Improvement of 360-degree Depth Estimation Network through Projection Method Modification

In depth estimation for 360-degree images, it is common to unfold the 360-degree images into planar images using the projection methods such as Equirectangular Projection (ERP) and Cubemap Projection (CMP) before estimating the depth. The projected images using each projection method exhibit unique distortions, which complicate depth estimation. To address this, some methods combine both projection methods to mitigate the effects of both distortions. However, in such methods, to realize the high depth estimation performance, an additional image processing (Spherical Padding) is required, which need to pay high computational cost. To solve this problem, in this research, we propose a novel depth estimation method with Gnomonic Projection (GP), which does not require the additional image processing and provides the projected images having less distortions compared to CMP. In the proposed method, we employ GP instead of CMP to reduce the computational cost and combine ERP and GP to improve the depth estimation performance. Extensive experiments demonstrate that our proposed method not only achieves higher depth estimation accuracy but also delivers faster estimation times.