Feature-based Approaches to Multi-View Image Stitching

Image stitching, the technique of merging multiple images with overlapping regions, has various applications in real-life scenarios such as security systems, autonomous vehicles, and sports analytics. Several image stitching methodologies have been developed, resulting in a wide range of traditional feature-based and deep learning approaches. In this paper, we utilise the Adobe Panoramic Dataset to assess eight approaches for stitching four overlapping images. We develop six novel feature-based stitching approaches, including AKAZE-SIFT-BF KNN-RANSAC. As an example approach, KAZE-SIFT-BF KNN-RANSAC utilises Accelerated KAZE (AKAZE) for feature detection, Scale Invariant Feature Transformation (SIFT) to generate feature descriptors, Brute Force with k-Nearest Neighbours and a ratio test for keypoint matching, and Random Sample Consensus (RANSAC) for matched keypoint outlier removal. Evaluation is performed using the CombSIQE metric and qualitative comparison to evaluate the quality of the resulting stitched images. Performance evaluation shows that our proposed approaches produce comparable quality stitched images to the optimised SOTA SuperPoint and SuperGlue approach in outdoor scenes, while remaining consistent in indoor scenes.