Multi-stage image retrieval based on feature augmentation with truncated polynomial weight

Keundong Lee, Seungjae Lee, Wonyoung Yoo

Abstract: In this paper, we propose an effective image retrieval method. Based on a conventional global image representation, multi-stage image retrieval pipeline with feature augmentation is constructed to improve retrieval accuracy. To suppress irrelevant images while boosting relevant images, a novel weighting scheme for feature augmentation is introduced. In addition, the relationship between database images is leveraged to update or re-rank shortlist of the retrieved images. The proposed method was evaluated on Google-landmarks dataset, and the experimental results validate the effectiveness of the proposed method.