Boosting the Performance of Generic Deep Neural Network Frameworks with Log-supermodular CRFs