Do we have to remove rgn images before training? I can see rgn images are not included in testing csv and submittion csv files. Why to include rgn images and why not?
I belive the rgn images were given to provide extra information. It's your choice to remove them before training or not(mindful of possible leakage). You can find a way to blend the rgn images into your solution.
Tbh removing the rgn images makes the lb score worse while the local validation is good. I don't know why we are having a bad lb score while removing the rgn and we dont have rgn images in the sample submission. Any explanations?
I released one of the encoders I pretrained during my runs with this particular architecture (xresnet34), linked to in the repo, in case you want to try it out in your own setup.
I belive the rgn images were given to provide extra information. It's your choice to remove them before training or not(mindful of possible leakage). You can find a way to blend the rgn images into your solution.
Tbh removing the rgn images makes the lb score worse while the local validation is good. I don't know why we are having a bad lb score while removing the rgn and we dont have rgn images in the sample submission. Any explanations?
As @TheVeloper said, you can blend them somehow. I thought about this early on when I joined in the competition and came up with this method: https://github.com/poppingtonic/semisupervised-rice-africa
I released one of the encoders I pretrained during my runs with this particular architecture (xresnet34), linked to in the repo, in case you want to try it out in your own setup.