$\theta$-MRF: Capturing Spatial and Semantic Structure in the Parameters for Scene Understanding