Reviews: SURGE: Surface Regularized Geometry Estimation from a Single Image

The paper proposes a method for recovering scene geometry from a single RGB image. This method uses a dense CRF with terms that enforce consistency between point-wise depth and normal estimates, using regularizers based on classification of planarity and presence of depth boundaries. Each of these estimates (depth, normal, planarity, edges) comes from a separate network proposed for each task in prior work. In addition to the geometry-terms in proposed DCRF-based model, the paper's contributions include using multiple passes through the depth and normal networks with dropout to derive'confidence' values of these metrics, and joint training to fine tune the depth and normal networks. While significantly engineered for its specific application domain, the paper does demonstrate a successful example of inference with a regularized objective, where different terms are predicted from trained neural networks.