PRAMGCN-Net: 3D Human Pose Estimation With a Parameterized Routing Adjacency Modulation Graph Convolutional Network

Human pose estimation reconstructs 3D human joint positions from monocular images or videos, enabling applications in healthcare, sports, AR/VR, and animation. The challenge lies in depth ambiguity, temporal coherence, and complex motion patterns. Recent deep learning advancements, such as SMPLs, GCNs, and Transformers, have improved estimation precision. Most techniques evaluate poses in 2D before converting them to 3D, often ignoring temporal dynamics and encountering depth ambiguity, resulting in multiple 3D poses from a single 2D setup. While GCNs are effective for joint interactions, their dependence on pre-defined graph structures limits their capacity to capture global dependencies and adapt to varying joint configurations. To address these limitations, we proposed PRAMGraphConv, a novel graph convolution approach that dynamically models joint relationships. It is characterized by an expert routing block that dynamically assigns importance weights and alters joint relational patterns in an adaptable manner. Furthermore, the model enhances its ability to capture intricate, non-local linkages while maintaining processing efficiency, addressing depth ambiguity, and scaling issues, through the incorporation of an ANModGCN. We integrate these components into PRAMGCN-Net, a unified framework to achieve robust performance in challenging occluded scenarios. We evaluate our model through extensive experiments on standard benchmarks, including Human3.6M, HumanEva, and MPI-INF-3DHP. The results demonstrate that the proposed method outperforms several state-of-the-art approaches in 3D human pose estimation, while maintaining high accuracy with an efficient, lightweight architecture.