Datasets:
Question on the full DisasterM3 dataset
Amazing work! I noticed that in the zip file of DisasterM3_Bench.zip (11GB), there are only test_images and masks folder. In the test images, there are roughly 6k images. Could I ask if you would provide the train set as well, or the full 26K imagery mentioned on the paper?
Many thanks for your interests.
I have already uploaded the train set.
I'm also very interested in this paper! As a beginner, I just downloaded the dataset from the provided link.
However, I'm a bit confused about the data size. The paper mentions the full dataset should include:
Instruct Set: ~21k images (17,190 optical + 3,798 SAR)
Bench Set: ~6k images (5,024 optical + 976 SAR)
But what I downloaded contains:
Instruct Set: 559 images
Bench Set: 7,278 test images
Could you please help clarify if:
The current release is a sample/subset for demonstration?
I might have missed some files or additional download steps?
The full dataset is available through a different channel?
Thank you so much for your guidance and for sharing this amazing work!
You may encounter an incomplete downloading process.
what are the subset json files. they are required to run_vllm but they're not available neither on the github repo nor hf
'subset' refers to the task name; I have combined them in the benchmark.json.
You could find the task types in the provided JSON.
The file sunda_tsunami_00000135_pre_disaster.png in the dataset is empty. For the disaster type recognition task, landslide samples are present in the training set, but there are no ground truth landslide annotations in the test set. In that case, how were the evaluation results for different disaster types obtained in Section 4.2 of the paper?
Many thanks for your feedback.
Let me check these
Many thanks for your feedback.
Let me check these
Thank you very much for your reply. I'm looking forward to seeing the finalized data.
I checked the data, and there might have been a misunderstanding in the explanation or naming. We do have landslide data; image scenes named shovi_landslide_* and kalehe_flooding_* represent landslide disasters. However, our disaster_type.json file doesn't query for these two events (to maintain sample balance, not all images have a disaster type query). But the landslide accuracy in Section 4.2 of our paper is calculated by statistically analyzing the relevant tasks for shovi_landslide_* and kalehe_flooding_*.
Besides, we updated the 'sunda_tsunami_00000135_pre_disaster.png'. Please have a check.
If you want to explicitly test the landslide types, I also updated new versions. Please check here:
https://drive.google.com/drive/folders/1KTiJKjJIA3gzbsxQPYArADG55Zfb8R52?usp=sharing