A Generative Model of Indoor Scenes

In this Master thesis project, we tackle the problem of 3D indoor scene understanding from a single image. Following existing approaches, we parameterize the problem as one of estimating the faces of a 3D cuboid. Towards this goal, we propose a generative model of the scene, which takes advantage of sophisticated image features while utilizing a simple prior learned from the statistics of room photographs. We utilize an adaptive Metropolis-Hastings sampling scheme for inference and demonstrate that our approach outperforms the state-of-the-art on the main two benchmarks that exist for this task. Moreover, our approach outperforms previously published generative models for this problem by more than 10% absolute error.