GraphPose-Fi: WiFi 3D Pose Estimation

• The paper introduces GraphPose-Fi, a WiFi-based framework that encodes the human skeleton as a graph to improve 3D pose estimation.
• It leverages a shared CNN encoder, a Lightweight Temporal and Spatial Attention module, and graph convolution with multi-head self-attention for robust feature aggregation.
• The method achieves state-of-the-art performance on MM-Fi by reducing MPJPE and enhancing PCK, especially for central body joints in cross-subject and cross-environment settings.

GraphPose-Fi is a WiFi-based 3D human pose estimation framework that estimates 3D joint coordinates from WiFi channel state information (CSI) while explicitly encoding the human skeleton as a graph instead of directly regressing joint coordinates from RF signals. The framework combines a shared CNN encoder for per-antenna subcarrier–time feature extraction, a Lightweight Temporal and Spatial Attention (LTSA) module for joint-wise aggregation across time and antennas, and a graph-based regression head that integrates Chebyshev graph convolutions with multi-head self-attention to capture both skeletal topology and global inter-joint dependencies. On MM-Fi, it reports state-of-the-art Mean Per Joint Position Error (MPJPE) and Percentage of Correct Keypoints (PCK) under Protocol 1, particularly in cross-subject and cross-environment settings (Chen et al., 24 Nov 2025).

1. Problem formulation and motivation

GraphPose-Fi addresses WiFi-based 3D human pose estimation, where the input is a short temporal window of CSI measurements and the target is a 3D body pose. The CSI sequence is represented as

$$\mathbf{H} \in \mathbb{C}^{N_r \times N_t \times S \times T},$$

where $N_r$ is the number of receive antennas, $N_t$ the number of transmit antennas, $S$ the number of subcarriers, and $T$ the number of time steps. The 3D pose label is

$\mathbf{Y} \in \mathbb{R}^{J \times 3},$

with $J$ the number of joints; in MM-Fi, $J = 17$.

The central premise is that human motion perturbs WiFi propagation through reflection, scattering, diffraction, and multipath variation. Because modern WiFi uses OFDM, the CSI on each subcarrier captures a complex frequency response between transmit and receive antennas. These responses change systematically with body pose and motion, so CSI sequences contain pose-relevant information even though the mapping from CSI to joints is indirect and mediated by RF propagation rather than visual appearance.

The motivation for GraphPose-Fi is twofold. First, WiFi-based HPE inherits the practical advantages typically associated with RF sensing: occlusion robustness, because RF signals penetrate many obstacles; privacy, because CSI does not directly encode appearance in the manner of images; and low deployment cost, because commodity WiFi hardware is widespread. Second, most earlier WiFi-based HPE systems use CNNs, LSTMs, transformers, or hybrid regressors that treat pose prediction as a black-box mapping from CSI to coordinates. GraphPose-Fi argues that this neglects the topological coupling imposed by the kinematic tree. Joints are not statistically independent outputs: knees depend on hips, elbows depend on shoulders, and central body segments constrain distal limbs. Explicit graph modeling therefore introduces a skeletal inductive bias that prior direct-regression pipelines leave implicit (Chen et al., 24 Nov 2025).

2. Input representation and overall architecture

GraphPose-Fi applies standard CSI preprocessing to obtain a real-valued tensor

$$\mathbf{Z} \in \mathbb{R}^{A \times S \times T}, \qquad A = N_r N_t.$$

The preprocessing consists of magnitude extraction $|H|$ and phase calibration to remove random phase offsets caused by hardware or packet timing imperfections. A sliding window of $N_r$0 CSI frames is aligned with one pose label.

For MM-Fi, the WiFi setup is $N_r$1 and $N_r$2, so $N_r$3, $N_r$4, and $N_r$5. Each sample is therefore

$N_r$6

The architecture has three stages. The first stage is a per-antenna CNN encoder. For each antenna $N_r$7,

$N_r$8

where $N_r$9 is the number of feature channels, $N_t$0 is the number of latent spatial positions, and $N_t$1 is a compressed temporal length. A critical design choice is that the latent spatial dimension is explicitly set equal to the number of joints. This creates a joint-aligned representation before graph reasoning is introduced.

The encoder is shared across antennas. That weight sharing enforces the same processing pipeline for each antenna view and encourages the network to learn antenna-invariant local patterns. Functionally, the encoder is a ResNet-style 2D CNN inspired by MetaFi++, operating on the $N_t$2 CSI slice as a 2D signal image. After all antennas are processed, the resulting feature maps are stacked and channel-reduced to form

$N_t$3

The second stage is LTSA, which aggregates features over time and antennas on a per-joint basis. The third stage is the graph-based regression head, which receives the final joint embeddings and predicts

$N_t$4

In the best experimental configuration, the graph head uses $N_t$5 GCN-based attention blocks (Chen et al., 24 Nov 2025).

