Where did you download the vgg16 pretrained weight?

[yonghaoxu]

Where did you download the vgg16 pretrained weight?

[lauraset]

Hi, @YonghaoXu. Thank you very much. I have obtained the similar results by self-training as reported in your paper. It is a practical technique.

[YonghaoXu]

Hi @lauraset, thanks for your interest in this work. The pretrained VGG-16 model is the same one that I used in a previous project on cross-domain semantic segmentation. You can download it here. I think directly using the official pytorch pretrained weight would also be fine as here we just want to have a good initialization to accelerate the training procedure.

[lauraset]

Hi, @YonghaoXu. Thank you for your detailed reply. I got it.
By the way, I used the well-trained weights of CRGNet.
When I run GenVaihingen.py, the results show that

The boundary seems poor. Is that right?

[YonghaoXu]

Yes, since CRGNet only uses point-level annotations, the generated pseudo labels may not be accurate, especially on the boundary regions. Thus, we propose to use the self-training technique to finetune the model with these inaccurate pseudo labels. The dense CRF can also help to further improve the accuracy of those boundary regions.