About the result of deepglobe dataset

[hly1999]

hi！
For the deepglobe dataset, I have some questions about the results of the replication.
This is the result of my own training backbone network and refined modules：

I have done many experiments and still can not reproduce the effect of the original paper

I want to know what's wrong and there are two groups of Coarse iou and Refinement iou. What do they represent respectively？
I hope to get your answer.Thank you very much!!!

[hmchuong]

Hi,

The first two lines are the final IoU while the last line is the IoU of a single image (and in this case, this is the last image). There is a problem with your baseline (or coarse) model since it has low result. Could you please share the config you use to train it?

[hly1999]

This is my config
config.pdf
Another problem is that the training refinement module saves the model of each epoch, but does not give the best one, so which one should be used in the final test?

[hly1999]

It's really a baseline problem. My training backbone network score can only reach 62 or 63, while the paper gives 67. Do you know why

[hmchuong]

How many GPUs you use to train the model? I used 8 V100 GPUs

[hly1999]

I trained the resnet backbone network to use three 3090 gpus. It seems that they are not fully occupied