Finalist's solution in the track of Oriented Object Detection in Remote Sensing Images, 2022 Guangdong-Hong Kong-Macao Greater Bay Area International Algorithm Competition.

This project is based on Jitto framework. Please follow the official installation documentation for installation.

Jianhong Han, Zhonghao Fang, Zhaoyi Luo

• Backbone
• Support Swin-Transformer Tiny/Small/Base/Large Backbone Network.
• Neck
• Support PAFPN network.
• Optimizer
• Support AdamW Optimizer.
• Some Useful Tools
• Support Model Ensemble.
• Support Soft-NMS, Class-Agnostic NMS.
• Support HSV Data Augmentation.

• Training Data Augmentation
  We use random combination of hsv, horizontal/vertical flip, rotation for data augmentation.
• Multi-scale training and testing
  The training images are scaled to 0.5,1,1.5 times and cropped to 1024x1024 for training and testing.
• Swin Transformer Backbone
  We use Swin-Transformer as backbone in Oriented R-CNN, S2ANet and ROI Transformer for better performance.
• Model Ensemble
  We merge the detection results from Oriented R-CNN, S2ANet and ROI Transformer for better performance.
• Test Time Augmentation
  We use extra random horizontal/vertical flip, random rotation for inference phrase.
• Soft NMS and Class-Agnostic NMS
  We use Class-Agnostic NMS for post-processtion. Soft-NMS used but not work.