Road Extraction Assisted Offset Regression Method in Cross-view Image-based Geo-localization

Cross-view image-based geo-localization aims to determine the location of ground-view images in satellite images. Most researches regard it as image retrieval problem. However, they don't provide exact location. In this paper, we consider it as a regression problem: given a ground-view image and satellite image, a siamese network finds the location offset between them. The proposed siamese network with fully connected layers can get the offset of satellite center with a ground-view image. Meanwhile, by training road extraction with Dice and Binary Cross Entropy loss, the network can perceive road location and get more accurate result. The experiment of our method gets 16.50 meters mean distance error, 28% less than VIGOR [1] on its dataset.