Neural Contourlet Network for Monocular 360° Depth Estimation

For a monocular 360° image, depth estimation is challenging because the distortion increases along latitude. To perceive distortion, existing methods devote to designing deep and complex network architecture. In this paper, we provide new perspective that constructs an interpretable sparse representation for image. Considering importance of geometric structure in estimation, utilize contourlet transform capture explicit cue spectral domain integrate it with implicit spatial domain. Specifically, propose neural consisting convolutional branch. encoder stage, design spatial–spectral fusion module effectively fuse two types cues. Contrary encoder, employ inverse learned low-pass subbands band-pass directional compose decoder. Experiments on three popular panoramic image datasets demonstrate proposed approach outperforms state-of-the-art schemes faster convergence. Code available at https://github.com/zhijieshen-bjtu/Neural-Contourlet-Network-for-MODE .