Digital Africa Plantation Counting Challenge
I am working on labeling app so I chose this competition to test labeling capabilities for object detection and image regression. Using oof predictions it was pretty easy to detect serious labeling errors. For training I fixed a few of them manually to get better sense of validation progress and not try to learn obvious mistakes.
Beside the obvious mistakes I also get that is difficult to count trees on the images. Smaller trees, incomplete trees around the borders were sometimes counted another times they were left out.
I trained a few efficientnetv2_rw_s models with relatively large resolution (640, 768, 1024px).
Best submission was a blend of 5 models.
Code might be helpful for other competitions. Please find more details at https://github.com/gaborfodor/zindi_trees_9th_place
Just a question what's the best single model score?
Just checked a few single models, they had 1.62 - 1.66 on private LB
Amazing, Thanks
How was your CV score @beluga ?
I used my fixed labels for validation so my scores were probably better than raw metrics. I had RMSE ~ 1.3 & MAE ~0.76 for the best models.
Thanks a lot and congratulations on your result !. I have a feeling that training with large image size contributed less to public LB and more to private LB . May be the type of images in public and private split was a bit different .
Tbh I don't know the answer, my local validation and public LB progress got misaligned quite quickly. I stopped experimenting at that point. It would be hard to tell what helped with such high label noise.