A WiFi-based system for estimating three-dimensional human pose reported lower errors than DT-Pose in an indoor evaluation where it was used without additional training on the target environment. The proposed system reported a mean per joint position error (MPJPE) of 144.8 millimetres, compared with 162.3 mm for DT-Pose. Its Procrustes Aligned MPJPE (PA-MPJPE), a structural measure calculated after global rotation, translation and scale are removed, was 105.5 mm versus 114.5 mm. The paper also reported 93.3% PCK@50 and described the two error results as improvements of 10.8% and 7.9%.
The document is an arXiv version 1 preprint dated 25 Aug 2026. MPJPE is the primary accuracy metric, while PA-MPJPE evaluates structural pose quality after global rotation, translation and scale are removed.
How the model is built
The framework combines three named elements: Physics-Guided Multipath-Aware Attention, Hierarchical Cross-Attention Fusion and Physics-Disentangled Domain Adaptation. The paper assigns these components to feature extraction, ambiguity resolution and cross-environment generalization. Training jointly minimizes pose regression, action classification and contrastive domain-alignment losses.
An information-theoretic analysis in the paper supports those design choices: it argues that physics-guided representations can retain more information about pose, action conditioning can reduce uncertainty about the pose given the action, and contrastive alignment can encourage features to cluster in an environment-invariant way. The analysis is theoretical, depends on modeling assumptions and does not report empirical mutual-information estimates.
Results across the datasets
On Person-in-WiFi-3D, the model reported 84.98 mm MPJPE and 55.5 mm PA-MPJPE, compared with DT-Pose's reported MPJPE of 90.0 mm. It also reported 92.0% PCK@50 and 74.1% PCK@20.
On the MM-Fi comparisons labeled P1, it again reported lower figures than DT-Pose. In S1, MPJPE was 156.2 mm and PA-MPJPE 72.4 mm, versus 166.5 mm and 104.1 mm for DT-Pose. In S2, the corresponding figures were 190.3 mm and 76.6 mm, compared with 215.1 mm and 106.6 mm.
In the MM-Fi cross-environment tests labeled S3, PA-MPJPE was 78.9 mm in P1, 77.8 mm in P2 and 80.4 mm in P3, while the cited DT-Pose P1 value was 106.2 mm. The analysis also notes that absolute MPJPE and PCK were not uniformly better than DT-Pose across all S3 protocols.
The evaluation covered three sources of data. Person-in-WiFi-3D included seven volunteers performing eight actions across three locations, with approximately 90,000 training samples, 7,800 test samples and 14-keypoint poses. MM-Fi included 40 subjects, 27 actions, four environments, 320,000 frames and 17-keypoint annotations. The self-collected real-world dataset covered eight actions across two indoor scenarios and 28,406 aligned frames, with WiFi sampled at 600 Hz and ground truth at 30 frames per second.
What the numbers leave open
No confidence intervals or repeated-run variability were reported for the comparisons. The real-world zero-shot result came from one self-collected deployment dataset, and no uncertainty estimate was reported for it.
An additional test examined imperfect action guidance. With action accuracy set at 60%, the model reported 186.8 mm MPJPE, compared with 189.5 mm without semantic guidance. Because the study used controlled action-accuracy injection, the result does not show how naturally occurring recognition errors would behave.
In an MM-Fi ablation, the full model reported MPJPE of 157.8, 190.9 and 335.5 mm across P3S1, P3S2 and P3S3. The variant without PGMA reported 168.9, 198.3 and 358.4 mm; the variant without the action branch reported 189.5, 224.5 and 363.9 mm.
The reported computational profile was described as consistent with real-time inference in the tested setup: 13.87 million parameters, 0.039 GFLOPs per inference and approximately 258.4 frames per second on the named GPU. The analysis flags inconsistent hardware naming elsewhere in the document and says no independent profiling was reported.
Paper data and sources
Original title: Physics-Informed WiFi Sensing for Robust 3D Human Pose Estimation in Mobile and Cross-Environment Settings
Authors: Kaixuan Huang, Yuanbo Chen, Guangjin Pan et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text