Rank diffusion for context-based image retrieval

Abstract

This paper presents an efficient diffusion-based re-ranking approach. The proposed method propagates contextual information defined in terms of top-ranked objects of ranked lists in a diffusion process. That makes the method suitable for large scale real-world collections. Experiments were conducted considering public image collections, several descriptors, and comparisons with state-of-the-art methods. Experimental results demonstrate that the proposed method provides high effectiveness gains with low computational costs.