Affine invariant salient patch descriptors for image retrieval

Abstract

Image description constitutes a major part of matching-based tasks in computer vision. The size of descriptors becomes more important for retrieval tasks in large datasets. In this paper, we propose a compact and robust image description algorithm for image retrieval, which consists of three main stages: salient patch extraction, affine invariant feature computation over concentric elliptical tracks on the patch, and global feature incorporation. We evaluate the performance of our algorithm for region-based image retrieval and image reuse detection, a special case of image retrieval. We present a novel synthetic image reuse dataset, which is generated by superimposing objects on different background images with systematic transformations. Our results show that the proposed descriptor is effective for this problem.