团队简介    研究论文    研究人员    发明专利    研究资助    联系我们    English
Sequential 3D Human Pose Estimation Using Adaptive Point Cloud Sampling Strategy
International Joint Conference on Artificial Intelligence, 2021

Zihao Zhang, Lei Hu, Xiaoming Deng, and Shihong Xia

Abstract
3D human pose estimation is a fundamental prob-lem in artificial intelligence, and it has wide ap-plications in AR/VR, HCI and robotics.How-ever, human pose estimation from point clouds stillsuffers from noisy points and estimated jittery ar-tifacts because of handcrafted-based point cloudsampling and single-frame-based estimation strate-gies. In this paper, we present a new perspective onthe 3D human pose estimation method from pointcloud sequences. To sample effective point cloudsfrom input, we design a differentiable point cloudsampling method built on density-guided attentionmechanism. To avoid the jitter caused by previ-ous 3D human pose estimation problems, we adopttemporal information to obtain more stable results.Experiments on the ITOP dataset and the NTU-RGBD dataset demonstrate that all of our con-tributed components are effective, and our methodcan achieve state-of-the-art performance.
Downloads

Sequential 3D Human Pose Estimation Using Adaptive Point Cloud Sampling.  International Joint Conference on Artificial Intelligence, 2021 [PDF] [Project]