# 185th place solution

Thanks to the organizers and kaggle team for hosting this competition! Also great thanks to
[@rwightman](https://www.kaggle.com/rwightman)
for his efficientdet-pytorch repo,
[@zfturbo](https://www.kaggle.com/zfturbo)
for his weighted-box-fusion repo, to
[@shonenkov](https://www.kaggle.com/shonenkov)
for his amazing public notebook based on which I started this competition and to everyone who has been active in the discussion forum during the competition. Finally congratulations to all the fair prize winners and to the fair medal winners.
Summary:
5 fold d2 trained on 1024x1024 + 5 fold d4 trained on 512x512
Heavy augmentations (Added more augmentations along with the alex's public kernel augmentations for example: Noise augment, Scale and crop, this gave me 0.01x-0.2x boost)
Used weighted-boxes-fusion for ensemble
5xTTA
Pseudo labeling(1 epoch, 1 round, training data + 2x pseudo test data)
All models ensemble with 5xTTA and wbf achieve 0.7563 public LB/ 0.6956 private LB. Adding pseudo labeling improves the score to 0.7595 public Lb/ 0.7050 private LB.
Did not work for me:
Cleaning the bounding boxes
Gradient accumulation
Cutmix and Mixup scores were not as good as the models without them
WBF threshold Tweaking
Tweaking efficientdet loss function hyper-parameters