Context and observation driven latent variable model for human pose estimation