Results of save_crop=True differs from result.boxes coordinates

Support YOLO
, ,
1

Hello!

I’ve been using YOLOv8 for a detection task. I want to process crops on-the-fly without saving crop images on the disk (it means i can`t use <save_crop=True> option).

However, the results of using the <save_crop=True> flag differ significantly from the bounding boxes in the result.boxes object.

You can see that in the attached images. On the left you can see results with save_crop=True, on the right you can see results of manually cropping original image with result.boxes coordinates:

Screenshot_3
Screenshot_31388×780 138 KB

This difference is crucial because the obtained crops from YOLO are then processed by another OCR model. The quality of the OCR is much better with the save_crop=True flag.
I’ve tried adding some pixels around the coordinates obtained from result.boxes, but the quality is still low. I think the problem is somehow related to resizing, but I can’t figure out how. I don`t resize original image before passing it down to YOLO.

This is how i get save_crop images:

results = yolo_model(image_name, save_crop=True, exist_ok=True, verbose=True, project="/work/", name="yolo_res", classes=[5,6,12])

This is how i get results.boxes images:

results = yolo_model(image_name, save_crop=False, save=False, verbose=True, classes=[5,6,12])
for bbox in yolo_results[0].boxes.xyxy.numpy():
       x_min, y_min, x_max, y_max = bbox
       cropped_image =  page_name.crop((x_min, y_min, x_max, y_max))

Please advise me on what could be the source of this problem and how I can fix it. TIA!

2

You can see the code here. It adds gain and pad.

The save_one_box function can also return the crop without saving:

from ultralytics.utils.plotting import save_one_box
crops = [save_one_box(b, results[0].orig_img, BGR=True, save=False) for b in results[0].boxes.xyxy]
2 Likes
3

Thank you so much! You saved my day!

1 Like