Feature Detectors and Descriptors for Fisher Vectors

Visual search and classification applications often use local features for image representation and description. Various detectors and descriptors have been developed for extracting these features. The local descriptors can be aggregated into a global image signature for a more compact representation. The global signature can be used in mobile applications where memory and computation time is critical. This paper investigates the suitability of detectors and descriptors for compression by Fisher Vectors. We find that the SURF descriptor is a very good choice as it is faster to compute and outperforms SIFT for a larger number of classes.