[datasets] Add multi-label classification annotations to OpenFire
frgfm opened this issue · comments
🚀 Feature
Add multi-classification targets to OpenFire dataset.
Motivation
Currently, OpenFire only contains binary classification target, which is not enough to reach production-ready performances. Extending the target to multi-label will help reaching better model performances.
Pitch
The JSON file with annotations will have to be updated with including targets for multiple task:
- binary classification: single boolean value per sample (True notifying the presence of a fire)
- multi-label classification: a few options
- list of present class per sample (eg.
['class1', 'class3']) - list of present class indices per samples (eg.
[0, 2]) - one-hot encoded vector of present classes (eg.
[True, False, True, ...])
@blenzi I'd like your opinion on this. Do you think we should move forward with this? Or discard it?
I was about to close it, but I wondered whether it might still be good to have a multi-label classification dataset (OpenFire or WildFire, or another).
Thanks for looking into this @frgfm. I have the impression this would not be extremely useful for the OpenFire dataset. If I understand correctly it requires new annotations, which will take long to complete. Maybe you want to wait some more results on WildFire to decide?
Agreed, good idea!
Considering latest training results by @MateoLostanlen are quite encouraging, and that the decision to have a multi-label dataset can only be made based on what has been annotated, I'm closing the issue for now.
Feel free to reopen if you disagree