fcdl94 / MiB

Hi,

Could you please share the code for generating data splits under data/ folder, it is convenient to use default splits such as (19-1, 15-5, 15-5s, 100-50), but It would be good to have the code, I am interested in some different splits.

BTW, I found the split voc 15-5s step-0 seems wrong, but I can use the one from 15-5, they should be the same.

Thanks.

Hi!
Yes, the split for 15-5s is missing since I always used the pretrained made on the 15-5 task. Sorry about that, find you solved. I will copy and paste also in the repo.

About the code for the splits:

For Pascal-VOC the code to generate the splits can be found directly in the dataset.

MiB/dataset/voc.py

Line 138 in e74fb8d

idxs = filter_images(full_voc, labels, labels_old, overlap=overlap)

Changing the overlap flag allows you to generate the "overlapped" version or the disjoint one.
For ADE-20K, the splits were made using a jupyter notebook which I need to clean up before publishing. I hope to be able to upload it before the weekend.

Hi!
I managed to clean the code and I uploaded the script to split the ADE20K dataset.
See it here: https://github.com/fcdl94/MiB/blob/master/ADE-Split.ipynb
However, I don't guarantee that you can re-create the same split used in my paper since I lost the actual seed to generate mine (but you can find the splits I utilized in the data folder of this repo).

Hope you can find it clear, but if you have any question, please comment on this issue.
Otherwise, if there are no additional comments, I will close the issue. 👍

Many thanks!!!

Code for generating data splits