Panorama Inspector: A Novel Methodology for Automated Error Detection in Stitched Images

Obtaining a panorama image by combining various images is called image stitching. Although improvements in stitching methods continue, and each new model causes fewer stitching errors, errors occur even when the state-of-the-art models are used, especially in samples taken in the real world with few key points, such as empty apartments. It is crucial to detect errors when they occur, specifically in systems like ours, where the result created by the stitching system is shown in a real-world property listing platform. In this way, it is possible to avoid presenting unsatisfying results, which contain errors, to the user. The novel and automated error detection method proposed in this paper detects whether an error has occurred, the location of the error within the panorama, and the magnitude of the area affected by the error with a clustering-based approach at the seam boundaries. The proposed approach achieved 86% precision, 85 % recall, and 85 % Fl score when evaluated on a real-world dataset. Thanks to our proposed approach, only the panorama images error-free can be delivered to the users, and the quality of the panorama images shown in the real-world platform can be kept above a certain level.