UniFuse: Unidirectional Fusion for 360° Panorama Depth Estimation

Learning depth from spherical panoramas is becoming a popular research topic because panorama has full field-of-view of the environment and provides relatively complete description scene. However, applying well-studied CNNs for perspective images to standard representation panoramas, i.e., equirectangular projection, suboptimal, as it becomes distorted towards poles. Another cubemap which distortion-free but discontinued on edges limited in field-of-view. This letter introduces new framework fuse features two projections, unidirectionally feeding only at decoding stage. Unlike recent bidirectional fusion approach operating both encoding stages, our scheme much more efficient. Besides, we also designed effective module scheme. Experiments verify effectiveness proposed strategy module, model achieves state-of-the-art performance four datasets. Additional experiments show that advantages complexity generalization capability.