VSL400: A Multi-view Dataset for Vietnamese Word-Level Sign Language Recognition

This research introduces VSL400, a multi-view video dataset created for the isolated wordlevel recognition of Vietnamese Sign Language (VSL). The dataset comprises 74,259 manually annotated video clips, encompassing 400 glosses, and was performed by 28 signers, including both deaf native users and trained hearing signers. Each sign instance was recorded concurrently from three synchronized RGB camera views (front, left, and right), thereby capturing the inherently three-dimensional character of sign articulation. To facilitate reproducible research, we provide a comprehensive preprocessing pipeline for boundary detection, temporal segmentation, and spatial normalization, along with structured metadata. All video data, annotations, and baseline processing code are publicly accessible under an open license. VSL400 addresses a significant resource deficiency for Vietnamese Sign Language and offers a practical reference dataset for the development and evaluation of robust recognition methods in low-resource environments.