Unsupervised Effectiveness Estimation for Image Retrieval Using Reciprocal Rank Information

In this paper, we present an unsupervised approach for estimating the effectiveness of image retrieval results obtained for a given query. The proposed approach does not require any training procedure and the computational efforts needed are very low, since only the top-k results are analyzed. In addition, we also discuss the use of the unsupervised measures in two novel rank aggregation methods, which assign weights to ranked lists according to their effectiveness estimation. An experimental evaluation was conducted considering different datasets and various image descriptors. Experimental results demonstrate the capacity of the proposed measures in correctly estimating the effectiveness of different queries in an unsupervised manner. The linear correlation between the proposed and widely used effectiveness evaluation measures achieves scores up to 0.86 for some descriptors.