Latent Diffusion-Inspired Domain-Invariant Human Pose Estimation from WiFi Signals

WiFi-based human pose estimation offers a privacypreserving alternative to vision-based methods but faces significant challenges under domain shifts caused by environment and user variability. We propose DiffPose<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>, a diffusion-inspired, domain-invariant regression framework that combines adapterenhanced U-Net architecture with uncertainty-guided learning and latent feature alignment for zero-shot cross-domain generalization. DiffPose incorporates three key components: (1) an adapter-equipped U-Net that predicts 2D heatmaps, log-variance uncertainty maps, and compact domain-invariant features; (2) a composite regression objective combining uncertainty-aware residual loss (RLE), pose adjacency constraints (PAM), and keypoint localization; and (3) a dual-domain alignment strategy using cosine similarity in the latent space and moment matching of projected features. Experiments on the MM-Fi benchmark show that DiffPose achieves 0.734 PCK@ 0.8 in unseen environments while also delivering improved uncertainty calibration (NLL = -1.446) and $\mathbf{1 4 \%}$ higher accuracy under cross-domain transfer. Our approach offers strong generalization as a practical and robust solution for privacy-sensitive human sensing.<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>Code is available in https://github.com/WiFiCSI-Video/DiffPose