Studying the compression performance of video descriptors

The main objective of this paper is to study the per- formance of a framework for encoding visual feature descriptor s. Local visual feature descriptors are employed in a number of computer vision tasks, e.g. image and video retrieval by visual search, object recognition and automatic annotation. In scenar ios strictly constrained in terms of storage capability, memory and network resources such as those observed in visual sensor networks and mobile visual search applications, compression may be imperative. We evaluate coding schemes for the two most used feature descriptors, namely Scale Invariant Feature Transform (SIFT) and Speeded Up Robust Features (SURF). The coding modes include intra- and inter-frame modes, with and without decorrelating transforms. They are tested in descriptors extra cted from video sequences with different content characteristics. A detailed rate-distortion analysis is conducted in order to assess the contribution of each coding mode. Also, is shown that rate- distortion optimization with all coding mode enabled leads to best results.