Object retrieval using adaptive rectangular windows

In this paper, we propose adaptive rectangular windows to improve object retrieval. The adaptive rectangular windows are created by G-means clusters, and solve the problem that G-means splits the database image into small irregular regions and decreases the retrieval accuracy. According to the spatial information in the clusters and the relative spatial relationship between the clusters, the rectangular windows adjust adaptively and represent the object regions with more spatial information. Afterwards, each adaptive rectangular window corresponds to an independent window vector, which increases the similarity between query object region and object regions of database images. The experimental results on the Oxford building dataset demonstrate that the adaptive rectangular windows improve the retrieval accuracy while ensuring the retrieval efficiency.