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.

Members

Qi Ming, Junjie Song, Yunpeng Dong.

Solution

  • 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.

  • Strong backbone
    Swin transformer is adopt in ORCNN and RoI Transformer for better performance.

  • Model ensemble
    We have merged the results from RoI Transformer, ORCNN, S2ANet, and ReDet.

  • Lower confidence
    Set the output threshold into 0.005.

Tried but didn’t work

  • Soft-NMS.
  • 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).

Detections

demo image

GitHub

View Github