Line level bounding boxes

[lololouuuu]

I would like to know to which point the bounding box precision is something that can be parameterizable or configurable.
Using PaddleOCR on python API you can get pretty good bounging boxes at line level but in the README I see that the bounding boxes are paragraph level.

Thanks :)

[nataell95]

Hi! TurboOCR already returns line-level bounding boxes by default. The "paragraph-level" in the README only refers to the optional ?as_blocks=1 aggregation, which mirrors PP-StructureV3's parsing_res_list.
For tighter/looser polygons around each line, that's the DB unclip ratio (currently hardcode).
Happy to expose it (and the other DB thresholds) as query params if useful let me know.