The rectangles of predictions shown in web are wrong, but the predictions are correct for the input image

[ErnestinaQiu]

The text area recognized by the OCR ml backend written by myself is misplaced on the web page, but the recognition result of the text content is correct. I use def convert_to_ls in https://labelstud.io/guide/predictions.html#Units-of-image-annotations to convert the unit.

[ErnestinaQiu]

[matharano]

I had the same issue until I realized the coordinate system of my model was different. It seems that label-studio uses X and Y values as the position of the top-left corner of the bounding box.

So, let's say you want to place a bounding box occupying the entire image. Then the coordinates would be:

• x = 0
• y = 0
• width = 100
• height = 100

At least it worked when I converted it to work that way.

[ErnestinaQiu]

Thanks for your sharing.
However, I tried another way to achieve it that I figured out the transformation functions between the wrong rectangle data I got from def convert_to_ls and the right data corrected mannually. It may sound a little strange and troublesome, but it works well.