Research map / 3D vision / Human pose estimation
Person-in-WiFi 3D: End-to-End Multi-Person 3D Pose Estimation with Wi-Fi
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces Person-in-WiFi 3D, a Transformer-based DETR-style end-to-end system that estimates multi-person 3D poses from Wi-Fi CSI using four laptop receivers/transmitters, trained with Kinect SDK pseudo-labels on a new 97K-sample dataset.
Problem
Multi-person 3D pose estimation from Wi-Fi signals was unsolved; prior Wi-Fi methods handled single-person 2D/3D or multi-person 2D only, and naive 3D extensions of bottom-up heatmap/PAF methods were storage-heavy, slow, and failed to converge.
Approach
Treats multi-person Wi-Fi pose estimation as set prediction: a CSI tokenizer with spatial-temporal embeddings and Transformer encoder feeds a DETR-style pose decoder with learnable queries (with Deformable-DETR-style cumulative offsets) and a PETR-inspired refine decoder, trained with Hungarian matching and focal loss.
Key results
- MPJPE of 91.7mm (1-person), 108.1mm (2-person), and 125.3mm (3-person) on the self-collected dataset; overall 107.2mm.
- Reproduced WiPose achieves 101.8mm MPJPE vs 91.7mm for Person-in-WiFi 3D on single-person data.
- Using 3 receivers gives 107.2mm MPJPE versus 148.5 to 173.0mm with fewer receiver configurations; refine decoder lowers MPJPE from 116.1 to 107.2mm.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Person-in-WiFi 3D instead of WiPose | Single-person 3D pose estimation · 1-person | MPJPE: 101.8 -> 91.7 | Helped |
| Add a Refine decoder (ablation) | Person-in-WiFi 3D · All test samples | MPJPE: 116.1 -> 107.2 | Helped |
| Add a Refine decoder (ablation) | Person-in-WiFi 3D · All test samples | MPJPE(v): 46.9 -> 49.4 | Hurt |
| Reduce the receiver configuration to R1+R2 (ablation) | Person-in-WiFi 3D · All test samples | MPJPE: 107.2 -> 148.5 | Hurt |
| Reduce the receiver configuration to R1+R3 (ablation) | Person-in-WiFi 3D · All test samples | MPJPE: 107.2 -> 142.8 | Hurt |
| Reduce the receiver configuration to R2+R3 (ablation) | Person-in-WiFi 3D · All test samples | MPJPE: 107.2 -> 144.2 | Hurt |
| Use receiver R2 only (ablation) | Person-in-WiFi 3D · All test samples | MPJPE: 107.2 -> 173.0 | Hurt |
| Add phase information to amplitude (ablation) | Person-in-WiFi 3D · All test samples | MPJPE: 192.4 -> 137.7 | Helped |
| Denoise phase information (ablation) | Person-in-WiFi 3D using amplitude and phase · All test samples | MPJPE: 137.7 -> 107.2 | Helped |
| Evaluate with a held-out person | Person-in-WiFi 3D · Cross-Person | MPJPE: 107.2 -> 131.6 | Mixed |
| Evaluate with a held-out environment | Person-in-WiFi 3D · Cross-Environment | MPJPE: 107.2 -> 626.4 | Hurt |
Surprising findings
- Naive 3D extension of Person-in-WiFi (3D heatmaps and PAFs) failed to converge and took about 6 hours per epoch, and a direct PETR application yields 219.0mm MPJPE.
- Cross-environment generalization degrades sharply, with MPJPE rising from 107.2mm to 626.4mm.
Limitations
- Ground-truth labels come from Azure Kinect Body Tracking SDK, which caps accuracy and performs poorly with more than 3 people.
- Cross-environment performance drops sharply (MPJPE 626.4mm) with no adaptation strategy.
- Hand and arm joints have large errors (182-203mm), and dataset diversity of arm poses is limited.
Experimental setup
Datasets: self-collected Wi-Fi CSI and Azure Kinect RGB-D dataset (97K+ samples, 7 volunteers)
Benchmarks: Person-in-WiFi 3D self-collected dataset
Metrics: MPJPE MPJDLE (horizontal, vertical, depth) cumulative distribution of MPJPE
Baselines: WiPose Person-in-WiFi (extended to 3D) PETR
Builds on: DETR Deformable DETR PETR PhaseFi Person-in-WiFi
Closest papers by meaning
- Interleaved Selective State Space Models for Efficient WiFi-Based 3D Multi-Person Pose Estimation (ICML 2026 · similarity 0.90)
- Through-Wall Human Pose Estimation Using Radio Signals (CVPR 2018 · similarity 0.81)
- SelfPose3d: Self-Supervised Multi-Person Multi-View 3d Pose Estimation (CVPR 2024 · similarity 0.80)
- SAT-HMR: Real-Time Multi-Person 3D Mesh Estimation via Scale-Adaptive Tokens (CVPR 2025 · similarity 0.80)
- RAPTR: Radar-based 3D Pose Estimation using Transformer (NeurIPS 2025 · similarity 0.79)
- $\text{Di}^2\text{Pose}$: Discrete Diffusion Model for Occluded 3D Human Pose Estimation (NeurIPS 2024 · similarity 0.79)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.