BSL-Static-48: A Signer-Independent Dataset of Anonymized Images and MediaPipe Hand Landmarks for Bangla Sign Language Recognition

BSL-Static-48 is a specialized Bangla Sign Language (BSL) dataset designed for Signer-Independent recognition. It encompasses 48 distinct classes, categorized into 10 digits (D0–D9) and 38 alphabets (L1–L38). This version (Version 2) implements a rigorous logic to prevent data leakage and ensure model generalizability.

Key Features & Methodology:
Data Source & Anonymization: Data was collected from 5 distinct volunteers (V01 to V05), totaling 14,568 original images. To ensure privacy, all images underwent facial blurring using an OpenCV-based DNN face detector, which did not compromise the accuracy of hand landmark detection.

Data Augmentation & Integrity: To double the dataset size and ensure hand-dominance invariance, horizontal mirroring was applied. To prevent Data Leakage, original images and their mirrored counterparts are strictly maintained within the same data partitions.

126-Dimensional Feature Extraction: Pre-processed features are provided in NumPy (.npy) format. Each sample is a 126-dimensional vector (21 landmarks per hand × 3 coordinates (x, y, z) × 2 hands). Landmark extraction using MediaPipe Holistic achieved a success rate of 99.97%.

Signer-Independent Protocol: This dataset provides a dedicated volunteer-based split:

Train Set (20,438 samples): Data from Volunteers V01, V02, V03, and V04.

Validation Set (2,956 samples): Data from Volunteers V01, V02, V03, and V04.

Independent Test Set (5,742 samples): Data strictly from Volunteer V05.

Transparency: Includes metadata CSVs mapping every feature file to its original volunteer and assigned split, along with quantitative verification reports.