Feature Alignment Method for Cross-View Image Geo-localization

Cross-view image geo-localization aims to estimate the geo-location of a ground-view image, via finding the matching geo-tagged satellite images in a reference satellite database. We observe that the spatial relative layouts and contents are different for two images, captured in the same location from different initial orientations. Especially, when the orientations of the ground view images and the satellite images for the same geo-location are not aligned, cross-view image matching methods will be dramatically affected. Therefore, this paper proposes a feature alignment method to align the spatial layouts of cross-view image pairs. The method mainly has two parts: the coarse-grained alignment based on local feature segmentations and the fine-grained alignment based on global features. Coarse-grained alignment is to segment the ground view image features and the satellite image features separately in the shallow part of the convolutional networks, rearrange each satellite sub-image block by using the similarity between local features, and then concatenate them to achieve first alignment. Fine-grained alignment is using global ground view image features to search matching features in the widened satellite global feature maps by sliding windows, and obtaining feature alignments for cross-view images. Our method alleviates the spatial layout differences between cross-view image pairs by feature alignment, and the experimental results on two datasets show that the proposed method outperforms the state-of-the-art algorithms for ground view images captured in different orientations.