Query Adaptive Fusion for Graph-Based Visual Reranking

Developing effective fusion schemes for multiple feature types has always been a hot issue in content-based image retrieval. In this paper, we propose a novel method for graph-based visual reranking, which addresses two major limitations in existing methods. First, in the phase of graph construction, our method introduces fine-grained measurements for image relations, by assigning the edge weights using normalized similarity. Furthermore, in the phase of graph fusion, rather than summing up all the graphs for different single features indiscriminately, we propose to estimate the reliability of each feature through a statistical model, and selectively fuse the single graphs via query-adaptive fusion weights. Fusion methods with either labeled data and unlabeled data are proposed and the performance are evaluated and compared by experiments. Our method is evaluated on five public datasets, by fusing scale-invariant feature transform (SIFT), CNN, and hue, saturation, hue (HSV), three complementary features. Experimental results demonstrate the effectiveness of the proposed method, which yields superior results than the competing methods.