Is it possible to not create overlapping bboxes, because that will help to identify the elements much easier. One such example like below -
This image was using Surya Streamlit version. But when I run the same pdf through marker, the extracted image is very different, and it is actually truncated. Marker and Surya, these two repos are in sync? (I think Surya repo is using more recent layout model than marker)
Is it possible to increase the bbox tolerance as an configurable argument, so, it can detect little more surrounding areas, when image detection is wrong and only gets cropped images
Is it possible to extract tables as images, rather than directly printing it on the md file, at least a configurable option? Because table detection is not top notch.
How to ensure, whatever is in table / image, text inside it is not again repeated in the md file?
homeant
commented
1 month ago
This image was using Surya Streamlit version. But when I run the same pdf through marker, the extracted image is very different, and it is actually truncated. Marker and Surya, these two repos are in sync? (I think Surya repo is using more recent layout model than marker)
Is it possible to increase the bbox tolerance as an configurable argument, so, it can detect little more surrounding areas, when image detection is wrong and only gets cropped images
Is it possible to extract tables as images, rather than directly printing it on the md file, at least a configurable option? Because table detection is not top notch.
How to ensure, whatever is in table / image, text inside it is not again repeated in the md file?