Hi,
Congratulations on your great work and thanks for releasing the code.

For the GTA5 to Cityscapes training code, I could not find the gta5_train.txt file. While I created it on my own, the code also does not map the GTA5 labels from 1-33 range to the Cityscapes label IDs (0-18) which leads to errors. I managed to fix that by performing the mapping. However, after training for around 4-5 epochs, the Cityscapes validation mIoU is very low (~0.5 to 1% only) and not increasing. I find this very odd since plain training on GTA5 usually quickly gives at least 15-25% mIoU. There is no model specified as initialization and I believe the PyTorch ImageNet-pretrained one is being used.

Kindly advise on whether some other initialization is required or some code additions are required or some preprocessing is required for the datasets. It would be very helpful if you could release the corrected code.

Hi, thanks for your great work and for releasing the code.

I want to draw a Gaussian kernel density estimation diagram like figure 1, can you please release the code or give me some advice on how you plotted it?

Thanks for reading.

Hello Mr. Chen!

thanks for the great work and repository!
Currently, I'm trying to reproduce your numbers stated in the paper. Using your provided code and checkpoints, I cannot achieve the same results.
In my case, I get an mIoU on Cityscapes if 36,141% on the ResNet50 checkpoint.
Does the provided checkpoint differ from the one used for generating the paper results?

Thanks.

Hi @chenwydj! Thanks for this outstanding work. Your public code really helps me a lot.

But I am a little confused about the model selection procedure, since the best model evaluated on the target domain (Cityscapes) is saved. This seems not a popular way to select models in domain generalization. Or is this the popular way in the current GTA5->Cityscapes task?

Pls let me know if I miss something. Thank you so much.

Hi there,

Thanks much for sharing the code and great work!
Quick question on the features used for kde - which method was used to convert the high-dimensional features after glocal average pooling for the gaussian kde? Is this TSNE or some other method?

Thanks again!
YS

Hi, Thanks for your great work!

I have a question about the segmentation model you used in your paper.
The paper on the segmentation model you referred to is as follows. https://arxiv.org/pdf/1706.05587.pdf
As far as I know, this paper is about Deeplabv3.
However, in the paper, it is marked as Deeplabv2.
Can you tell me what is the difference between the segmentation model you used and deeplabv3?

Thank you for reading.

Hi Team,
Thank you for the great work and releasing the code.

We are wondering about the Es: hyperspherical energy of features part in the paper, however in the code base we couldn't find it.

It would be great, if you can guide us or have the implementation part of it to get the E_s part , similar to Table 1 (https://arxiv.org/pdf/2104.02290.pdf)