CGIAR Crop Damage Classification Challenge
Helping Africa
$10 000 USD
Completed (~2 years ago)
Classification
1148 joined
347 active
Start
Oct 27, 23
Close
Jan 28, 24
Reveal
Jan 28, 24
Key to win : Data Issues & Label Errors ?
Data Ā· 21 Jan 2024, 10:56 Ā· 8
for those who done data cleaning is the percentage of mislabeled data worth cleaning it?
This dataset is whole crap, images are all mixed up. Some G(good) class contains samples with barren land, some drought class have full of vegetation, and it's too hard to even manually differentiate between ND, DR, WD and other for some images. If these kind of mislabelled images are in test set too, then it's just luck to get good score in lb.
@Nayal_17 what's the trick to get to 0.56.... range???
I would not say that everything is right, I also thought so, but when I saw corn that seemed to have dried up from the drought, but this is corn that it's time to mow in the fall
@Koleshjr I am using the public fast ai notebook, with different models(two model ensemble), and basic hyperparameter tuning, there is nothing new about it may be just overfitting , I don't know.
Thanksss @Nayal_17
How do one clean data in this case?
I think more than 45% of the data have noise
cleaning is your key to win