wangwenhao0716 / isc-track2-submission Goto Github PK

Hello! Congratulations with the best results and thank you for such a good and well-documented solution!

I have a question regarding score normalization. In the paper, you wrote:

scores of different queries are not comparable.

Why are matching confidences not comparable between queries and why do we need score normalization?

My best guess is that intra-cluster variability is different for different objects in our hyperspace.
For example, two classes: images of the sky and images of a human face. The first cluster is "more packed" and has less variability. For a query with a face, (cosine) distance of 0.8 might indicate that we have an exact match, while for the other class with less variability the same distance is too low for an exact match.

如题～