Using this approach, we have created qualitatively correct 3-d models for 64.9% of 588 images downloaded from the internet, as compared to Hoiem et al.'s performance of 33.1%. Further, our models are quantitatively more accurate than either Saxena et al.