$\mathcal {S}^{2}$Net: Accurate Panorama Depth Estimation on Spherical Surface

Monocular depth estimation is an ambiguous problem, thus global structural cues play important role in current data-driven single-view methods. Panorama images capture the complete spatial information of their surroundings utilizing equirectangular projection which introduces large distortion. This requires method to be able handle distortion and extract context from image. In this paper, we propose end-to-end deep network for monocular panorama on a unit spherical surface. Specifically, project feature maps extracted onto surface sampled by uniformly distributed grids, where decoder can aggregate distortion-reduced maps. Meanwhile, cross-attention-based fusion module fuse skip connection enhance ability obtain context. Experiments are conducted five datasets, results demonstrate that proposed substantially outperforms previous state-of-the-art All related codes will open-sourced upcoming days.