Any reason for dual softmax matcher for inference?

[ducha-aiki]

Hi,

I am wondering, can we just use mutual nearest neighbor matching for speed at inference time? Dual softmax normalizes the scores, but doesn't change the actual neighbors (matches).

[Parskatt]

We have a confidence threshold for the matches based on the ds scores. If you remove that youll get more outliers.

Otherwise yes, I see no major issue.