I don’t understand why use similarity normalization.
If the Euclidean distance of the descriptor is used to measure the similarity, then the similarity is independent of other reference images.So we don’t need similarity normalization, right?
@dd-mike hi,can you help me?
Hi @ZaneRo. Sure, can you tell me a bit more about the similarity normalization you’re referring to?
Is this something in the baseline or the competition paper? Or something we’ve said on the discussion forum? Please point me to a specific reference.
It’s in the competition paper: “The 2021 Image Similarity Dataset and Challenge”
@dd-mike hi, above is my reply.
Got it, thanks for clarifying, @ZaneRo.
There is no requirement to use similarity normalization. The baseline models described in the competition paper use similarity normalization to improve the accuracy of the image pair scores, but you don’t have to do the same.
It sounds like there might be some confusion about why similarity normalization is being used. Similarity normalization is not intended to ensure independence from other reference images. But in applying similarity normalization, the baseline models must ensure that they are still following competition rules, and they do this by computing the normalization factor using the training set.
I hope this helps.
ok,it helps. Thank you!