Group Burstiness Weighting for Image Retrieval

In Bag-of-Word based image retrieval, burst phenomenon is a common issue and should be carefully addressed for improving retrieval accuracy. Current state-of-the-art solutions, e.g., the intra- and inter- image burstiness weighting methods, ignore burstiness problem in query image. In this paper, a group burstiness weighting approach is proposed to address this issue by introducing penalties to burst features of query image. Specifically, burst features are detected at query side such that different groups consisting of burst features can be determined. Then, penalties are imposed on the detected burst features when computing images similarity. It is worthwhile to highlight that the proposed approach is compatible with current burstiness processing methods and effective to improve their performance for image retrieval. Experimental results over several public datasets demonstrate that the proposed approach can well fit for existing burstiness processing methods and significantly improve the performance of image retrieval in terms of accuracy, especially for retrieving landmark images.