QVPR / Patch-NetVLAD

Code for the CVPR2021 paper "Patch-NetVLAD: Multi-Scale Fusion of Locally-Global Descriptors for Place Recognition"
MIT License
517 stars 72 forks source link

About score of the similarity #7

Closed noah003 closed 3 years ago

noah003 commented 3 years ago

hi, It looks like the score is not normalized, how to distinguish between larger and lower scores

StephenHausler commented 3 years ago

Hi noah003,

The similarity score is normalized, please see https://github.com/QVPR/Patch-NetVLAD/blob/051127d03f347e5dec72e3368e06d823ad906c51/patchnetvlad/models/local_matcher.py#L167 and https://github.com/QVPR/Patch-NetVLAD/blob/051127d03f347e5dec72e3368e06d823ad906c51/patchnetvlad/models/local_matcher.py#L119-L126 The normalization function ensures that all patch sizes are weighted according to the patch weights configuration and not biased in any other way.

Please let me know if this answers your query.