HFPose: Hierarchical Masking and Global-Local Fusion for 3D Human Pose Estimation

Although monocular 3D human pose estimation has progressed substantially in recent years, it still struggles under occlusion conditions because the hierarchical motion dependencies among joints and the discriminative fusion of joint features are often overlooked. Motivated by these challenges, this paper proposes Hierarchical Masking and Global-Local Fusion for 3D Human Pose Estimation (HFPose), which introduces a Hierarchical Masking (HM) strategy and a Global-Local feature Extraction and Fusion (GLEF) module. The HM module partitions joints into three hierarchical levels according to the kinematic chain and performs randomized masking during training, thereby augmenting occlusion-related samples and enhancing spatial modeling performance under joint occlusion conditions. The GLEF module extracts both local and global joint features and introduces scalable adaptive fusion weights to mitigate the imbalance between local and global information, thereby improving the model’s ability to capture complex spatial dependencies among joints. Experiments on Human3.6M, MPI-INF-3DHP, and 3DPW datasets validate the effectiveness of HFPose. Code is available at https://github.com/Lxg-233/HFPose.