Congratulations ASSAZZIN! Your solution is very impressive to me. I will try to learn it. I also applied binary classification but it's only good in Public LB and I don't use it in Private because my CV score is reduced.

BTW, with such deep density of pbw that you have applied segmentation. I feel quite surprised!!

You mean Ground Truth or Predictions ?

• Ground Truth : Center of each worm bounding box
• Prediction : (Unet prediction mask) > 0.1

I was referring to the ground truth you used to train the Unet. Got it thanks.

• The Classifier will predict if an image contain a worm or no,the idea came from this ZindiWeekendz hackathon : Hack_Link. This Classifier is used because i found that Yolo predict some false worms. So the classifier reduce False Positives.
• U-Net prediction is not always perfect , there is some examples U-net output something like heat map

• For that I just generated the largest bounding box that contains worms in an image , then transform the original bounding box coordinates to match the new cropped image

I see, good example. Thank you. Is it the prediction with the rule all the pixels > 0.1? And if to try alpha * max(image) instead of 0.1 searching alpha with the help of Optunà or findContours and algebra?

Or alpha * max(ground truth bounding box) or maybe trying to skeletonize the prédiction in an attempt to find peaks of probability values... or something else... just to get rid of the 0.1 in favour of something more flexible.

Yes ,

my final mask = the prediction with the rule all the pixels > 0.1