Real-time Depth Estimation for Aerial Panoramas in Virtual Reality

With the emergence of consumer-level 360° panoramic cameras, omnidirectional RGB images and videos are now easy to be captured either in the hand-held mode or from a drone. Although previous works achieve plausible results in room-sized panoramic datasets, they are limited in indoor scenes with little rotations. In this paper, we present a real-time depth estimation method for more challenging aerial panoramas, where the viewing angle is changing rapidly and the lighting condition is more complicated. Our graph convolutional network(GCN)-based framework makes full use of the global connection information of the omnidirectional images, and is trained with extensive outdoor data. Experiments show that our method is robust to estimate the depth of outdoor aerial panoramas captured from various angles accurately.