Supervised Earth Mover's Distance Learning and Its Computer Vision Applications

The Earth Mover's Distance (EMD) is an intuitive and natural distance metric for comparing two histograms or probability distributions. It provides a distance value as well as a flow-network indicating how the probability mass is optimally transported between the bins. In traditional EMD, the ground distance between the bins is pre-defined. Instead, we propose to jointly optimize the ground distance matrix and the EMD flow-network based on a partial ordering of histogram distances in an optimization framework. Our method is further extended to accept information from general labeled pairs. The trained ground distance better reflects the cross-bin relationships, hence produces more accurate EMD values and flow-networks. Two computer vision applications are used to demonstrate the effectiveness of the algorithm: first, we apply the optimized EMD value to face verification, and achieve state-of-the-art performance on the PubFig and the LFW data sets; second, the learned EMD flow-network is used to analyze face attribute changes, obtaining consistent paths that demonstrate intuitive transitions on certain facial attributes.