Bounding boxes on images during yolo training goes all over the place

[AScythe]

I used roboflow to annotate my images (see comment for sample image)

then i download my annotated image dataset from roboflow.

when i train yolo using the the dataset, i check the training and validation images after training, and i saw that the bounding boxes go all over the place, see example below:

i checked the bounding boxes on the dataset that i downloaded to see whether they are on correct places, and it seems they are… but when i run the training, the training shows me that the bounding boxes are not in the right place… and this makes the trained model detect wrongly

if someone knows why this bug is happening and how to solve it, i will be very grateful…

[AScythe]

this is example of the correct annotated image

[Automatez]

Hey @AScythe I’m kind of thinking you have polygon bounding boxes in that image but YOLO is looking for rectangle bounding boxes. So unless you did something more to the images (like converting the polygons) it probably would not work well.

[AScythe]

thank you for your reply…
what do you mean, if i used “polygon”, it is not supported in training yolo model?
does it mean, i have to use only rectangles (with 4 corners) or any shaped polygon as long as they have 4 corners (rhombus and the likes)?

[Automatez]

You would need rectangles and they need to be square to the image (no rotation).