SwinTransformer + OBBDet
The sixth place winning solution (6/220) in the track of Fine-grained Object Recognition in High-Resolution Optical Images, 2021 Gaofen Challenge on Automated High-Resolution Earth Observation Image Interpretation.
Off-line date augmentation
We use random combination of affine transformation, flip, scaling, optical distortion for data augmentation.
Multi-scale training and testing
The training images are resized into sizes of 600, 800, and 1024 for training and testing.
Swin transformer is adopt in ORCNN and RoI Transformer for better performance.
We have merged the results from RoI Transformer, ORCNN, S2ANet, and ReDet.
Set the output threshold into 0.005.
Tried but didn’t work
- Adjust NMS threshold.
- Class-agnostic NMS.
- Mosaic, and mix up for data augmentation.
- Oversample the categories with fewer instances.
- Train the detectors for specific classes with low AP.
- Multi-scale training and testing on SwinTransformer-based detectors (even dropped by about 1% mAP).