Question regarding IoUs of pretrained HRNet Proto about protoseg

[tfzhou]

@ksakmann I guess that there are some issues about your testing. I remember that some guys reported the problem in other projects but I cannot recall the details. I will check and come back to you later.

[ksakmann]

@ksakmann I guess that there are some issues about your testing....

I thought so too at first, but when I take the predicted labels saved in the output label dir and run a different IoU evaluation script on them, I arrive at the exact same numbers as above, i.e. an mIoU of 85.6%

If you want to reproduce it:

• in tester.py remove the palletized output Flag 'P' in fromarray(label_img_, 'P'):

208                    label_img_ = Image.fromarray(label_img_)

• set CSUPPORT=False in cityscapes_evaluator.py to use plain numpy.

[ksakmann]

Sorry for the confusion, my bad. I blindly copy pasted the instructions for data preparation from openseg.pytorch which have the flag coarse set to True. So the numbers I got were for the coarse labels.