Reproduce “Generalization of Text Line Segmentation for HTR in Historical Documents” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Gayan H. Pathirage, Stephan M. Unter, Simon Corbillé, Elisa H. Barney Smith
The authors evaluate text-line detectors across historical collections, measuring box overlap and whether crops support handwriting recognition. With CRNN recognizers trained separately for each dataset, they find Hi-SAM’s zero-shot crops have a 2.90-percentage-point average character error rate gap from ground-truth boxes across six out-of-domain datasets; 10 labeled target pages reduce it to 1.69 points. Results vary by collection, and layouts are simple.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.