Cross-View Isolated Sign Language Recognition with Graph

In the Cross-View Isolated Sign Language Recognition (CV-ISLR) Challenge, we propose a graph-based method for recognizing cross-view sign language. Our analysis reveals three key observations: (1) sign language recognition (SLR) requires greater emphasis on the connections between different parts of the human body; (2) relationships, such as those between facial expressions and hand gestures, are crucial for understanding the meaning of sign language; (3) spatiotemporal relationships, particularly between a signer's hands, exhibit consistency across viewpoints, which can aid recognition from new perspectives. Based on these findings, we model RGB and depth videos as graphs, where edges represent the relationships between different body parts and spatiotemporal features. To further enhance recognition performance, we adopt ensemble techniques that combine models trained on different modalities. Our method achieved a recognition accuracy of 33.97% on the RGB-D dataset and secured second place in the RGB-D track of the CV-ISLR challenge. This demonstrates the effectiveness of graph-based modeling for cross-view sign language recognition.