KANMultiSign: Kolmogorov-Arnold Networks for Efficient Sign Language Pose Animation
Researchers have introduced KANMultiSign, a novel framework designed to translate HamNoSys sign language notation into two-dimensional human pose sequences. This system addresses the challenge of creating scalable and accessible sign language animations by employing a coarse-to-fine generation strategy with multi-scale supervision. The model first establishes global structural coherence through an intermediate body-hand-face scaffold before refining fine-grained hand articulations for detailed finger movements. A key innovation involves integrating Kolmogorov-Arnold Network (KAN) modules into a Transformer backbone, utilizing learnable univariate function primitives to map discrete symbols to continuous kinematics efficiently. Experiments across Polish, German, Greek, and French sign language corpora demonstrate that KANMultiSign significantly reduces joint error compared to baseline models while using substantially fewer parameters. Ablation studies reveal that multi-scale supervision is the primary driver of accuracy improvements, whereas KAN integration offers a compact, efficient alternative for modeling. This research highlights a promising direction for efficient, high-quality sign language production systems, with code scheduled for public release to support further development in accessible technology.
Wire timeline
KANMultiSign: Kolmogorov-Arnold Networks for Efficient Sign Language Pose Animation
Researchers have introduced KANMultiSign, a novel framework designed to translate HamNoSys sign language notation into two-dimensional human pose sequences. This system addresses the challenge of creating scalable and accessible sign language animations by employing a coarse-to-fine generation strategy with multi-scale supervision. The model first establishes global structural coherence through an intermediate body-hand-face scaffold before refining fine-grained hand articulations for detailed finger movements. A key innovation involves integrating Kolmogorov-Arnold Network (KAN) modules into a Transformer backbone, utilizing learnable univariate function primitives to map discrete symbols to continuous kinematics efficiently. Experiments across Polish, German, Greek, and French sign language corpora demonstrate that KANMultiSign significantly reduces joint error compared to baseline models while using substantially fewer parameters. Ablation studies reveal that multi-scale supervision is the primary driver of accuracy improvements, whereas KAN integration offers a compact, efficient alternative for modeling. This research highlights a promising direction for efficient, high-quality sign language production systems, with code scheduled for public release to support further development in accessible technology.
cs.AI updates on arXiv.org