On the Earth Mover's Distance as a histogram similarity metric for image retrieval

In this paper, the performance of the Earth Mover's Distance (EMD) vs /spl chi//sup 2/ distance as histogram similarity metrics for image retrieval is evaluated experimentally. Ground truth is generated in two ways: through a novel approach of extracting frames from video shots and through random sampling. Quantitative evaluations of the retrieval performance with regular partitioning, clustered and adaptive binning histograms and with various number of bins are given. Meanwhile the influence of ground distance functions to the performance of the EMD is also studied. The findings of the paper may constitute guidelines in adopting the EMD.