Continuous Sign Language Recognition (CSLR) typically relies on sentence-level annotations, which are costly and difficult to obtain. In this paper, we propose a unified boundary-sensitive Informer–GCN framework that enables CSLR using supervision derived exclusively from isolated sign datasets. The proposed architecture integrates (i) a boundary-aware Informer encoder regularized by smoothness, curvature, prediction, and consistency constraints to encode transition dynamics, (ii) a relationally adaptive Graph Convolutional Network (GCN) constructed using signal-aware adjacency modeling based on PCC, PLV, COH, frequency, and distance features, and (iii) a Cross-Domain Diffusion with Relational Cross-Attention (CD-DRCA) mechanism that refines graph topology through diffusion-based relational denoising. The model is trained solely on isolated sign samples and evaluated on continuous sequences using a sliding-window inference strategy with a Softmax confidence threshold and duplicate-suppression post-processing to stabilize boundary detection. Extensive experiments on ASLLRP and PSLS datasets demonstrate the effectiveness of the proposed approach under CSLR problem. Specifically, under random split of datasets, the proposed framework achieves 90.80% accuracy on ASLLRP and 92.6% on PSLS. Additionally, under signer-independent evaluation, the proposed framework achieves 88.9% accuracy on ASLLRP, outperforming the strongest baselines, and 91.3% on PSLS while maintaining strong robustness across signers. Moreover, robustness analysis shows that the proposed CD-DRCA reduces performance degradation under 20% keypoint masking from 7.9% (fixed graph) to 4.2%Ḟurthermore, statistical significance testing confirms that the improvements are consistent (p < 0.05). These results demonstrate that boundary-sensitive temporal modeling combined with diffusion-refined relational learning provides a practical, robust, and scalable solution for CSLR without requiring sentence-level supervision.
Rastgoo et al. (Mon,) studied this question.