With M. Ahmed and S. I. Moyeen — paper under review. Started as my undergraduate thesis at RUET (Jan 2023 – May 2024), advised by Sumaya Ishrat Moyeen.

Problem & motivation

Bangla Sign Language has very few public resources, and what exists is mostly isolated signs: one clip, one label, neutral start and end poses. Real signing is continuous — signs blend into each other, and a usable system has to keep up with the signer in real time. We built both halves of the missing infrastructure: a continuous BdSL dataset (9,010 videos, 530 word classes) and a real-time recognition pipeline evaluated with latency as a first-class metric.

The dataset

The dataset has its own page — stats, methodology, benchmark protocol, and citation. It will be released alongside the paper.

Approach

The pipeline extracts holistic landmarks per frame (MediaPipe Holistic — hands, pose, and face), then LSTMs over the landmark streams produce a running word sequence, with an attention mechanism that improves sentence-formation accuracy in continuous signing. Landmarks rather than raw pixels keep per-frame cost low enough for real time on consumer hardware. Full training details and ablations are in the paper.

Pipeline diagram: video frames pass a hand-landmark extractor, the landmark sequence feeds a temporal model, and the model emits a running sequence of recognised words
Landmarks in, words out — per-frame cost stays low enough for real time. (Placeholder figure — final diagram coming with the paper.)

Results

Baseline results on the real-time benchmark (to be published with the paper)
ModelTop-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.

Demo

Recognition demo

A recorded demo of the full pipeline is planned once the paper is public. TODO: add YouTube/HF Space URL.

Meanwhile, the first stage runs live in your browser — try the hand-tracking demo on the dataset page. Nothing is uploaded; it’s the same landmark representation the recognition model consumes.

Limitations & what I’d do next