Sharing Solutions and Ideas

Hi all,
I thought it would be nice to share ideas and solutions. Here are mine;
Leaderboard rank 27:

Yolo v3 from Gluon.
5 models from Kfold split
Weight Initialization from 2020 kaggle competition
Weighted Boxes Fusion
Pseudo Labelling tested but dropped for final submission. In hindsight, should have done it and not optimized ADA for public leaderboard.

Also tried FasterRCNN but YOLO did well in my experiments

Leaderboard rank 10:

ensemble with wbf:
effdet6_768 (2 folds)
effdet6_896 (1 folds)
effdet6_896 (finetuned with pseudolabeled data)
effdet6_896 (trained from scratch with pseudolabeled data)
postprocess: delet boxes after fbw with conf < 0.3 (aproximetly)

1 Like