sensor images are used as data set and descriptors are evaluated using four evaluation criteria including Recall, Precision, positional accuracy and speed. The evaluation results indicate that there does not exist one descriptor which outperforms the other descriptor for all scene types and all types of transformations, but in average DAISY and SIFT show the best performance.