Rethinking Supervised Depth Estimation for 360deg Panoramic Imagery

Lu He, Bing Jian, Yangming Wen, Haichao Zhu, Kelin Liu, Weiwei Feng, Shan Liu; Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, 2022, pp. 5173-5181

Depth estimation from a single 360 panorama image is a difficult task. It is an ill-posed problem to estimate depth maps from an RGB panorama image due to the intrinsic scale ambiguity issue. To mitigate the scale inconsistency issue in the ground truth depth map, we propose a simple yet effective method to normalize the depth data based on estimated camera height. In addition, we design a multiple head planar-guided depth network, to provide more geometric constraints for depth estimation. Experimental results show that our relative depth estimation task is more accurate than the absolute depth estimation task, and our proposed model produces state-of-the-art performance on both Matterport3D and Stanford2D3D datasets.