Training details on semantic segmentation benchmark of EntitySeg

I have a doubt related to how you have trained the model for semantic segmentation. Have you used the images high res or low res, since the annotations for this kind of data are coherent with images_lr?
