OIDC-Net: Omnidirectional Image Distortion Correction via Coarse-to-fine Region Attention

Omnidirectional cameras have recently received significant attention in panoramic imaging systems such as virtual reality (VR) technology; however, the strong geometric distortion in omnidirectional images severely affects the object recognition and semantic understanding. In this paper, we propose an automatic omnidirectional image distortion correction approach powered by a unified learning model (OIDC-Net). This approach is applicable for almost all types of omnidirectional cameras, requiring nothing more than a distorted image. A crucial and challenging ingredient for reconstructing the real physical scene is to estimate the heterogeneous distortion coefficients in an appropriate camera model. To address this issue, we present a novel coarse-to-fine region attention mechanism to alleviate the difficulty of predicting all coefficients simultaneously. With the proposed cascade structure and deep fusion strategy, the ambiguous relationship among these heterogeneous distortion coefficients has been incrementally perceived. Our experimental results show significant improvement over the state-of-the-art methods in terms of visual appearance, while maintaining a promising quantitative performance.