Annotations created with negative label values

Community Help
1

The project is object detection using Oriented Bounding Boxes.

If using Bounding Box level augmnetations like Bounding Box: Flip: Horizontal, Bounding Box: 90° Rotate: Clockwise, Counter-Clockwise to augment the annotated dataset the label txt files created end up with some negative lable values which when trained on the framework presented at this link https://blog.roboflow.com/yolov5-for-oriented-object-detection/ throws this error:

train: WARNING: …/yolov5_obb/dataset/Pineapple-6/train/images/scene00062_jpg.rf.a718a2ded3f8c3123482c94f00d5060f.jpg: ignoring corrupt image/label: negative label values [ -0.41199], please check your dota format labels
train: WARNING: …/yolov5_obb/dataset/Pineapple-6/train/images/scene00062_jpg.rf.feac0b19e6e81873a5af99de7867fa22.jpg: ignoring corrupt image/label: negative label values [-1.7603e-05], please check your dota format labels
train: WARNING: …/yolov5_obb/dataset/Pineapple-6/train/images/scene00168_jpg.rf.75af23dc982c07b171b86cb7be4067ba.jpg: ignoring corrupt image/label: negative label values [ -21.147 -5.264], please check your dota format labels
train: WARNING: …/yolov5_obb/dataset/Pineapple-6/train/images/scene00168_jpg.rf.8455485aca7a940a14b497f1467f8226.jpg: ignoring corrupt image/label: negative label values [ -3.1118], please check your dota format labels
train: WARNING: …/yolov5_obb/dataset/Pineapple-6/train/images/scene00168_jpg.rf.f908926844f3546560d5daff3765075f.jpg: ignoring corrupt image/label: negative label values [ -10.186 -27.568], please check your dota format labels
train: WARNING: …/yolov5_obb/dataset/Pineapple-6/train/images/scene00171_jpg.rf.529b7789994ee625af28850063414896.jpg: ignoring corrupt image/label: negative label values [ -0.93258 -11.287 -20.121], please check your dota format labels

is this a problem or I am misunderstanding something? Just as a note I exported the dataset for YOLOv5_obb (and apparently that shouldn’t have negative lable values)

2

Thanks for reporting. We’ll have a look and see what we can find.

In the meantime I’d recommend trying the official YOLOv5 instance segmentation model (or YOLOv8 instance segmentation which is even newer); YOLOv5 OBB is pretty old/obscure now and, to my knowledge, not actively maintained.

3

Thanks for that. The issue is for this particular project we really don’t need to know the object’s shape at all. It is of no interest. The only thing I am looking for is the orientation. This is why the obb seemed the perfect solution. The instance segmentation labelling would be a huuuge job which I cannot really justify , especially in the 1st stage when I am just trying to build a prototype to see if this can be done .