Authors: Yongsa Chen

Affiliation: Lenovo

Description: Use the Adaboost method, base model is DBNet_resnet18_fpn.
step:
1. Approximate original polygons with coarser ones to speed up the train.
2. Filter out some bad annotated samples in the training set. (training: validation is 9:1).
3. Use the Adaboost method, after every training process, adjust the weights of the samples, and resample to prepare training data for the following training process.
4. We use 4 DBNet_resnet18_fpn models generated by the Adaboost method.
5. Use SGD and ReduceOnPlateauParamScheduler, 20 epoch for each model.