Sign Language Recognition (SLR) faces dual challenges of variable signing speeds and spatial intricacy. Indian Sign Language (ISL) is inherently bimanual, yet progress is hampered by a lack of large‑vocabulary benchmarks and the inability of conventional models to cope with irregular signing velocities. To tackle this, we propose a structurally‑aware hybrid framework for efficient gloss‑level ISL recognition.
An entropy‑driven SVD sampler dynamically extracts informative keyframes from video streams, reducing computational redundancy. Because this subsampling creates uneven physical time gaps, the architecture incorporates Continuous Dual Positional Encoding together with a distance‑aware Elastic Transformer to preserve the natural signing rhythm.
On the spatial side, a Kinematically‑Conditioned Adaptive GCN leverages unified bone‑joint features to explicitly model bimanual geometry. To validate the approach, we build a native ISL dataset comprising 120 classes and 9,453 videos, and apply Physical Node Dropout and ArcFace optimization to mitigate subject‑specific memorization. The system achieves 89.8% accuracy, establishing a strong baseline for complex sign language modeling.
Review