HopFIR：面向3D人体姿态估计的逐跳图Former与组内关节联合细化
HopFIR: Hop-wise GraphFormer with Intragroup Joint Refinement for 3D Human Pose Estimation
AI总结：
本文提出HopFIR架构，通过逐跳图Former发现潜在关节协同作用，并结合组内关节细化利用肢体先验信息，在Human3.6M上以32.67 mm的MPJPE大幅超越现有方法，且能显著提升SemGCN和MGCN等GCN方法的性能。
摘要：
2D到3D人体姿态提升是3D人体姿态估计（HPE）的基础，图卷积网络（GCNs）已被证明天然适用于对人体骨骼拓扑结构进行建模。然而，当前基于GCN的3D HPE方法通过聚合邻居信息来更新节点特征，而没有考虑不同关节协同作用中关节之间的交互。尽管一些研究提出引入肢体信息来学习运动模式，但关节之间潜在的协同作用（例如维持平衡）很少被研究。我们提出逐跳图Former与组内关节联合细化（HopFIR）架构来解决3D HPE问题。HopFIR主要由一个新颖的逐跳图Former（HGF）模块和一个组内关节细化（IJR）模块组成。HGF模块按k跳邻居对关节进行分组，并对这些组应用逐跳transformer式注意力机制，以发现潜在的关节协同作用。IJR模块利用先验肢体信息进行外围关节细化。大量实验结果表明，HopFIR大幅优于SOTA方法，在Human3.6M数据集上的平均每关节位置误差（MPJPE）为32.67 mm。我们还证明，最先进的基于GCN的方法可以从所提出的逐跳注意力机制中受益，性能显著提升：SemGCN和MGCN分别提高了8.9%和4.5%。
英文摘要：
2D-to-3D human pose lifting is fundamental for 3D human pose estimation (HPE), for which graph convolutional networks (GCNs) have proven inherently suitable for modeling the human skeletal topology. However, the current GCN-based 3D HPE methods update the node features by aggregating their neighbors' information without considering the interaction of joints in different joint synergies. Although some studies have proposed importing limb information to learn the movement patterns, the latent synergies among joints, such as maintaining balance are seldom investigated. We propose the Hop-wise GraphFormer with Intragroup Joint Refinement (HopFIR) architecture to tackle the 3D HPE problem. HopFIR mainly consists of a novel hop-wise GraphFormer (HGF) module and an intragroup joint refinement (IJR) module. The HGF module groups the joints by k-hop neighbors and applies a hopwise transformer-like attention mechanism to these groups to discover latent joint synergies. The IJR module leverages the prior limb information for peripheral joint refinement. Extensive experimental results show that HopFIR outperforms the SOTA methods by a large margin, with a mean per-joint position error (MPJPE) on the Human3.6M dataset of 32.67 mm. We also demonstrate that the state-of-the-art GCN-based methods can benefit from the proposed hop-wise attention mechanism with a significant improvement in performance: SemGCN and MGCN are improved by 8.9% and 4.5%, respectively.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
