Hi,
I have a dataset of some large images ( > 2000 px). I need to train an yolo-obb model, so I’ve applied an rotation of ± 15° as augmentation technique.
The problem is that after applying this rotation, the image quality is quite reduced.
Original image :
After rotation :
How can I fix that ?
Hi! I’m asking our internal team if there’s a way to maintain image quality during rotation. In the meantime, are you trying to actually train the model at the >2000 pixel resolution? Is it possible that at target resolution (ie maybe 640?) the degradation is less visible?
Hi! All our augmentations resize to a max of 2000x2000 prior to doing the augmentation so there is an expectation of a downgrade. If you need higher resolution you can crop prior to uploading + augmenting, assuming ultimately you’re running at a lower resolution.
If your model really is running at >2000 resolution, I would try training without the augmentation and see what happens! 
may not be necessary for good results and would save some headache
This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.
