[Feature] The performance of transformer-based rotation object detection is not as good as ResNet50.

[ZhenboZhao77]

What's the feature?

Recently, I conducted experiments comparing different models and found that using transformer-based models for rotation object detection did not perform well. I recall reading that the QKV mechanism might not be well-suited for objects with large scale variations. Could anyone share more detailed papers or explanations on the disadvantages of QKV in the application of rotation object detection?

Any other context?

No response