Podcast · Past conversation
Document Layout with PP-DocLayoutV3
31 min · Watch on YouTube ↗
Joe Barrow and I read and discussed the RT-DocLayout paper behind PP-DocLayoutV3. We worked through the methods, looked at the experiments and examples, and talked through our questions about the research.
We focused on how the model finds document regions, puts them in reading order, and handles warped, curved, or rotated pages. We also discussed where we’d use these ideas in document pipelines.
What we discussed
- Boxes: How it finds and classifies document regions (paragraphs, headings, tables, etc.)
- Reading order: How it compares pairs of regions to work out a reading sequence in complex layouts.
- Warping and rotation: How the segmentation approach works, and how training sets were created to make it generalize better to distorted pages and curved, warped, and tilted documents.
Explore PP-DocLayoutV3 and its examples ↗
Pre-talk context
Curved text and segmentation
Why a rectangle can include neighboring text, and how PP-DocLayoutV3 follows the shape of a region.
Context from the paper
PP-DocLayoutV3 is the layout stage of an OCR pipeline and the curved outlines are segmentation boundaries. The paper tests document images, but there are many other use cases for this feature
Read the source ↗Putting blocks in reading order
How the model compares pairs of blocks to order a page.
Context from the paper
PP-DocLayoutV3 finds document regions and predicts reading order together. It compares pairs of regions in parallel, estimates how many regions come before each one, and sorts them into a sequence. The paper’s experiment found learning these tasks together reduces reading-order errors compared with training the ordering model separately.
Read the source ↗