search

naver / r2d2

Other
461 stars 86 forks source link

Reliability loss converges to 0.5 instead of 0. #16

Closed MarviB16 closed 4 years ago

MarviB16 commented 4 years ago

Hello,

I am using your proposed framework to match historical and present satellite imagery. However, the reliability loss does not converge to 0, but to 0.5. Which results in an AP of 0 and a reliability score of 0. I would guess that is due to the initial AP being too low.

Did you experience any similar behavior while testing? Or do you have an idea why this could happen?

jrevaud commented 4 years ago

Hi

No, I've never seen this behavior. As you said, this is most likely due to the fact the the matching problem is harder on your data than on our data.

What you can do is change the value of \kappa in the reliability loss. This represents the threshold above which an AP is considered to be "good" -- by default it is equal to 0.5 (see ReliabilityLoss(base=0.5)). So you can try to lower it to 0.1 or even below, and see if it helps for the convergence.

Another solution would be to 1- step 1: train without the reliability (ie using PixelAPLoss() instead of ReliabilityLoss()) 2- step 2: continue training with the reliability

Chjx395 commented 1 year ago

Hello,

I am using your proposed framework to match historical and present satellite imagery. However, the reliability loss does not converge to 0, but to 0.5. Which results in an AP of 0 and a reliability score of 0. I would guess that is due to the initial AP being too low.

Did you experience any similar behavior while testing? Or do you have an idea why this could happen?

Hi there,

I have the same problem. Have you solved the problem?