Yunsheng Li

I try SGD but the performance is a little worse than using Adam.

I don't think so. I use same labels as other papers

Hi, I appreciate your interest in our work. However, the code for key point detection is still under processing. We will release soon. you can refer to the code for...

I upload the labels to Google Drive. You can download them with the following link [synthia_mapped_to_cityscapes](https://drive.google.com/open?id=1ShCoaYMP1G5U3fLPCvfedQ-BFGsA2zK4)

Just update the link. Sorry about it.

It is a little weird. It should not be that bad. Could you share my the learning parameters you use? if you use my code, you can get the file...

I train with all classes but only report 16 or 13 in the paper.

You use the hyper parameters for DeepLab ResNet101 to train VGG16. I suggest you to first test the model I upload to make sure everything is correct.

Yes, it is not that stable during training especially for the small model and the first hundreds of epochs. However, it will become more stable in the last 100 epochs...

You can find the parameters I use to train VGG model from my paper.