International Journal of Scientific Research and Engineering Development

This paper presents the VSL (Vietnamese Sign Language) dataset, a comprehensive collection of Vietnamese sign language videos designed for sign language recognition research with focus on vocabulary coverage. The dataset consists of 6,046 original video recordings covering 3,782 unique Vietnamese sign language gestures, collected from the QIPEDC project with appropriate permissions from 11 signers with diverse signing styles. Unlike previous studies, we propose a rigorous VocabularyCoverage-First evaluation protocol, using strategic stratified splitting (80% training, 10% validation, 10% test) based on original videos before data augmentation to completely eliminate data leakage. We apply 5 carefully selected transformation techniques to address class imbalance (64% of classes have only 1 sample), expanding the training set to 29,022 samples (4,837 original × 6 variants). Experimental results demonstrate realistic performance ranging from 42.18% (baseline LSTM) to 58.92% (Video Swin Transformer) on vocabulary-complete test sets. Critically, incorporating 468 facial landmarks to capture non-manual markers improves accuracy from 3.67% to 8.81% absolute gain, affirming the essential importance of these grammatical components in sign language. This dataset provides a solid and honest foundation for future Vietnamese sign language processing research, explicitly acknowledging the extreme challenge of 2,422 single-sample classes (38.42% accuracy) as the research bottleneck requiring Few-shot Learning approaches.

Ha Manh Dung, Nguyen Vo Hung, Nguyen Khanh Dang, Pham Thi Huong Nhai,"Towards Realistic Vietnamese Sign Language Recognition: A Large-Scale Dataset and Rigorous Evaluation Protocol" International Journal of Scientific Research and Engineering Development, V9(1): Page(454-468) Jan-Feb 2026. ISSN: 2581-7175. www.ijsred.com. Published by Scientific and Academic Research Publishing.