Codebase for Global Wheat Detection.

• training-efficientdet-v2.ipynb

I used this script to train my final model.

With psuedo labeling, this model can achieve LB/PB 0.761/0.707. Ensemble further boosts to LB 0.765.

• eval-efficientdet-sources.ipynb

  I used this script to evaluate mAP by data sources.

  Since it is hard to tell model performance with concatenated mAP, data source split tells more about it.

I mainly worked on Centernet training for this challenge. (But did not use it for final submission)

The implementation is based on camaro's repo.

Although worse than Effdet for wheat, Centernet is much more easier to customize.

• config/3x3_traincrop_mixup.yaml

should be the best model setup which uses rx101 for backbone and fpn.

• centernet_train.ipynb

  Is the main training script.