Four rounds of sharp review ... sounds very much like Felix Dittrich :-) Thank you, Saad Umer, for your contribution and your kind words. In terms of new developments, we are currently working on a layout detection model that is not only on-par with the current SOTA, but also robust to rotations. More info coming soon!
Recently a redaction use case landed on my desk i.e. blacking out sensitive info (names, salaries, account numbers) on documents before sharing. Since you're editing the user's original file, you need to know exactly where every word sits, down to the pixel. I surveyed the open-source OCR landscape and the bounding-box situation surprised me. The newest, most impressive models like Baidu's Unlimited-OCR released just weeks ago, DeepSeek-OCR output beautiful Markdown but zero word geometry. If the model can't tell you where a word is, you can't redact it. Mindee's docTR which is now maintained by t2k GmbH was the sweet spot. It provided a balance of accuracy with real word-level boxes, and a single straighten_pages=True flag that handles deskewing internally no DIY preprocessing pipeline. With one caveat i.e. those boxes came back in the internally straightened coordinate space, an image the user never sees. The transform was computed and thrown away. Best boxes in the ecosystem, unusable on the original document. So I fixed it upstream. My PR which is merged for docTR 1.1.0 adds preserve_original_coords=True: the straightening transform is captured as an invertible affine matrix and every geometry (words, lines, blocks, tables) is remapped back to the input image, validated to sub-pixel accuracy. Big thanks to Felix Dittrich for four rounds of sharp review. PR: https://lnkd.in/dgs6H9pv #OpenSource #OCR #ComputerVision #DocumentAI #Python