How to improve the performance of the model in matching key point

[foxkw]

Hi, Thanks for your great job.
I have a question :How to improve the performance of the model in matching similar key point descriptors. like buildings

[wangch-g]

The duplicate patterns in local regions are hard to be matched accurately.
I think you can try to use some matching algorithms, like NN+RANSAC, SuperGlue, etc., to filter the outliers, or improve the architecture of the model (attention mechanism?) to tackle this problem.