Overview

Most sign language datasets — and almost all Bangla Sign Language resources — contain isolated signs: one clean clip, one label. Real signing isn't like that. BdSL-530 was collected to close that gap for BdSL: 9,010 video clips covering 530 word classes, recorded as continuous signing so that models must handle coarticulation, transitions between signs, and signing speed — the things that actually make real-time recognition hard. The dataset ships with a benchmark protocol that scores both recognition quality and per-frame latency, because a sign language system that lags behind the signer is not usable.

The dataset accompanies our paper “Implementation of Bangla Sign Language (BdSL) Recognition in Continuous Real Time” (T. Hasan, M. Ahmed, S. I. Moyeen — under review).

Collection methodology

Full collection details — signer demographics, capture setup, annotation protocol, and quality control — are documented in the paper and will be expanded here once it is public. In summary:

Vocabulary: 530 word classes selected to cover everyday communicative needs. TODO: add the selection criteria from the paper.
Recording: continuous signing sequences rather than isolated, pre-segmented clips. TODO: camera setup, resolution, frame rate.
Signers: TODO: number of signers and demographics, pending paper release.
Annotation: word-level labels aligned to the video. TODO: annotation tooling and agreement protocol.

Sample clips

Muted, looping examples will appear here once sample clips are cleared for public release.

Greeting sequencesample clip pending release

Question phrasesample clip pending release

Numbers in contextsample clip pending release

Everyday vocabularysample clip pending release

Two-handed signssample clip pending release

Continuous sentencesample clip pending release

Benchmark

Baseline results on the real-time benchmark (to be published with the paper)
Model	Top-1 (%)	Latency (ms/frame)
MediaPipe Holistic + LSTM-attention (ours)	—	—
Frame-level CNN baseline	—	—

Benchmark numbers will be added when the paper is public. The protocol measures recognition quality and per-frame latency on continuous signing.

Try the hand-tracking demo

The recognition pipeline starts from hand landmarks. You can run that first stage right here: it loads MediaPipe in your browser and tracks your hands from your webcam. Everything runs locally — no video leaves your device, and nothing loads until you click.

Requires a webcam and a modern browser; the model (~6 MB) is fetched from a CDN when you start.

Download / access

The dataset will be released once the accompanying paper completes peer review. Until then, access requests are handled individually — email me with your affiliation and intended use. Email me with your affiliation and intended use.

License

To be finalised at release (research-only licence expected).

How to cite

If you use BdSL-530, please cite:

BibTeX

@misc{hasan2026bdsl530,
  author = {Hasan, Tahir and Ahmed, M. and Moyeen, S. I.},
  title  = {{BdSL-530}: A Continuous {Bangla Sign Language} Video Dataset},
  year   = {2026},
  note   = {Dataset accompanying ``Implementation of Bangla Sign Language (BdSL) Recognition in Continuous Real Time'', under review}
}