Bharat Scene Text Dataset (BSTD) v17.58
BSTD is a benchmark for Indian-language scene text: 6,582 scene images from Wikimedia Commons with 126,292 manually annotated text regions (polygons, transcriptions, and script labels). It covers 11 Indian languages (Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu) and English. About 19.8K further regions in other scripts, such as Urdu and Meitei, are annotated in the full images but have no word crops.
Project Links
Configurations
| Configuration | Train | Test | Task |
|---|---|---|---|
detection |
5,263 images | 1,319 images | Localize text regions in full images |
end_to_end |
5,263 images | 1,159 images | Localize and transcribe text in full images |
recognition |
77,715 crops | 28,746 crops | Transcribe a cropped word |
script_identification |
77,715 crops | 28,746 crops | Classify the script of a cropped word |
detection uses all source images with their original train/test splits.
end_to_end uses the source train split for training. Its test split is the 1,159 test
images whose text is entirely in the 12 supported scripts, matching the IJDAR end-to-end
test set. In both splits, regions whose text or script label is UNK are removed; images
left with no regions are kept with an empty annotation list.
recognition and script_identification read the same word-crop shards and differ only in
the label used. The paper reports 77,726 / 28,752 crops; 17 crops failed extraction and are
not included.
Fields
Full-image configs (detection, end_to_end)
| Field | Description |
|---|---|
image |
Scene image |
image_id |
Image identifier, e.g. A_image_1005; key into the license manifest |
image_name |
Original file path, e.g. A/image_1005.jpg |
| Region annotations | Per region: polygon as [[x, y], ...], transcription, and script label |
Word-crop configs (recognition, script_identification)
| Field | Description |
|---|---|
image |
Cropped word image |
| Transcription | Text of the word (label for recognition) |
script_language |
Script label, trimmed for filtering (label for script_identification) |
script_language_raw |
Script label exactly as in the original crop manifest |
source_image_id |
Full image the crop was cut from; key into the license manifest |
source_polygon_id |
Region within that image |
Load the Dataset
from datasets import load_dataset
repo_id = "Bhashini-IITJ/Bharat-Scene-Text-Dataset"
end_to_end_test = load_dataset(repo_id, "end_to_end", split="test")
recognition = load_dataset(repo_id, "recognition", split="train")
# All crops cut from one image (filters on the ID column without decoding images)
image_words = recognition.filter(
lambda sid: sid == "A_image_1005", input_columns="source_image_id"
)
Load Per-image License Metadata
import json
from huggingface_hub import hf_hub_download
license_path = hf_hub_download(
repo_id, "BSTD_v17_58_licenses.json", repo_type="dataset"
)
with open(license_path, encoding="utf-8") as file:
licenses = json.load(file)["images"]
image_license = licenses[end_to_end_test[0]["image_id"]]
crop_source_license = licenses[recognition[0]["source_image_id"]]
License and Attribution
Images come from Wikimedia Commons and keep their original licenses: mostly CC BY-SA
(3.0/4.0), CC BY, GODL-India, CC0 and public domain. None are NonCommercial or NoDerivatives.
BSTD_v17_58_licenses.json holds one record per image with its license, URL, author,
Commons page, restrictions and verification status. Join full-image rows by image_id and
crop rows by source_image_id.
When you redistribute images or crops, credit the author and license listed for each image.
Derivatives of ShareAlike, GFDL or Free Art License images must keep the same license. Some
images carry non-copyright restrictions noted on Commons (personality rights, trademarks,
insignia); these are listed in each record's restrictions field.
Annotations (polygons, transcriptions, script labels) are released under CC BY-SA 4.0.
To report a licensing issue or request removal of an image, contact the maintainers below.
License verification status
Licenses were checked against Wikimedia Commons on 2026-10-14. Statuses describe metadata only and are not legal clearance.
| Status | Images | Meaning |
|---|---|---|
verified_match |
5,537 | Archived license matches current Commons metadata |
verified_current_only |
788 | Taken from Commons; no archived record |
public_domain |
228 | Commons marks the image public domain |
archive_only_source_unavailable |
25 | Commons page not found; archived license only |
no_canonical_url |
2 | License has no standard URL |
license_changed |
1 | Archived and current licenses differ; both kept |
unresolved |
1 | No license information found |
Version
v17.58 updates the v17.57 annotations released on GitHub. It corrects misspelled script
labels (for example, engish → english); polygons and transcriptions are unchanged.
Citation
@article{de2026bharat,
title={{B}harat {S}cene {T}ext: a novel comprehensive dataset and benchmark for {I}ndian language scene text understanding},
author={De, Anik and Penamakuri, Abhirama Subramanyam and Yadav, Rajeev and Rathore, Aditya and Shah, Harshiv and Sharma, Devesh and Agarwal, Sagar and Kumar, Pravin and Mishra, Anand},
journal={International Journal on Document Analysis and Recognition (IJDAR)},
pages={1--19},
year={2026},
publisher={Springer}
}
Acknowledgement and Contact
Supported by MeitY, Government of India (S/MeitY/AM/20210114) under NLTM-Bhashini.
Contact: Abhirama Subramanyam, Anik De.
- Downloads last month
- 21