Yolov8-obb generated labels are not rectangles and do not match the original oriented rectangles

Object Detection, Project Link here: Sign in to Roboflow

Image in question for the example is here: Sign in to Roboflow


Was training a yolov8-obb model and when I went to draw rectangles to evaluate the results I was comparing the rectangles in the ground truth labels from roboflow against the predictions … and I found out that the rectangles in the ground truth data aren’t really rectangles.

It’s a little hard to see in the image but the white border around the marked rectangles in the picture are parallelograms instead of rectangles. The same is true for the black rectangles and the smaller gray ones around the symbols in the middle of the shorter edge of the rectangles.

I checked and rechecked the labels, and even verified that the rectangles stored on roboflow’s servers are accurate (they are, and your visualization online looks correct too).
Somehow while generating the oriented rectangle data there seems to be a bug creating this issue.

Please fix this! I can’t use your service to train the models otherwise.

You should be able to repro by downloading yolov8-obb format of version 6 of this project, which will include that image as an example. I’ve verified that all the images in this project have the same issue.

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