3. Lightweight Temporal and Spatial Attention

LTSA is designed to reweight evidence over time and across antennas while remaining substantially lighter than full self-attention over all antenna–time tokens. Its input is

$N_t$6

The temporal attention stage first averages over feature channels: $N_t$7 A point-wise $N_t$8 3D convolution is then applied on the $N_t$9 volume, followed by a softmax across the temporal dimension: $S$0 These weights define a weighted temporal aggregation,

$S$1

yielding

$S$2

The spatial attention stage then aggregates across antennas. After channel averaging and a point-wise $S$3 convolution on the $S$4 array, a softmax is applied across antennas: $S$5 The antenna-weighted aggregation is

$S$6

with

$S$7

After transposition and LayerNorm, the graph head receives

$S$8

The design intent is explicit joint-wise evidence selection. For each joint, LTSA can emphasize the most informative temporal segments and the most informative antenna observations rather than pooling them uniformly. In the reported ablation on MM-Fi P1–S1, Global Average Pooling gives 161.6 mm MPJPE, Per-Joint Multi-Head Self-Attention gives 160.8 mm, and LTSA gives 160.6 mm. The reported interpretation is that LTSA slightly outperforms PJ-MHSA while being much lighter computationally (Chen et al., 24 Nov 2025).

4. Graph-based pose regression

The graph head models the human body as a graph whose nodes are joints and whose edges are skeletal connections. Let $S$9 denote the adjacency matrix, $T$0 the diagonal degree matrix, and

$T$1

the normalized Laplacian. GraphPose-Fi uses a Chebyshev polynomial-based GCN, ChebGConv, with rescaled Laplacian

$T$2

The Chebyshev polynomials are defined recursively: $T$3

$T$4

Given node features $T$5, the graph convolution update is

$T$6

where $T$7 is the polynomial order and $T$8 are learnable weights. This yields localized graph filters whose receptive field spans up to $T$9-hop neighborhoods.

The regression head has three conceptual stages. First, an initial ChebGConv maps the LTSA output

$\mathbf{Y} \in \mathbb{R}^{J \times 3},$0

to

$\mathbf{Y} \in \mathbb{R}^{J \times 3},$1

Second, a stack of $\mathbf{Y} \in \mathbb{R}^{J \times 3},$2 GCN-based attention blocks processes these embeddings. Each block contains two ChebGConv layers and one multi-head self-attention layer across joints. The graph convolutions inject local topological structure, while MHSA captures global dependencies that need not follow the skeletal adjacency alone. Third, a final ChebGConv regresses the 3D coordinates,

$\mathbf{Y} \in \mathbb{R}^{J \times 3},$3

The ablation study quantifies the effect of explicit graph reasoning. Replacing the graph head with an MLP yields 167.8 mm MPJPE on MM-Fi P1–S1. A graph head with $\mathbf{Y} \in \mathbb{R}^{J \times 3},$4 blocks reduces this to 161.8 mm, $\mathbf{Y} \in \mathbb{R}^{J \times 3},$5 achieves the best result at 160.6 mm, and $\mathbf{Y} \in \mathbb{R}^{J \times 3},$6 slightly degrades to 161.5 mm. The paper attributes the improvement over the MLP head to explicit skeletal modeling and suggests that the small degradation at six blocks reflects overfitting or diminishing returns (Chen et al., 24 Nov 2025).

5. Training objective, dataset, and protocol

GraphPose-Fi is trained with mean squared error over joints: $\mathbf{Y} \in \mathbb{R}^{J \times 3},$7 No additional regularizers or auxiliary losses are described.

The reported training configuration uses AdamW with initial learning rate $\mathbf{Y} \in \mathbb{R}^{J \times 3},$8, weight decay 0.02, cosine decay to near zero, 50 epochs, and batch size 256. The implementation is in PyTorch and training is reported on an NVIDIA GH200 GPU.

Experiments are conducted on MM-Fi. The dataset contains 40 subjects, 14 daily activities, and 4 different rooms. The WiFi system operates at 5 GHz with 40 MHz bandwidth, using one transmit antenna and three receive antennas. Each sample therefore contains CSI of size $\mathbf{Y} \in \mathbb{R}^{J \times 3},$9, corresponding to receive antenna, subcarrier, and temporal frame axes. Ground-truth pose annotations consist of 17 joints in 3D coordinates.

