discussion about difference between re-rank and qb-norm

ioanacroi / qb-norm

Cross Modal Retrieval with Querybank Normalisation

https://vladbogo.github.io/QB-Norm/

MIT License

52 stars 2 forks source link

discussion about difference between re-rank and qb-norm #5

Closed onlyonewater closed 1 year ago

onlyonewater commented 1 year ago

Hi, I notice that re-rank [1] as a post-processed technology is very popular in cross-modal retrieval, I think qb-norm is also a re-rank technology. Its implementation is a little different from re-rank. This is my understanding, is it right?

[1] Re-ranking Person Re-identification with k-reciprocal Encoding. CVPR, 2017

ioanacroi commented 1 year ago

Hi!

There are many works that enhance similarity search for retrieval, including re-ranking. QB-Norm focus specifically on tackling the problem of hubness in cross-modal embeddings. Intuitively, QB-Norm aims to scale down the similarity between a video and a query if that video is highly retrieved. This means that combining QB-Norm with a re-ranking method (such as [1]) could potentially further increase the performance.