Evaluation of Local Features for Near-Uniform Scene Images

Image stitching requires accurate matching of visual features to achieve good alignment. However, feature-based matching often has poor result particularly when image content is fairly near-uniform and thus it remains a challenging problem to be addressed. When the current state-of-the-art feature detectors unable to detect sufficient reliable corresponding keypoints, the output stitched images often suffer from misalignment, projective distortion and visible artefact. This paper presents a new experimental evaluation using especially near-uniform images for the performance of some well-known feature detectors, such as Harris, SIFT, SURF, BRISK and KAZE. In addition, we have also introduced RC/S o score to compare spatial distribution of the correct matched keypoints in overlapping region between images. The results show that the best performed local feature detector is KAZE. However, none of the tested feature detectors can reach more than 50% spread of the overlapping region.