Do you know that pepper_septoria bboxes are duplicates of tomato_septoria? Pepper_septoria bboxes perfectly overlay Tomato_Septoria bboxes in a tomato image. How did you go about this ?
So, the idea is that a tomato bounding box shouldn't be in the same image with a pepper bounding box since these are different crops - which makes sense in a normal world. Another observation was that the bounding boxes perfectly align, they have the same value, meaning they are duplicates, the only difference just being their class name. However, my observation was that when I removed the duplicates, the public leaderboard score dropped. So, my thinking is that these duplicates also exists in the test set held out by zindi.
"However, my observation was that when I removed the duplicates, the public leaderboard score dropped. So, my thinking is that these duplicates also exists in the test set held out by zindi."
I tthink he has already shared his solution. He left the dataset untouched. He didn't remove those contradictory classes because the public leaderboard score dropped. You can also do same.
Do these bounding boxes seem to mark out a meaningful area when you pick out an image or it's just garbage.
So, the idea is that a tomato bounding box shouldn't be in the same image with a pepper bounding box since these are different crops - which makes sense in a normal world. Another observation was that the bounding boxes perfectly align, they have the same value, meaning they are duplicates, the only difference just being their class name. However, my observation was that when I removed the duplicates, the public leaderboard score dropped. So, my thinking is that these duplicates also exists in the test set held out by zindi.
I am having the same issue especially considering tomato and pepper. If it's not too much to ask please share your solution if you come up with one
"However, my observation was that when I removed the duplicates, the public leaderboard score dropped. So, my thinking is that these duplicates also exists in the test set held out by zindi."
I tthink he has already shared his solution. He left the dataset untouched. He didn't remove those contradictory classes because the public leaderboard score dropped. You can also do same.