ClapperText: A Benchmark for Text Recognition in Low-Resource Archival Documents

ClapperText Dataset Overview

ClapperText is a benchmark dataset for text detection and recognition in visually degraded and low-resource archival conditions. It is derived from 127 World War II–era film segments containing clapperboards that record structured production metadata such as date, location, and camera operator identity.

The dataset contains:• 9,813 annotated frames• 94,573 word-level text instances• 67.4% handwritten words and 1,566 partially occluded instances• Both full-frame and cropped word images• Precise 4-point polygon annotations to support spatially accurate OCR applications

ClapperText is designed to advance robust optical character recognition and document understanding in historical and visually challenging archival materials.

Folder Meaning

detection/imgs – Full-frame images for text detection• Subfolders train, val, test: disjoint video splits• test_keyframes: manually verified frames used in reported benchmarks

detection/annos – JSON annotations following the same split structure

recognition/imgs – Cropped word images derived from the detection data• train, val: disjoint video splits• test_keyframes: manually verified frames used in reported benchmarks

recognition/annos – Corresponding CSV annotations for cropped word images

Each subfolder (e.g., 8332_2_46_100_T) represents:• 8332 → Segment index in the HISTORIAN source dataset• 2 → Shot number within that segment• 46, 100 → Start and end frame indices• T → Shot type (here, T = text)

Data Summary

Split
Videos
Frames
Word Annotations
Handwritten (%)
Occluded (%)

Train
18
1,122
17,749
72.3
4.8

Val
8
527
4,983
67.6
1.6

Test
101
8,164
71,841
66.2
0.9

Total
127
9,813
94,573
67.4
1.7

Recommended Usage

• Use detection/test_keyframes and recognition/test_keyframes for validation to ensure comparability with benchmark results reported in the paper.• Training is limited (18 videos) to reflect low-resource historical scenarios.• Evaluation metrics:– Detection: Hmean @ IoU = 0.5– Recognition: Word Recognition Accuracy (case- and symbol-normalized)

Acknowledgements

Supported by the Austrian Science Fund (FWF) – doc.funds.connectProject No. DFH 37-N: Visual Heritage: Visual Analytics and Computer Vision Meet Cultural Heritage.

More Information

Further details are available at:https://github.com/linty5/ClapperText

For questions, please contact: [e-mail]