Misalignment of the orientations between a 360 camera and the scene results in a wavy and distorted spherical panorama image, which may look unstable and have poor perceptual quality. To automatically correct such mis-oriented 360 panoramas, this paper proposes a novel upright adjustment framework based on a convolutional neural network. Instead of directly predicting the 3D rotation of the camera on a given panorama image, our method estimates the rotation by analyzing the projected 2D rotations of multiple images sampled from the panorama. To accurately estimate the rotations of 2D sampled images, we train a 2D roll estimation network using a large-scale labeled image dataset generated by cropping 360 spherical panoramas with various view orientations. Experimental results demonstrate that the proposed method accurately and robustly handles upright adjustment of rotated panoramas while outperforming the previous methods on test datasets that consist of a variety of scenes.
展开▼
