README.md

Common OCR Dataloaders

I created a collection of dataloaders for recognition task. Although detection should be easily implemented, this dataloaders only incorporate recognition (it's free, don't complain). For internal CVC distribution only amolina@cvc.uab.cat has the original .zip file containing all the data weakly curated.

Make sure to read Terms and Conditions for each dataset before using it.

Dataset Wiki:

• Esposalles: Historical catalan handwritten recognition with word and line level annotations. 4 75% cross validation sets are provided as well. (source)[http://dag.cvc.uab.es/the-esposalles-database/]

  • Train Words 75% of 23588 (4-Fold Cross validation)
  • Train Lines: 75% of 2288

• COCOText: English and non-English COCO Images annotated with OCR and manually corrected. Scene text recognition and a few handwritted images. (source)[https://bgshih.github.io/cocotext/]:

  • Train words: 163476 (97557 ilegible)
    1. English: 157579
    2. Not English: 5897
  • Val Words: 37650

• FUNSD: English documents for form text (printed and some handwritten) text recognition. (source)[https://bgshih.github.io/cocotext/]

  • Train entities: 22512
  • Test entities: 8973

• George Washington: English historical handwritten recognition with word and line level annotations (aachen splits provided). (source)[https://fki.tic.heia-fr.ch/databases/washington-database]

  • Train Words: 2433
  • Train lines: 325
  • Val lines: 168
  • val words: 1293
  • test words: 1168
  • test lines: 163

• HierText: Handwritten and printed scene text recognition with hierarchical annotation (paragraph, line, words). (source)[https://github.com/google-research-datasets/hiertext]

  • Train words: 925545
    1. Handwritten: 63100
    2. Other: 862445
  • Validation words: 191433
  • Train lines: 435138

• Historical Maps: English printed recognition in historical maps. (source)[https://weinman.cs.grinnell.edu/research/maps.shtml#data]

  • Train words: 75% of 25478 (4-fold cross validation)

• IAM: Good old IAM. Handwritten author-wise text recognition at line and word level. (source)[https://fki.tic.heia-fr.ch/databases/iam-handwriting-database] Using aachen partitions as found https://github.com/jpuigcerver/Laia/tree/master/egs/iam

  • Train (aachen split) words: 14331
  • Val (aachen split) words: 413
  • Test (aachen split) words: 1422

• IIIT5K: Scene text recognition in english. (source)[https://cvit.iiit.ac.in/research/projects/cvit-projects/the-iiit-5k-word-dataset].

  • Train words 3000
  • Test words 2000

• MLT19 Dataset: Multilingual scene text. (source)[https://rrc.cvc.uab.es/?ch=15]

  • Train words: 82662
    1. latin: 51296
    2. arabic: 4013
    3. chinese: 2545
    4. japanese: 4920
    5. korean: 5447
    6. bangla: 3169
    7. hindi: 3033
    8. symbols: 1295
    9. mixed: 179
    10. None: 6765
  • Val words: 29336 (from 4-fold cv, test evaluation can be found in rrc site)

• Parzival: Historical german binarized documents with word and line level annotation. (source)[https://fki.tic.heia-fr.ch/databases/parzival-database]

  • Train words: 5872
  • Test words: TBA
  • Validation words: 2936
  • Train lines: 2237
  • Test lines: TBA
  • Validation lines: 912

• Saint Gall: Historical handwritten latin documents at line level (there's word-level annotatins but they don't work for me :( ). (source)[https://fki.tic.heia-fr.ch/databases/saint-gall-database]

  • Train lines: 707
  • Test lines: 468
  • Validation lines: 235

• SROIE: Superset (i think) of FUNSD for receipts. (source)[https://rrc.cvc.uab.es/?ch=13]

  • Train entities: 34930
  • Test entities: 19386

• SVT: Street view images for scene text recognition. A link doesn't seem to work, idk why but i have the images. (source)[https://tc11.cvc.uab.es/datasets/SVT_1]

  • Train words: 647
  • Test words: 257

• TextOCR: OCR-Extracted annotations for VQA. (source)[https://textvqa.org/textocr/dataset/]

  • train words: 714770
  • val words: 107802

• TOTAL TEXT: Totally cool dataset for curved text recognition (interesting for domain shift reasons). (source)[https://github.com/cs-chan/Total-Text-Dataset/tree/master/Groundtruth]

  • Train words: 10589
  • Test words: 2547

• XFund: multilingual FUNSD with 199 pages per language. (source)[https://github.com/doc-analysis/XFUND/releases/tag/v1.0].

  • Training entities: 71999
    1. DE: 8632
    2. ES: 11449
    3. FR: 8816
    4. IT: 12215
    5. JA: 9005
    6. PT: 11654
    7. ZH: 10228
  • Validation entities: 25013

Note: Some measurements may be messed up, give me time I spent +3 whole days on that and i get easily overwhelmed :(