CODEWITHSUNDEEP
deep-learningobject-detectionbounding-boxyolov8roboflow
gülsemin
gülsemin

Reputation: 47

What kind of default image preprocessing is applied in Yolov8 model?

I am currently working with Ultralytics Yolov8 model for object detection. I trained my model with a custom dataset. I annotated my dataset on Roboflow and exported in Yolov8 format. The training result looks good. However there is a mismatch between the coordinates of bounding boxes of my manual annotation on Roboflow and that of prediction.

For example I have this image called image_1.jpg which is annotated on Roboflow and is part of training set. Roboflow defined the coordinates (as the part of the label):

[0 0.36953125 0.39609375 0.54765625 0.7875]

where the last 4 numbers showing the x1,x2,y1,y2 coordinates of the bounding boxes. However when I passed this image into predict method (just for checking) using the following piece of code

model = YOLO(‘best_weights.pt’)
result = model.predict(path_of_example_image, save = True)
boxes = result[0].boxes.xyxy

I got the coordinates:

tensor([[ 22.5558, 9.7429, 619.7601, 345.7614]])

When I draw the bounding box with the predicted coordinates, the area covers the object completely (with .99% accuracy on the (correctly) predicted class). On the other hand, when I draw the bounding box with the coordinates from manual annotation, the object is not covered at all as, obviously, all coordinates are below 1. So I just wonder what kind of preprocessing is applied within the model so that the coordinates from Roboflow's labelling make sense and my training result is actually good?

Upvotes: 0

Views: 2598

Answers (2)

fassage
fassage

Reputation: 26

Yolov8 tensor can output normalised values too. FYI, xyxy is not yolo format, you would need xywh instead. Just change this line:

boxes = result[0].boxes.xyxy

to:

boxes = result[0].boxes.xywh

Upvotes: 1

Brad Dwyer
Brad Dwyer

Reputation: 6504

The inputs are normalized values (range of 0 to 1) which means you need to multiply them by the width and height of the image to get pixel coordinates.

Upvotes: 1

Related Questions