Test results are consistent with the reported values in the paper ?

[LinaShanghaitech]

Hi, I've attempted to test your provided best caffemodel in the validation set. Following the results you reported in SSL, the overall acc., mean acc. and mIoU should be 84.36%, 54.94%, 44.73%, respectively, but I used the provided code and only got the 82.06%, 51.64%, 40.95%. Besides, I've also tried other methods (deeplab_v2 , attention without ssl ) according to your training strategies and all of them get the lower results. Are there any changes to the validation set ? Thx !