Abstract
The purpose of cross-view image matching is to match images acquired from the different platforms of the same target scene and then help positioning system to infer the location of the target scene. With the rapid development of drone technology, how to help Drone positioning or navigation through cross-view matching technology has become a challenging research topic. However, the accuracy of current cross-view matching models is still low, mainly because the existing public datasets do not include the differences in images obtained by drones at different heights, and the types of scenes are relatively homogeneous, which makes the models unable to adapt to complex and changing scenes. We propose a new cross-view dataset, SUES-200, to address these issues.SUES-200 contains images acquired by the drone at four flight heights and the corresponding satellite view images under the same target scene. To our knowledge, SUES-200 is the first dataset that considers the differences generated by aerial photography of drones at different flight heights. In addition, we build a pipeline for efficient training testing and evaluation of cross-view matching models. Then, we comprehensively evaluate the performance of feature extractors with different CNN architectures on SUES-200 through an evaluation system for cross-view matching models and propose a robust baseline model. The experimental results show that SUES-200 can help the model learn features with high discrimination at different heights. Evaluating indicators of the matching system improves as the drone flight height gets higher because the drone camera pose and the surrounding environment have less influence on aerial photography.
Abstract (translated)
URL
https://arxiv.org/abs/2204.10704
