-
Notifications
You must be signed in to change notification settings - Fork 3k
Description
Is your feature request related to a problem? Please describe.
Right now, even if you took extra effort to annotate rotated bounding boxes, the only option is to export to (normal) YOLO format, which excludes the rotation as a whole and only preserves the axis-oriented coordinates. OBB YOLO adds an extra angle to be used for object detection and has a different dataset format.
Describe the solution you'd like
Ability to extract object detection annotated data in OBB YOLO format as per the ultralytics dataset guide.
Describe alternatives you've considered
I have found an alternative that involved extracting the data in CSV format (or JSON) [as they are the only formats that keep the rotation angle] and then writing a Python script to convert it to my desired format.
Additional context
All in all, working with rotated bounding boxes is not well supported in Label Studio. For instance, an affine transformation was needed to extract & transform the rotated bounding box of the image to a vertical sub-image (for visualization, or any other use case). Label Studio does not consider the rotation in all of their demo examples which is understandable but a change would be appreciated. I would love to help in that regard as I have done it.