Open 666zz666 opened 4 years ago
training use a small size and eval use original image size is a common strategy used in segmentation. In training, we use random crop to augment datasets. So it is not necessary to use very big image size or same with validation. Anyway, you can try to modify the resolution to see if it can get better result.
You use the resolution of 769 769 when training cityscapes but use 1025 2049 when validation。 If there be a gap between the two results