Answers are grounded in this paper and cite their sources.
Review tracks thirty years of sign language recognition advancements, revealing major technical gaps in continuous sentence processing and non-manual cue integration.
Over the past three decades, the focus of research in sign language recognition (SLR) and sign language translation (SLT) has shifted significantly. Early research relied primarily on wearable hardware sensing solutions such as wired data gloves and electromagnetic trackers, while the current trend has shifted to data-driven, vision-based deep learning frameworks such as 3D convolutional neural networks (3D-CNNs) and label-free pose estimation have been extremely helpful. This article attempts to review the evolution of SLR and translation over the past thirty years, focusing primarily on Australian Sign Language (Auslan). The article will explore how foundational technologies have evolved into specialized models where they address the unique challenges of sign language, using models such as SAM-SLR and SignVTCL. Finally, the article will discuss some problems with current evaluation methods. A significant issue is the large performance gap between recognizing individual gestures and recognizing consecutive sentences. One issue is co-articulation which refers to signs influence on each other when produced in sequence, causing assimilations or drops that puzzle recognition models. Another is the presence of non-manual signals (facial expressions, head movements) which carry crucial grammatical information (questions, negation, etc.) but are not explicitly annotated in gloss sequences. In addition, the differences in speech rate and style among sign language interpreters are substantial; for example, the speech rate in news videos is much faster than in the lab dataset, posing a significant challenge to the model's time alignment capabilities. Furthermore, currently used word-translation-based evaluation criteria struggle to fully reflect the rich meaning of Auslan.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: