IIIT-INDIC-HW-WILD

IIIT-INDIC-HW-WILD (Indic-HW-Wild) is a large collection of unlabeled, word-level handwritten images in nine Indic languages, gathered "in the wild" from handwritten notes, answer sheets and documents shared publicly on the internet, plus word images cropped from five existing page-level handwriting datasets. It was introduced in SemiHastakshar: Generalizable Indic Handwritten OCR through Semi-Supervised Learning (ICVGIP 2025) as the unlabeled corpus for semi-supervised handwritten text recognition.

• 2,450,068 word images, 9 languages, no transcriptions
• 2,348,149 images are included in this repository; 101,919 images from PHDIndic_11 and IIITS-THDB are not redistributed and are rebuilt from those datasets with the included script (see Rebuilding)
• Real-world variety: ruled and plain paper, scans and phone photos, uneven lighting, blur, bleed-through
• Project page: https://lalithaevani.github.io/SemiHastakshar-page/ · Code: https://github.com/LalithaEvani/SemiHastakshar
• The data is unlabeled. Some crops may contain partial words, several words, digits, symbols or non-text marks.

Contents

Language	Code	Word images	Collected by us	From existing datasets	In this repo	Rebuild from source
Hindi	hin	1,628,801	1,628,801	—	1,628,801	—
Telugu	tel	301,170	257,976	IIITS-THDB 37,787 · PHDIndic_11 5,407	257,976	43,194
Bengali	ben	163,492	39,711	BN-HTRd 83,006 · BanglaWriting 15,536 · Bongabdo 14,207 · PHDIndic_11 11,032	152,460	11,032
Tamil	tam	146,652	140,462	PHDIndic_11 6,190	140,462	6,190
Kannada	kan	73,323	71,628	PHDIndic_11 1,695	71,628	1,695
Gujarati	guj	60,852	46,022	PHDIndic_11 14,830	46,022	14,830
Malayalam	mal	57,062	50,800	PHDIndic_11 6,262	50,800	6,262
Punjabi (Gurmukhi)	pan	9,362	—	PHDIndic_11 9,362	—	9,362
Odia	ory	9,354	—	PHDIndic_11 9,354	—	9,354
Total		2,450,068	2,235,400	214,668	2,348,149	101,919

Odia and Punjabi come entirely from PHDIndic_11, so they are available only through the rebuild step.

Files

metadata.csv
checksums.sha256
rebuild_from_sources.py
images/bengali.tar  gujarati.tar  kannada.tar  malayalam.tar  tamil.tar  telugu.tar
images/hindi/hindi_00.tar … hindi_17.tar      (Hindi split into ~2 GB parts)

Each tar is uncompressed and holds <language>/<file>.png. Extract into images/ to get the paths used in metadata.csv:

huggingface-cli download evanilalitha/IIIT-INDIC-HW-WILD --repo-type dataset --local-dir IIIT-INDIC-HW-WILD
cd IIIT-INDIC-HW-WILD
sha256sum -c checksums.sha256
for t in images/*.tar images/hindi/*.tar; do tar -xf "$t" -C images/; done

To get a single language only, download metadata.csv and that language's tar(s) (e.g. --include "images/telugu.tar" "metadata.csv").

metadata.csv

Column	Description
file_name	Path after extraction, e.g. images/telugu/tel_iiit-indic-hw-wild_0000052.png
image_id	7-digit identifier
language, lang_code	Language name and ISO 639-3 code
dataset	Origin of the image: iiit-indic-hw-wild (collected by us) or the source dataset (bnhtrd, banglawriting, bongabdo, phdindic11, iiits_thdb)
license	License of that image (see below)
archive	The tar file that contains the image; empty for images rebuilt from source
source_file, x, y, w, h	Rebuilt images only: page file in the original dataset and the crop box in its pixels
pixel_sha256	Rebuilt images only: checksum used to verify the rebuilt image

File names follow {lang_code}_{dataset}_{image_id}.png. Images are PNG word crops of varying size; there are no labels.

