TSL-ONE-S: A Real-World Thai Sign Language Dataset With Deep Learning Benchmarks

This paper presents TSL-ONE-S, a new word-level dataset for Thai sign language recognition, developed to address existing resource limitations and support communication for the deaf community. We present the creation, detailed annotation, and evaluation of a new dataset aiming to provide a resource for research on Thai sign language recognition. The dataset comprises 4,152 video samples covering 184 commonly used glosses, recorded from 29 signers with diverse demographics. The recordings were made in naturalistic settings with varied lighting and backgrounds, reflecting real-world conditions. Data were collected from nine schools across Thailand to ensure regional and dialectal diversity. A standardized annotation process was applied, including consistent labeling, unique signer identification, and automated video trimming using MediaPipe for pose and hand landmark detection, followed by manual validation to ensure accuracy. Signer-independent evaluations were performed using deep learning models, Inflated 3D Convolutional Neural Network (I3D), and Sign Pose-based Transformer (SPOTER), achieving top-1 classification accuracies for the TSL-ONE-S dataset of 79.69% and 80.83%, respectively. SPOTER, a pose-based transformer model, demonstrated faster training and greater robustness to environmental noise due to its skeletal-based representation.