Probability of correcting issues with additional training or tuning

[gpadvorac]

We have an app where it's absolutely critical that the bounding boxes are correct however sometimes they are way off as shown in this screenshot:

https://snipboard.io/6ckDKC.jpg

We are using DocLayout-YOLO based on YOLOv10.

What are the chances of being able to tune or train the model to get this correct and not make these kind of mistakes? I can probably produce 1000 pages that have mistakes like this which could be annotated.

Is it reasonable to expect to overcome this problem?

[JulioZhao97]

Hello! This kind of problem is the x-axis shrinking problem, which is largely solved in pypi version, for details you acn refer to #96.

Specifically, you should use code in pypi (do not use code in github), and use 2501 model