Problem when extracting dataset in YOLOv8 format

Community Help
1

I am having an issue when extracting a dataset in the YOLOv8 format. When I start training the model, I noticed that the bounding boxes for one of the classes are completely incorrect, even though they appear to be correct in the Roboflow dataset. I suspect that the problem may be occurring during the extraction process. Can anyone help me with this issue? I have spent 5 days annotating images on Roboflow and have not been able to complete my project.

In the left photo, you can see an example of a training batch inside YOLOv8 with a dataset extracted from Roboflow. The coordinates for class 1, named “gousset,” are incorrect. The right photo shows what the detection for class 1 should look like.

image
image1920×957 137 KB

  • Project Type: Object detection
  • Operating System & Browser: Windows
  • Project Link App & Universe
2

Hi @Dhia_Barhoumi

Do the images showing up incorrectly on the train_batch.jpg images show up correctly on the generated images for the dataset version? Are all the images off or just a certain subset?

3

@Dhia_Barhoumi Have you resolved your question? Because I am also facing the same issue.

4

Try Auto-Orient on the Preprocessing stage. It solved this problem for me.

1 Like