Video 3D Human Pose Estimation Guided by Action Category Feature
Xueqing Ma · 2024
Video 3D human pose estimation plays an significant role in the field of computer vision. To address the ambiguity, the occlusion and the complexity problems, existing methodologies strive to leverage abundant context information. However, they failed to make full use of the action category information, which can provide useful context clues. Therefore, in this paper, we estimate 3D pose guided by the introduced category feature Within our framework, we devise two parallel extractors - one for the pose feature and the other for the category. Specifically, for the sake of the effectiveness, we further augmented the framework with a classification module supervising the category information. Guided by this reliable category features, our method enables accurate pose refinement facing complex samples. Our proposed approach achieves newly state-of-art performance on the authoritative dataset, i.e. Human3.6M.