Gesture skeleton recognition and key technique node automatic annotation for traditional craftsmanship based on improved temporal graph convolutional network
Year: 2026
Type: article
Abstract: Traditional craftsmanship, as a critical component of intangible cultural heritage, encodes essential procedural knowledge within the hand gestures of skilled artisans. Digitally preserving these fine-grained manipulation techniques demands robust recognition and precise temporal annotation methods that existing approaches have not adequately addressed. This paper proposes an improved temporal graph convolutional network (IT-GCN) for traditional craft hand gesture skeleton recognition and key technique node automatic annotation. The proposed architecture incorporates three enhancements over standard spatiotemporal graph convolution: adaptive graph topology learning that captures latent inter-joint dependencies beyond fixed anatomical connectivity, multi-scale temporal convolution that accommodates heterogeneous motion rhythms, and channel attention recalibration that suppresses redundant features. Building upon the learned frame-level representations, a two-stage annotation strategy combining feature-space divergence peak detection with classification-confidence verification is introduced to automatically identify key technique nodes—temporal instants where meaningful skill transitions occur. A dedicated dataset, CraftGesture-18, comprising 2,160 skeleton sequences across 12 gesture categories from ceramic pottery, bamboo weaving, and wood carving, is constructed for evaluation. Experimental results show that IT-GCN reaches 93.7% recognition accuracy, which places it 1.3 percentage points above BlockGCN, the strongest baseline in our comparison, and 2.7 points above the widely used CTR-GCN reference. The annotation module attains an F1-score of 84.5% against the consensus ground truth and 82.0% when it is scored against a fully independent expert annotation, with a mean temporal offset of 2.6 frames. Because a held-out craft contributes gesture categories that the classifier has never seen, cross-craft transfer additionally requires a small labelling budget, and we quantify how recognition and annotation grow with it. Ablation analysis confirms the complementary contributions of each proposed component, with adaptive topology learning providing the most substantial individual gain.
Source: Scientific Reports
Author: Ye Yuan
Institution: Central South University
Language: English
Subfield: Human-Computer Interaction
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7203835574
FWCI:
Cites:
Cited by:
Related to: