search

prs-eth / DGInStyle-SegModel

Downstream semantic segmentation evaluation of DGInStyle.
Apache License 2.0
21 stars 2 forks source link

Reproduce Tab. 2 results #4

Open perladoubinsky opened 5 days ago

perladoubinsky commented 5 days ago

Hello,

Thank you for the nice work !

I was able to reproduce the results of Tab. 1 but I am having trouble for Tab. 2 (especially line 3 and 4). Could specify the settings you used for training DAFormer on Cityscapes?

In addition, for the results DAFormer + DGInStyle (line 4), how much synthetic data did you use (6000 images or same as the number of Cityscapes images)? The data is taken from the same generated dataset as for Tab. 1 results?

Thanks!

yurujaja commented 3 days ago

Hello,

For Tab. 2, the generated data differs from the dataset generated in Table 1. In this case, Cityscapes is used as the source domain data for the whole pipeline.

perladoubinsky commented 2 days ago

Ok I see, thank you. For training DAFormer on Cityscapes without the synthetic data, can you specify the settings that differ from GTA training (rare class sampling, shade etc). I cannot reproduce the results of Tab.2 line 3.

Thank you!