Abstract:Existing Transformer models excel at modeling sequential dependencies, whereas they are less effective at capturing graph structures when dealing with non-Euclidean human skeletal keypoint data. By contrast, graph convolutional networks (GCNs) are well-suited for capturing topological relationships and local structural information among keypoints. Based on this, this study proposes a novel Transformer architecture——GET. The proposed method integrates GCN and Transformer into a unified framework that simultaneously models graph structures and global context, thus enhancing the representation learning of human skeletons and overcoming the limitation of existing Transformer-based 3D human pose estimation methods, where the Q, K, and V vectors in the self-attention mechanism are derived only through simple linear mappings. Furthermore, GET preserves the flexibility of Transformers while incorporating GCNs to encode structural priors of the human skeleton, allowing the model to jointly capture local joint relationships and global spatiotemporal dependencies. Extensive experimental results demonstrate that the proposed GET achieves state-of-the-art performance on the Human3.6M and MPI-INF-3DHP datasets, validating its effectiveness and generalization capability for modeling non-Euclidean structured data.