Spatial consistency based selective reranking for content based object retrieval

Reranking is one of the commonly used methods to improve the initial ranking performance for content based object retrieval. In this paper, we propose a spatial consistency based selective reranking method to boost the performance of traditional reranking. After deriving the query's top results, we measure the spatial consistency degree of each query-result image pair via visual words spatial verification. Then the images which are most consistent to query image are selected as positive examples for current query images, and utilized in pseudo relevance feedback reranking. On the other hand, for queries whose top results are less consistent to the query images, they are excluded from reranking to avoid unnecessary performance deterioration. Experimental results on two datasets demonstrate the effectiveness of our method in object retrieval applications.