VITA-Group/GLNet

Help for DeepGlobe dataset

Closed this issue · 6 comments

Balabala-Hong commented

Hello, good work. I am trying to run your code on the DeepGlobe dataset, but I have difficulty in getting the DeepGlobe dataset as the websites didn't respond to my request for two weeks. Is it convenient for your to provide a copy of DeepGlobe dataset on val set and test set? My personal email is bingwangnpu@163.com.
Besides, I found there is some gap in the evaluation results in your paper and benchmark results in the official website, is there some difference on your evaluation methods?

chenwydj commented

Hi! Thank you for your interest!

  1. Sorry that following the challenge policy I cannot share the dataset to others.
  2. We report our accuracy on the Deep Globe validation dataset, since the challenge is already closed. We use the exact same metric code in this repo for the evaluation in our paper. I think there may be some domain shift between the validation and test set.
nidetaoge commented

Wuyang, congratulations on your work~

Balabala-Hong commented

Hi! Thank you for your interest!

  1. Sorry that following the challenge policy I cannot share the dataset to others.
  2. We report our accuracy on the Deep Globe validation dataset, since the challenge is already closed. We use the exact same metric code in this repo for the evaluation in our paper. I think there may be some domain shift between the validation and test set.

Hello, after waitting for nearly a month, I was accepted to access the dataset. As you mentioned you use the same metric code for evaluation on val set, but I didn't found label fliles corresponding to the val set, yet your codedataset_val = DeepGlobe(os.path.join(data_path, "crossvali"), ids_val, label=True), how can you evaluate val set without corresponding labels?

  • 👍3
lamhoangtung commented

Hi @chenwydj . Thanks for your great work. I trying to recreate your result for DeepGlobe datasets.
I'm facing the same issue as @Lelouch-Ice where I can't find any label files corresponding to the val set, and the test set, yet your code:

https://github.com/chenwydj/ultra_high_resolution_segmentation/blob/e3ad351d2291ec03b6108dd17b9be57fa89ba4ed/train_deep_globe.py#L53-L56

How can you evaluate val set without corresponding labels ? Where did that crossvali and offical_crossvali folder come from ? Also the path for training data is also different from your code ...

Did you do something to the dataset before using it ? If yes, what exactly are those step ?

Looking forward for your answers. Thanks a lot

  • 👍2
  • 👀4
dexception commented

Can anyone share a pretrained model so that others can just run and see if it is worth the effort ?

chenwydj commented

Hi! We updated the instructions for training and evaluation in the Readme, and also uploaded the pretrained model for the Deep Globe. Please take a look and give a try! :)

  • 👍1