BSTD is a benchmark for Indian-language scene text: 6,582 scene images with 126,292 manually annotated text regions (polygons, transcriptions, and script labels). Labels cover English and Indic languages, including Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, and Telugu.

Project Links

Configurations

Configuration	Rows	Main fields
`detection`	5,263 train + 1,319 test images	Image, polygon coordinates, text regions
`script_identification`	One row per word crop (same crops as `recognition`)	Crop image, source image/polygon IDs, and script label
`recognition`	77,715 train + 28,746 test word crops	Crop image, transcription, language/script label
`end_to_end`	5,263 train + 1,159 test images	Full image, polygons, text, and script labels

Detection uses all source images. End-to-end training uses the source train split; testing uses the subset of test selection for supported languages. Script identification and text recognition use the word-crop Parquet shards.

Full-image rows include image_id and image_name. Crop rows include source_image_id and source_polygon_id, linking each crop back to its full-image annotation.

Loading and Per-image License Metadata

import json
from datasets import load_dataset
from huggingface_hub import hf_hub_download

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")

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"]]

The repository includes one license/provenance record per source image. Join full-image rows by image_id and crop rows by source_image_id. The dataset has mixed per-image image licenses; the manifest is metadata, not a dataset-wide license grant. Review the status, archive/current values, creator/credit, and restrictions for each image before reuse. Commons metadata does not establish which terms applied when an image was collected.

Per-image License and Attribution

The repository includes BSTD_v17_58_licenses.json, with one record per source image. Join full-image rows using image_id; join recognition/script-identification crops using source_image_id. The manifest keeps the selected license_name/license_url, archived values (archive_license_*), current Commons values (current_license_*), attribution details, and a per-image license_status.

Current status counts from the manifest:

Status	Images	Meaning
`verified_match`	5,537	Archived license name and URL match current Commons metadata
`verified_current_only`	788	License metadata recovered from Commons; no archive record
`public_domain`	228	Commons identifies the image as public domain and provides a rights basis
`archive_only_source_unavailable`	25	Archived claim retained; current Commons page was not found
`no_canonical_url`	2	License name is known, but there is no canonical license URL
`license_changed`	1	Archived and current Commons licenses differ; both values are retained
`unresolved`	1	No license information was found

These statuses summarize metadata and do not establish which terms applied when an image was collected or grant dataset-wide redistribution rights. Review the linked Commons page, creator/credit, and status for each image before reuse. In particular, do not treat archive_only_source_unavailable or unresolved images as cleared for redistribution.

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.

Citation

@article{de2026bharat,
  title={Bharat scene text: a novel comprehensive dataset and benchmark for indian 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: -

Total file size:

35.8 GB