A Multi-view Dataset for Vietnamese Word-Level Sign Language Recognition

------------------------------------------------------------------
Basic information
------------------------------------------------------------------

1. Journal article: A Multi-view Dataset for Vietnamese Word-Level Sign Language Recognition

2. DOI:10.5281/zenodo.17943573.

3. Contact information:
   Name: Matej Sindelar
   Institution: VSB – Technical University of Ostrava
   E-mail: matej.sindelar@vsb.cz
   ORCID: orcid.org

4. Dataset publication date: 2026-03-09

5. Place of publication: Ostrava, Czechia

6. Dataset Description

VSL400 is a standardized video dataset for Vietnamese Sign Language (VSL) word-level recognition. It comprises 74,259 manually annotated video clips representing 400 isolated glosses, performed by 28 signers. Each signing instance was recorded simultaneously from three synchronized RGB views (front, left, and right), enabling both single-view and multi-view learning.

All videos were processed using a consistent preprocessing pipeline to ensure uniformity for benchmarking. The clips have an average duration of 2.61 seconds, with a total recording time of approximately 53.99 hours. The dataset is organized into front_view, left_view, and right_view directories, where corresponding recordings share a common six-digit identifier and are accompanied by JSON metadata containing gloss labels and signer information.

Facial regions were de-identified and direct personal identifiers were removed. Because the videos contain human signing performances that may retain residual biometric and behavioral information, video access is managed through controlled access and requires agreement to the Data Usage Agreement, which is publicly available at: zenodo.org.

VSL400 is intended to support academic and research and development activities related to isolated VSL recognition, including RGB-based and pose-based modelling, single-view recognition, cross-view evaluation, missing-view robustness, and multi-view learning.