The third place winning solution (3/220) in the track of Fine-grained Object Recognition in High-Resolution Optical Images, 2021 Gaofen Challenge on Automated High-Resolution Earth Observation Image Interpretation.

• models
  config location:Under folder config/fair1m
  Strong backbone: add swin transformer as backbone
  storng head: adapt double head to ROI transformer and ReDet
• Training trick
  config location:Under folder：config/fair1m
  trick：Fine-tuning with SWA and cosine learning rate
• Trick
  file location： competition_trick
  model emsemble: WBF（make WBF support oriented bounding box）
  version tool： Visualize objects for analysis

• models
  ReDet, Oriented RCNN, ROITranformer
• Tricks
  Multi-scale training and testing: Limited by the computing power using two scales of 1.0 and 1.5, and crop images into 1024*1024 with stride 800
  Lower confidence: Set the output threshold into 0.005.
  Off-line date augmentation： Random combination of affine transformation, flip, scaling, optical distortion for data augmentation.

install: readme_install.md
Dataset format：only support DOTA (Changed the label format of the FAIR1M dataset to DOTA format during the competition)