360D: A dataset and baseline for dense depth estimation from 360 images

Abstract

We present a baseline for 360^o dense depth estimation from a single spherical panorama. We circumvent the unavailability of coupled 360^o color and depth image datasets by rendering a high quality 360^o dataset from existing 3D datasets. We then train a CNN designed speci_cally for 360^o content in a supervised manner, in order to predict a 360^o depth map from a single omnidirectional image in equirectangular format. Quantitative and qualitative results show the need for training directly in 360^o instead of relying on traditional 2D CNNs.