This is an unoffical implementation of CVPR 2019 accepted paper "SOSNet: Second Order Similarity Regularization for Local Descriptor Learning" , and because of the training process was not public, I will try to follow with my understanding and use it in our dataset.
Our goal is to verfiy the ability of "cross source Matching" of SOSNet in Remote sensing situation.
dataset introduction: training data Collection-10000 is a synthetic dataset with nearly 10,000 image from Google Earth Software.
test dataset named "Collection_1" (aka Collection-850) contain 850 synthetic images. test dataset named "NWPU-ChangAn" is the real-world dataset with nearly 150 photo.
we test there ability of matching with historical satellite image.