Repeatability Performance Evaluation of Local Feature Detectors on Pyramid Images

摘要

To generate corresponding points for large size satellite, aerial or high resolution close-range image, it is always needed to downscale the original image to low resolution image via pyramid image generating to detect local features. So this paper studies the affection of the pyramid image on the repeatability performance of the state-of-the-art local feature detectors. For generating pyramid image with ground truth data, a method for calculating the homography parameters between pyramid images of stereo pair with homography parameters is proposed. Benchmark data sets with ground truth are used to generate pyramid images and derive homography parameters directly. Experiment results revealed that pyramid image can reduce the affection of scale change and image distortion to some extent.