Digital Africa Plantation Counting Challenge
This competition has been highly questioned in terms of data intrigity. I personally believe that perhaps the presence of wrong label in the test data as observed in the training data leads to the great leaderboard reshuffle in the private score.
It is not a new thing for position to change when the private score got revealed but this particular competition is one that can be easily evaluated locally. It has a lot of clear zero images and not too much images to check if indeed your model performs well. But the labeling errors like labeling a 0 as 29 among others in the train set, that may 🤔 also be in test data is a challenge on the data intrigity.
I really recommend @Zindi to confirm if the test data (private labels) used in this competition is free of error.. On the other hand, no issue with train data error, as this is normal but it is important to test a model on clean dataset and not on wrongly labeled one.
Once again, congratulations 🎊.
I am sharing your feeling. I don't know why my rank decreases of 7 on public leaderboard (1.967763457) to 52 on private leaderboard (1.747855829).
What were the CV for your models and what image sizes did you use ? I didnt experience shake up or down but probably my models were not that good to start with
I used a bunch of non tuned NN ,svr and lgbm models with image size 256x256 and just did a simple averge of that . The selected model was not my best public LB .
My discussion is not just to challenge my position but primarily to clarify if @Zindi ensures the intrigity of the private data despite all comments about the concern raised during the competition. And my comment becomes necessary observing the great shakeup for such a simple evaluation task (manual counting).
To answer your question, my cv score was around 1.5 to 1.78 . Image sizes used for training includes 384, 1024, 1536 and 2024.
No no ,I understood your intent . I was just curious of your solution as i could not even get a good score in Public LB.. So you must have done something right . Even with right label my models probably were not as good in predicting the cases where there were only partial image visible and few palm trees ..I saw the models were confused on those cases with respect to the right count .
Wow I expected quite some shakeup. I even stopped competing when I realized how high the training label noise is.
Actually, each tip you share geometry dash lite is a good and practical tip to apply while uploading articles. I think by following the basic lessons you can firmly grasp the background knowledge. Please continue to share values with the community.
I noticed the models got the right count wrong in those specific instances.