The results of reimplemented models

[guangmingboy]

I want to know the results of re-implemented models.

[speedinghzl]

PSPNet ~78.5 Deeplabv3 ~78.9 on cityscapes val dataset with single scale testing.
And we have uploaded the document to arxiv and it may take a little time to be public.

[speedinghzl]

The document is here https://weiyc.github.io/assets/pdf/toolbox.pdf

[lzrobots]

Are these results (pop 78.5, deeplabv3 78.9) based on the default cross-entropy loss?
So I assume the performance can be improved if OHEM loss is adopted (ohem flag=true)?

[speedinghzl]

Yes, these results can be obtained without OHEM. If you want to use OHEM, please refer to this issue.