Post
61
Reading order: the quiet problem after you've found the lines 📜
On a historical page with more than one column, finding the text lines is largely solved. Deciding the order to read them in is not, and that order is what a recogniser is ultimately given. On multi-column pages a perfectly recognised page can still come out in the wrong order.
I built a small, learning-free step that recovers the order, and benchmarked it against htrflow, Microsoft's Table Transformer and PaddleOCR on 198 real handwritten tables (the USS Jeannette Arctic logbook).
A few honest findings:
- On dense handwritten tables, most detectors barely find the lines at all, so detection, not ordering, is the harder part there.
- Given the lines, the simple step is level with much heavier table models, and they all reach the same ceiling. A shared open problem, not solved by any one of them.
Try it: abhishekjha1008/historical-reading-order-demo
Full study: https://valhtrdata01.z1.web.core.windows.net/
Dataset and code: abhishekjha1008/historical-table-reading-order
Comments and pointers welcome.
#HTR #OCR #DigitalHumanities #HistoricalDocuments #LayoutAnalysis
On a historical page with more than one column, finding the text lines is largely solved. Deciding the order to read them in is not, and that order is what a recogniser is ultimately given. On multi-column pages a perfectly recognised page can still come out in the wrong order.
I built a small, learning-free step that recovers the order, and benchmarked it against htrflow, Microsoft's Table Transformer and PaddleOCR on 198 real handwritten tables (the USS Jeannette Arctic logbook).
A few honest findings:
- On dense handwritten tables, most detectors barely find the lines at all, so detection, not ordering, is the harder part there.
- Given the lines, the simple step is level with much heavier table models, and they all reach the same ceiling. A shared open problem, not solved by any one of them.
Try it: abhishekjha1008/historical-reading-order-demo
Full study: https://valhtrdata01.z1.web.core.windows.net/
Dataset and code: abhishekjha1008/historical-table-reading-order
Comments and pointers welcome.
#HTR #OCR #DigitalHumanities #HistoricalDocuments #LayoutAnalysis