Rebuilding PHDIndic_11 and IIITS-THDB images

The 101,919 word images cropped from PHDIndic_11 and IIITS-THDB are not included, because those datasets are distributed by their authors under their own terms. To obtain them:

1. Get the original datasets from their authors: PHDIndic_11 (IEEE DataPort) and IIITS-THDB (segment.zip).
2. Run, from the repository folder:

pip install numpy pillow opencv-python pandas
python rebuild_from_sources.py --metadata metadata.csv \
    --phdindic11 /path/to/PHDIndic_11 \
    --iiits-thdb /path/to/segment.zip

The images are written to their file_name paths under images/, next to the extracted tar contents, and each is verified against pixel_sha256. Use --languages odia punjabi to rebuild only some languages.

Licenses

This is a compilation. Each image carries the license of its origin (dataset / license columns). Filter on dataset to use only the parts whose terms fit your use.

dataset	Images	License / terms	Source
iiit-indic-hw-wild	2,235,400	CC BY 4.0	This work
bnhtrd	83,006	CC BY 4.0	BN-HTRd, Mendeley Data
bongabdo	14,207	CC BY 4.0	Bongabdo, UCI ML Repository
banglawriting	15,536	CC BY-NC 3.0 (non-commercial)	BanglaWriting, Mendeley Data
iiits_thdb	37,787	Not redistributed — rebuild from the original dataset; its terms apply (non-commercial research and educational use)	IIITS-THDB
phdindic11	64,132	Not redistributed — rebuild from the original dataset; its terms apply (available from its authors for research use)	PHDIndic_11, IEEE DataPort

Word images from BN-HTRd, Bongabdo and BanglaWriting were cropped by us from their page images and are redistributed here with attribution under their original licenses. For PHDIndic_11 and IIITS-THDB only crop coordinates are distributed. Rights in all source-dataset images remain with their authors; if you use them, cite the original dataset papers below. For commercial use, use only the CC BY 4.0 parts (iiit-indic-hw-wild, bnhtrd, bongabdo).

Citation

@inproceedings{evani2025semihastakshar,
  title     = {SemiHastakshar: Generalizable Indic Handwritten OCR through Semi-Supervised Learning},
  author    = {Evani, Lalitha and Mondal, Ajoy and Jawahar, C. V.},
  booktitle = {Proceedings of the Sixteenth Indian Conference on Computer Vision, Graphics and Image Processing (ICVGIP '25)},
  year      = {2025},
  publisher = {ACM},
  doi       = {10.1145/3774521.3774605}
}

If you use images from the source datasets, also cite:

• BN-HTRd: M. A. Rahman, N. Tabassum, M. Paul, R. Pal, M. K. Islam. BN-HTRd: A Benchmark Dataset for Document Level Offline Bangla Handwritten Text Recognition (HTR) and Line Segmentation. In Computer Vision and Image Analysis for Industry 4.0, CRC, 2023.
• BanglaWriting: M. F. Mridha, A. Q. Ohi, M. A. Ali, M. I. Emon, M. M. Kabir. BanglaWriting: A multi-purpose offline Bangla handwriting dataset. Data in Brief, 2021.
• Bongabdo: A. Ghosh. Towards Full-page Offline Bangla Handwritten Text Recognition using Image-to-Sequence Architecture. IEEE Silchar Subsection Conference (SILCON), 2023.
• PHDIndic_11: S. M. Obaidullah, C. Halder, K. C. Santosh, N. Das, K. Roy. PHDIndic_11: page-level handwritten document image dataset of 11 official Indic scripts for script identification. Multimedia Tools and Applications 77(2):1643–1678, 2018.
• IIITS-THDB: C. Andrew, S. Reddy, V. Pulabaigari, U. Pal. Text Independent Writer Identification for Telugu Script Using Directional Filter Based Features. ICDAR 2017 (Workshops), pp. 65–70.

Acknowledgements

Supported by MeitY, Government of India, through the NLTM-Bhashini project. Created at CVIT, IIIT Hyderabad.