Person-in-WiFi 3D: Unified Model for 3D WiFi Perception

WiFi sensing offers a privacy-preserving and occlusion-robust alternative to camera-based human perception, making it attractive for applications such as smart homes, elderly care, and virtual reality. Although recent studies have achieved remarkable progress in WiFi-based single-person 2D/3D pose estimation, multi-person 2D pose estimation, and single-person mesh reconstruction, multi-person 3D human perception remains largely unexplored. In this paper, we present Person-in-WiFi 3D, the first WiFi sensing system capable of end-to-end multi-person 3D pose estimation and mesh reconstruction. Person-in-WiFi 3D advances prior work in two aspects. First, it employs multiple WiFi transmitters and receivers to capture richer 3D spatial reflections from multiple individuals, enabling robust perception in complex multi-person scenarios. Second, it introduces a DETR-inspired architecture with Hungarian matching and a coarse-to-fine hierarchical refinement strategy that progressively exploits global, instance-level, and keypoint-level features for accurate human localization and reconstruction. These designs enable efficient, scalable, and high-fidelity 3D human perception.To support research in this area, we construct Wiception3D, the first large-scale dataset for WiFi-based multi-person 3D human perception. The dataset contains over 97,000 frames collected from seven participants across diverse multi-person activities in a 4 m × 3.5 m environment, with annotations for both 3D pose estimation and 3D mesh reconstruction. Experimental results show that Person-in-WiFi 3D achieves a mean 3D keypoint localization error of 93 mm and a mesh reconstruction error of 41 mm, delivering performance comparable to camera- and millimeter-wave radar-based systems while retaining the inherent advantages of WiFi sensing.