Closed Xuyihaoby closed 2 years ago
The angle definition in OBBDetection is different from AerialDetection and S2ANet. I let the angle of the rotating box range in [-pi/2, pi/2] for the convenience of angle regression. And the angles are also counterclockwise to fit the roi_align_rotated.
The angle range doesn't have a large impact on results. But some operations like the projection of dx and dy can largely influence the final results. This repo makes a comparison of different representations.
The angle definition in OBBDetection is different from AerialDetection and S2ANet. I let the angle of the rotating box range in [-pi/2, pi/2] for the convenience of angle regression. And the angles are also counterclockwise to fit the roi_align_rotated.
The angle range doesn't have a large impact on results. But some operations like the projection of dx and dy can largely influence the final results. This repo makes a comparison of different representations.
Thank you for answering the question!!!! The data of repo shows that projection is very important.
Do you mind providing the log of training s2anet?
Hello!I was wondering whether the representation of the rotate box is same as the original model,when you reappear the model.And whether the representation has a large impact on the final results.