Improved Learning Of Structural Support Vector Machines: Training With Latent Variables And Nonlinear Kernels