About Results provided in the article

[VoyageWang]

Hi there! I am interested in your work. Your work is also brilliant. I have a question about the results you provided in the article in Table 11. When you do experiments in the Potsdam dataset, these results are from the train set or test set. I did not reach the performance listed there. And did you rework these results of previous works in table 11 like Unet FPN and so on? If you did, these results are pretty good. Are these results from the train set or the test set? Can you help me? That will help me a lot.

[zyxu1996]

Thank you for your praise of our work. All results in Table 11 were evaluated in the test set according to official recommendations. The results of Unet, FPN, etc. are our applications using our codebase. Hope to help you.

[VoyageWang]

Thanks for your reply! I also have another question: in figure 1, you compare the parameters and FLPOs between different models. I wonder what your parameters stand for. Did you mean the trainable parameters or the model size (MB). Can you help me again? That will help me more.

[zyxu1996]

During the model inference phase, the fvcore package in tools folder is used to calculate parameters. Of course, you can take them as trainable parameters. They are usually the same. In addition, parameters are usually linearly related to the model size (MB).

[VoyageWang]

Thanks for your reply! There were other questions arose. When I used your code, I train on my device. I directly run train.py instead of using the shell. The problem was that: the evaluation indexes here are all zero. It does not work. how can I solve that? I noticed that your code is trained on four GPUs, how can I change to a single GPU? I just changed the satement from os.environ["CUDA_VISIBLE_DEVICES"] = "0,1,2,3"to os.environ["CUDA_VISIBLE_DEVICES"] = "0". Is that correct?

[zyxu1996]

This code will only evaluate the model in the last three epochs. See here test_epoch = args2.end_epoch - 3

[VoyageWang]

We have run the model, but the model performance is poor. Every epoch will predict a value. But its performance is poor as the model is trained. As you can see in the figure. Can you give us some suggestions?

[VoyageWang]

Thank you! The problem was solved. I want to ask how to visualize the output features like fig 12 in the article. Can you share the codes? Thanks a lot!!!