Learning to Estimate Scenes from Images