An Error-Activation-Guided Blind Metric for Stitched Panoramic Image Quality Assessment.

Image stitching is one key enabling component for recent immersive VR technology. The quality of the stitched images greatly affects VR experiences. Evaluation of stitched panoramic images using existing assessment tools is insufficient for two reasons. First, conventional image quality assessment (IQA) metrics are mostly full-referenced, while panorama reference is hard to obtain. Second, existing IQA metrics are not designed to detect and evaluate errors typical in stitched images. In this paper, we design an IQA metric for stitched images, where ghosting and shape inconsistency are the most common visual distortions. Specifically, we first locate the error with a fine-tuned convolutional neural network (CNN), and later refine the locations using an error-activation mapping generated from the network. Each located error is defined by both its size and distortion level. Extensive experiments and comparisons confirm the effectiveness of our metric, and indicate the network's remarkable ability to detect error patterns.