Lower accuracy of PSPNet than the reported numbers of this repos

[xiaomengyc]

Hi, I trained PSPNet with ResNet101 as the backbone. The mIoU values are 78.29/85.74/95.18 which are lower than the numbers in the readme file (0.7907/0.8636/0.9534). I use the same config as the author provided and do not change anything. Does somebody get the same accuracy as the reported? My PyTorch version is 1.3.1. Don't know if PyTorch will affect the accuracy a little.

[swjtulinxi]

hi，i got the same problem,have you solved it

[xiaomengyc]

No, I haven't. Still do not where the problem is.

[swjtulinxi]

the author use resnet101 and ms predict ,,do you use the same batchsize and cropsize?

[xiaomengyc]

Yes, I use the default configuration. Batchsize and cropsize should be the same as the author's implementation. Are your results same to my numbers?

[swjtulinxi]

my results is lower than yours,because i don't have enough gpu, i also know that many other people can 't reach author's result ,yours seems to be the best one so far, i think your results seems that you don't use ms predict,

[xiaomengyc]

Yes, I didn't use MS prediction, because the author's results 0.7907/0.8636/0.9534, also do not use MS prediction. They are indicated by ss and ms in the tables.

[CharlesPikachu]

maybe you can try this repo: https://github.com/SegmentationBLWX/sssegmentation

[bea-CC]

Trained with PASCAL VOC2012 segmentation dataset, I got 0.7843/0.8593/0.9520（ss）、0.7938/0.8657/0.9540（ms）with TITAN Xp（4 GPUs）.
but I spent 10.5h， is it reasonable?
And I don't understand why the paper said "we achieve 82.6% accuracy.", It's not the same as here.

[shashankskagnihotri]

I got Eval result: mIoU/mAcc/allAcc 0.6570/0.7378/0.9222 (ss) for PASCAL VOC2021 on 8 nvidia rtx 2080 gpus
Seems much lower. I used the default hyperparameters mentioned in the configs for pspnet50.
Did anyone try other hyperparameters and got better results?