Hello, first thank you for providing the code !

Could anybody tell me how the augmentation used in deeplabv3+ paper** is done?

I have encountered some problem with mIoU issue.
I tried out the code you provided with PASCAL dataset without augmentation, and got low mIoU(around 68%) compared to those written in paper which is around 78% and more.

And I thought it might be due to the lack of dataset (since with the augmentation, dataset is 10 times bigger like illustrated on the paper.)
image

I check the [76] paper* but couldn't find how the augmentation was done.
So it will be so much help if I could get the code or the illustration of how to do the augmentation.
(* Hariharan, B., Arbel ́aez, P., Bourdev, L., Maji, S., Malik, J.: Semantic contours
from inverse detectors. In: ICCV. (2011))

Thank you in advance!

**btw the paper I am talking about is "Encoder-Decoder with Atrous Separable
Convolution for Semantic Image Segmentation" ECCV 2018 by Liang-Chieh Chen