search

naver / r2d2

Other
461 stars 86 forks source link

question about two losses #33

Open Jemmagu opened 3 years ago

Jemmagu commented 3 years ago

Hi @jerome-revaud , I'm quite interested in your two losses. You have tested the aachen localization result with limited keypoints, the result decreases by 3~6% (10k kpts to 2k), it seems acceptable. And i'm curious is peaky loss works so that you can using limited kpts? maybe because peaky loss ensures discriminativeness in a patch? if i wanna localize with limited kpts, do you have any further suggestions/strategies that i can use?

thanks in advance!

Jemmagu commented 3 years ago

another question is that, i added your repeatability loss in my training pipeline, and this loss doesn't decrease, but the training result do improves. quite confused why that happened...