Context based re-ranking for object retrieval

Abstract

We propose a simple but effective re-ranking method for improving the results of object retrieval. Our method considers the contextual information embedded in a dataset. This is based on the observation that if there are multiple images containing the same object in a dataset, then these images can often be grouped into clusters. We make the following two contributions. Firstly, we gain this contextual information by a random dimension partition of the dataset. This enables online query model expansion if needed. Secondly, we use the collected contextual information to refine the initial retrieval results by taking into account the context in which each retrieved image occurs. Experimental results on several datasets demonstrate the effectiveness of our method in both accuracy and computation cost: our method refines retrieval results without relying on low-level feature matching or re-issuing the query.