Current benchmarks for sign language recognition (SLR) focus mainly on isolated SLR, while there are limited datasets for continuous SLR (CSLR), which recognizes sequences of signs in a video. Additionally, existing CSLR datasets are collected in controlled settings, which restricts their effectiveness in building robust real-world CSLR systems. To address these limitations, we present Isharah, a large multi-scene dataset for CSLR. It is the first dataset of its type and size, collected in an unconstrained environment using signers' smartphones. This setup resulted in high variations of recording settings, camera distances, angles, and resolutions. This variation helps with developing sign language understanding models capable of handling the variability and complexity of real-world scenarios. The dataset consists of 30,000 video clips performed by 18 deaf and professional signers. Additionally, the dataset is linguistically rich as it provides a gloss-level annotation for all dataset's videos, making it useful for developing CSLR and sign language translation (SLT) systems. This paper also introduces multiple sign language understanding benchmarks, including signer-independent and unseen-sentence CSLR, along with gloss-based and gloss-free SLT.

Pose-based approaches for sign language recognition provide light-weight and fast models that can be adopted in real-time applications. This article presents a framework for isolated Arabic sign language recognition using hand and face keypoints. We employed MediaPipe pose estimator for extracting the keypoints of sign gestures in the video stream. Using the extracted keypoints, three models were proposed for sign language recognition: Long-Term Short Memory, Temporal Convolution Networks, and Transformer-based models. Moreover, we investigated the importance of non-manual features for sign language recognition systems and the obtained results showed that combining hand and face keypoints boosted the recognition accuracy by around 4% compared with only hand keypoints. The proposed models were evaluated on Arabic and Argentinian sign languages. Using the KArSL-100 dataset, the proposed pose-based Transformer achieved the highest accuracy of 99.74% and 68.2% in signer-dependent and -independent modes, respectively. Additionally, the Transformer was evaluated on the LSA64 dataset and obtained an accuracy of 98.25% and 91.09% in signer-dependent and -independent modes, respectively. Consequently, the pose-based Transformer outperformed the state-of-the-art techniques on both datasets using keypoints from the signer’s hands and face.