questions about the backbone on vgg16 model

[feipanir]

Hi, thank you for sharing this code. But I have a question that, have you ever tried running your code on the model with vgg-16 backbone? It seems that it works well on Resnet-101, but when we try it on vgg-16, the performance was dropping. It would be very helpful if you could share some experience.

Thank you.

[vikolss]

Hi, Unfortunately we did not try using vgg-16 backbone, since most existing works focus on Resnet-101 backbone. So I'm afraid I don't have any insight to share. Is the performance dropping more here than for other similar works implementing both backbones?

[feipanir]

We use a common Vgg-16 as backbone to train on our own dataset. Existing work performance on this dataset is validation mIoU > 30.00. While using DACS, we observe the val mIoU scores as follows: Iter_id Perfermance 5000 mIoU: 19.42 10000 mIoU: 15.68 15000 mIoU: 12.91 20000 mIoU: 12.44 25000 mIoU: 10.95 30000 mIoU: 9.92 35000 mIoU: 10.58 40000 mIoU: 9.51 45000 mIoU: 8.80 50000 mIoU: 7.30 55000 mIoU: 7.89 60000 mIoU: 6.25 65000 mIoU: 7.65 70000 mIoU: 5.44 ..... Still we don't figure out why it is dropping continuously.

Regards,

Fei

[vikolss]

I see. I currently don't have the opportunity to conduct these experiments myself, unfortunately. The performance of DACS is dependent on the distribution of classes in the data, so a possible reason could be that the method is not suitable for the data set that you are using. Have you tried vgg-16 backbone with GTA5/SYNTHIA -> Cityscapes to see if the same thing happens there?

Viktor