A Content-Aware Metric for Stitched Panoramic Image Quality Assessment

One key enabling component of immersive VR visual experience is the construction of panoramic images-each stitched into one large wide-angle image from multiple smaller viewpoint images captured by different cameras. To better evaluate and design stitching algorithms, a lightweight yet accurate quality metric for stitched panoramic images is desirable. In this paper, we design a quality assessment metric specifically for stitched images, where ghosting and structure inconsistency are the most common visual distortions. Specifically, to efficiently capture these distortion types, we fuse a perceptual geometric error metric and a local structure-guided metric into one. For the geometric error, we compute the local variance of optical flow field energy between the distorted and reference images. For the structure-guided metric, we compute intensity and chrominance gradient in highly-structured patches. The two metrics are content-adaptively combined based on the amount of image structures inherent in the 3D scene. Extensive experiments are conducted on our stitched image quality assessment (SIQA) dataset, which contains 408 groups of examples. Results show that the two parts of metrics complement each other, and the fused metric achieves 94.36% precision with the mean subjective opinion. Our SIQA dataset is made publicly available as part of the submission.