Generalization on RGB Image

[r-deo]

Hi,

We are trying to replicate this use case on 3 channel RGB image. We have sampled the images from US geography. However, when I trained the model, 50% predictions are not correct (More than 500 Km apart).

I am using moco resnet trained on three channel image.

Do I need to make any modification in the Position encoder part to constrain the predictions to US geography?

Thanks,
Rahul

[konstantinklemmer]

Hi @r-deo and sorry for the delay! If you constrain the area of your data to just the US, you will want to set the $L$ parameter (legendre_polys) in the location encoder higher (i.e. $L>=40$) to reflect smaller scale changes. What scale are you currently using?

[r-deo]

Currently I am running it with L=40. I will try to increase and check the result.