When studying bounding box-based detectors, it's not clear to me if data augmentation includes adding random rotations.

If random rotations are added, how is the new bounding box calculated?

  • 1
    $\begingroup$ I don't have much practical experience with them, but, I guess, if you rotate the images, you just need to rotate the boxes (i.e. labels) too. Check this (which may not be a reliable implementation: I don't know!). $\endgroup$ – nbro Nov 15 '20 at 22:43
  • $\begingroup$ The problem is the predicted boxes are not rotated - so would you just create a new box from the rotated images bounding box corners? This works but might make the predicted box a little larger,. $\endgroup$ – FourierFlux Nov 15 '20 at 22:47

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.