Abstract
In recent years, image retrieval has become an important building block for various robot vision systems. Applications include object recognition and place recognition. In those approaches, a set of local feature descriptors such as SIFT (Scale Invariant Feature Transform) are extracted from the visual images and used for indexing and retrieving the view images. A difficulty arises in dealing with a large size image set because the time and spatial cost required increase in proportion to the number of local features per image. To address the scalability issue, we propose a novel database augmentation technique. In it, useless features that have low discriminativity and repeatability arc detected in images and eliminated to reduce time and spatial cost. The proposed database augmentation technique is verified in image retrieval experiments.
