Yolov5 txt format seems to include segmentation mask instead of bounding box

Community Help
1

Please share the following so we may better assist you:

  1. Project type (Object detection)

What happened:

  1. Downloaded yolov5 txt format of Sign in to Roboflow
  2. Looked at file, for objects that I annotated with a polygon, I am getting a line like:
    8 0.7591515229166668 0.17062089895833335 0.5930737843749999 0.17920242291666666 0.6006457177083333 0.3286219083333333 0.766723456250 0.32004038333333334 0.7591515229166668 0.17062089895833335 which looks like a closed polygon string instead of a bounding box.

In the past when I did this I would get bounding boxes and not polygon masks. Does yolov5 automatically handle cases like this now? Can I get only bounding boxes if I want? I’m trying to construct a pose task like in https://github.com/TexasInstruments/edgeai-yolov5/tree/yolo-pose or in yolov7 pose library and I don’t think they accept a polygon annotation.

2

Yes, it does as of v6.0. And it gets you better training results by enabling additional augmentations like Copy/Paste.

You can read more about using polygons to improve object detection models here.

Can I get only bounding boxes if I want?

Not in YOLOv5 format, but you can export as a different YOLO format like v4 that doesn’t support polygons & it’ll give you the surrounding bounding boxes.