Self-supervised Indoor 360-Degree Depth Estimation via Structural Regularization

Estimating 360 degrees depth information has attracted considerable attention due to the fast development of emerging 360 degrees cameras. However, most researches only focus on dealing with the distortion of 360 degrees images without considering the geometric information of 360 degrees images, leading to poor performance. In this paper, we conduct to apply indoor structure regularities for self-supervised 360 degrees image depth estimation. Specifically, we carefully design two geometric constraints for efficient model optimization including dominant direction normal constraint and planar consistency depth constraint. The dominant direction normal constraint enables to align the normal of indoor 360 degrees images with the direction of vanishing points. The planar consistency depth constraint is utilized to fit the estimated depth of each pixel by its 3D plane. Hence, incorporating these two geometric constraints can further facilitate the generation of accurate depth results for 360 degrees images. Extensive experiments illustrate that our designed method improves delta(1) by an average of 4.82% compared to state-of-the-art methods on Matterport3D and Stanford2D3D datasets within 3D60.