PODM-mmHPE: Exploiting Physical Optics-Based mmWave Data Augmentation via Diffusion Model for Generalizable Human Pose Estimation

The scarcity of high-quality millimeter-wave (mmWave) radio frequency (RF) data for human sensing poses significant challenges to the generalization and robustness of deep learning models in this domain. The acquisition of RF data is constrained by the need for specialized equipment and specific application scenarios, resulting in a significant gap in both quantity and diversity compared to visual data. To address this critical bottleneck, we propose PODM-mmHPE, a novel framework that leverages Physical Optics (PO) and Diffusion Models (DM) to synthesize high-quality human RF data for mmWave Human Pose Estimation (mmHPE). Our framework introduces two key innovations: (1) a PO-based RF simulation pipeline that generates coarse RF data from 3D human mesh models by modeling electromagnetic wave interactions with body surface geometry and (2) a DM-based refinement module that transforms the coarse data into realistic RF heatmaps, capturing the complexity and noise of real-world RF signals. Furthermore, we develop map2pose, a Graph Convolutional Network (GCN)-based model that establishes an end-to-end mapping from RF heatmaps to 3D human poses, serving both as a downstream task benchmark and a validation mechanism for the quality of our synthetic data. Experiments demonstrate PODM-mmHPE’s effectiveness with 23.9% and 25.1% MPJPE reductions on HuPR and mmHPE datasets under synthetic augmentation. This validates our physics-guided synthesis enables generalizable mmWave pose estimation.