After annotating my images and creating a new data version (with or without image preprocessing), I noticed that the bounding boxes on the cars are not maintained. When I review the annotated images on Roboflow, the bounding boxes appear correctly (as shown on the right side of the attached image). However, after exporting the dataset in YOLO8 format and reviewing the images, the bounding boxes are either misplaced or missing entirely (as shown on the left side of the attached image).
Here’s a summary of my workflow:
Annotate images in Roboflow.
Create a new data version (with or without image preprocessing).
Export the dataset in YOLO8 format.
Download the zip file to my computer.
Review the exported images and observe the bounding box discrepancies.
Has anyone else encountered a similar issue or have any insights on why this might be happening? Any advice or solutions would be greatly appreciated!
Thank you!
Project Type: Object detection
Operating System & Browser: Macbook / Google Chrome
I have also gone through the discussion on the misaligned bounding boxes after augmentation (Misaligned bounding boxes after augmentation - #13 by leo) but nothing seems to work for me. At the end of that thread, it looks like it just worked for them, but I can’t seem to replicate the success.
Could you share how you are visualizing these discrepancies? We aren’t aware of any current issues with exports.
Information on the workspace and project (or Universe link if your project is public) and the export format you’re using would also be helpful in debugging this issue.
After investigating this issue, I found that the problem lies with the annotations themselves. YOLOv8 (Standard YOLOv8 Dataset) expects annotations to be in the form of bounding boxes (rectangles). However, the annotations I used included a mix of rectangles and polygons. This discrepancy likely causes the bounding boxes to be misplaced or missing in the exported dataset for YOLOv8.
To confirm the idea, I converted the polygons to bounding boxes to standardize the annotation.
Another option would be to export as YOLOv8-oriented bounding boxes dataset.
Suggestion:
When exporting a dataset into YOLOv8 that was annotated with something other than rectangles, I suggest showing a warning to the user, as this would be a helpful feature to have.
