I'm working on a projet that uses PaddleOCR to get bounding boxes. It's far from perfect, but it's open source and good enough for our requirements. And it can mostly handle a 150 MB single-page PDF (don't ask) without completely keeling over.