Supplementary Materials for Surface Normals in the Wild

Here we provide additional details on the choice of parameters used in our experiments. For all NYU images, we use the one ground truth focal length shared by all images. For all SNOW images, we assume the same focal length as NYU. Focal length is not needed to predict depth at test time. In training, for SNOW images focal length is often available in EXIF, which could be used to further improve our results. The parameter λ in Eq.(1) of the paper is selected by performing hyper-parameter optimization on a held-out validation set. They are listed in Tab. 1 The parameter τ in Eq.(3) of the paper equals ln(1.02). It is chosen to be consistent with the threshold used in [4] to generate the ground truth relative depth. We use the same τ for all experiments including those on SNOW.