Images are flipped in YOLOv8 dataset export, affecting label coordinates

Community Help
,
1

Issue Description:
When exporting a dataset in YOLOv8 format (ZIP file), 133 out of 1523 images appear to be flipped. An example of such an image and the associated label is attached. Due to the flipped image, the coordinates of the objects become incorrect, rendering the dataset unusable for further work. This issue did not occur previously.

Steps to Reproduce the Issue:

  1. Open the project on Roboflow at the following link: STADA KZ Prod General.
  2. Export the dataset in YOLOv8 format (ZIP file).
  3. Unzip the ZIP file and review the images.
  4. Notice that several images are flipped, as shown in the attached example.

Expected Behavior:
All images should be saved in their original orientation so that the label coordinates remain correct.

Project Type: Object detection
Operating System & Browser: macOS, Google Chrome
Project Universe Link or Workspace/Project ID: STADA KZ Prod General

Question:
Could you please help? What should I do to resolve this issue?

069-7E88DB39-D85D-472E-93A7-1D69926BEBE2_png.rf.e204f47e30f689169f4c80ccad8acbbc
069-7E88DB39-D85D-472E-93A7-1D69926BEBE2_png.rf.e204f47e30f689169f4c80ccad8acbbc1920×1440 319 KB

2

@Jacob_Witt, Could you please provide some advice on this bug?

3

@Дария_Новикова - thanks for the tag! Will take a look.

4

Hi @Mikhail_Korotkin! I see you didn’t use the auto-orient preprocessing.

Can you try generating a new version with auto-orient enabled? I guess it’ll manage your image orientation metadata properly.

5

Hi @iurisilvio

Thank you for your suggestion! However, enabling “auto-orient” seems to compress my images by a factor of 10, which I believe could affect the training process. I’m concerned this might reduce the quality of the model’s learning. Could you please advise if there is a way to use auto-orient without this compression, or if there is another recommended approach?

Thank you for your support!

6

Hi @Mikhail_Korotkin, it is not expected that auto-orient compress your images.

We resize images only when you apply augmentations.

I see all versions you generated don’t have auto-orient enabled. I recommend you try it and test a model to understand if the lower quality images affect your results, in general it works nice.

7

This topic was automatically closed 21 days after the last reply. New replies are no longer allowed.