Gravity Alignment for Single Panorama Depth Inference
Matheus A. Bergmann, Paulo G. L. Pinto, Thiago L. T. da Silveira, Cláudio R. Jung · 2021
Monocular depth inference methods based on 360° images allow 3D reconstruction of entire rooms with a single capture. However, most state-of-the-art approaches assume gravity-aligned images and are highly sensitive to camera rotations. Such limitations result in poor depth estimates, which may jeopardize further 3D-based applications. Here, we present a pipeline for spherical single-image depth inference supplied by a novel rotation correction module. We show that our gravity alignment module can improve existing single-image depth estimation methods, being also useful for aligning color and depth to the horizon, which is highly desirable in many applications.