关键词: 三维人体姿态估计, 无监督域适应, 生成对抗网络, 注意力机制

Abstract: Existing 3D human pose estimators perform well on a single dataset but are limited by the single pose structure of the training data, resulting in insufficient generalization to cross-domain experiments. Existing methods mitigate this deficiency by increasing pose diversity, but their generated poses often lack genuine validity. Moreover, there is still a significant gap between the global positions of poses in the target and source domains. To address these issues, a domain adaptation algorithm for 3D human pose estimation is proposed. Specifically, a spatial node attention module within the GAN network is introduced to encourage the generation of more natural human poses.In parallel, a pose position correction module is incorporated to facilitate alignment between the generated poses and the target data domain. Furthermore, a stochastic end-to-end mixing training strategy is proposed to enable the pose estimator to learn both old and new data information simultaneously, thereby enhancing the stability of the estimator's training. As a generative domain adaptation approach, the present algorithm can be efficiently applied to a variety of two-stage 3D human pose estimators. Through cross-scene and cross-dataset experiments, the results show that the proposed algorithm achieves the best current performance on several benchmark datasets. In particular, on the 3DHP dataset, the algorithm optimizes the MPJPE and AUC metrics by 1.7% and 1.4%, respectively, compared to the optimal work, verifying its effectiveness in improving the generalization of the 3D human pose estimator.

Key words: 3D human pose estimation, unsupervised domain adaptation, generative adversarial network(GAN), attention mechanism