Evaluation follows MM-Fi Protocol 1 with three splits. S1 is a random split with train:test $J$0. S2 is a cross-subject split with 32 training subjects and 8 unseen test subjects. S3 is a cross-environment split with training on three environments and testing on the remaining unseen environment. These progressively test base performance, subject generalization, and environment generalization.

The reported metrics are MPJPE, Procrustes Aligned MPJPE (PA-MPJPE), and PCK at thresholds 10, 20, 30, 40, and 50. MPJPE is the average Euclidean distance between predicted and ground-truth joints; PA-MPJPE applies a rigid similarity alignment before computing that distance; PCK measures the proportion of joints below a thresholded normalized error (Chen et al., 24 Nov 2025).

6. Empirical performance and error profile

On MM-Fi Protocol 1, GraphPose-Fi is compared against MetaFi++, HPE-Li, and DT-Pose. In S1, GraphPose-Fi reports 160.6 mm MPJPE, improving over MetaFi++ at 174.5, HPE-Li at 172.6, and DT-Pose at 168.0. Its PA-MPJPE is 105.0 mm, compared with 112.9 for MetaFi++, 102.0 for HPE-Li, and 102.4 for DT-Pose. For PCK@50, GraphPose-Fi reaches 89.3, compared with 88.1, 87.8, and 88.9 respectively. The reported pattern is therefore lower MPJPE and higher PCK across thresholds, with PA-MPJPE remaining slightly higher than HPE-Li and DT-Pose.

In S2, the cross-subject setting, GraphPose-Fi reports 210.5 MPJPE, compared with 214.8 for MetaFi++, 221.4 for HPE-Li, and 221.1 for DT-Pose. Its PCK@50 is 86.3, compared with 86.9, 84.3, and 84.8. The paper characterizes this as the best MPJPE with strong PCK under unseen-subject generalization.

In S3, the cross-environment setting, GraphPose-Fi reports 302.7 MPJPE and 67.2 PCK@50, compared with 341.8 and 57.3 for MetaFi++, 361.1 and 52.3 for HPE-Li, and 326.9 and 61.0 for DT-Pose. This is the most difficult protocol, and the reported improvement over the best baseline is from 326.9 to 302.7 mm in MPJPE, together with an increase of approximately 6.2 PCK points over DT-Pose.

Per-joint analysis on S1 shows that the largest gains occur for central body joints. Bot Torso improves from 116.4 to 93.9, Center Torso from 117.4 to 100.4, Upper Torso from 139.7 to 123.7, Neck Base from 166.1 to 150.9, and Center Head from 168.9 to 150.8. Limb joints also improve, but hands and elbows remain comparatively difficult: right hand from 372.7 to 360.4, left hand from 377.6 to 365.2, right elbow from 257.8 to 251.7, and left elbow from 252.6 to 244.6. The paper interprets this as evidence that graph modeling particularly benefits central joints, where structural consistency and global context are critical, while distal joints remain challenging because WiFi has limited spatial resolution (Chen et al., 24 Nov 2025).

7. Limitations, significance, and position in the literature

Several limitations are either stated directly or implied by the results. The cross-environment setting remains difficult, with MPJPE still around 300 mm, indicating sensitivity to domain shift in RF conditions. Fine joints such as hands and elbows remain hard to estimate, reflecting ambiguity in small limb motions and the limited spatial resolution of WiFi sensing. The experiments use a single WiFi system configuration with one transmitter and one receiver carrying three antennas, so behavior under different hardware layouts or reduced antenna counts is not fully characterized. The system is also single-modality: it uses WiFi only and does not fuse RGB or radar.

Within RF sensing, GraphPose-Fi is positioned against earlier WiFi or RF pose systems such as WiSPPN, WiPose, GoPose, MetaFi++, Person-in-WiFi 3D, HPE-Li, and DT-Pose. Earlier systems mainly employ CNN-based, recurrent, or transformer-based direct regression. GraphPose-Fi introduces explicit graph modeling of the skeleton into the WiFi domain by aligning latent features with joint indices and applying graph convolution over that axis. The paper relates this design to graph-based pose modeling in computer vision, including GraFormer, DiffPose, and 2D-to-3D lifting with GCNs, while emphasizing that the RF setting lacks intermediate 2D keypoint supervision and therefore requires end-to-end learning from CSI to joint features to 3D pose (Chen et al., 24 Nov 2025).

A broader research context is that graph-structured pose reasoning is appearing in distinct modalities and tasks. GRPose, for example, applies graph relations to pose-guided human image generation by constructing pose, latent, and fused graphs within a diffusion-based control pipeline, rather than using a fixed skeletal graph for RF pose regression (Yin et al., 2024). This suggests a wider methodological trend: pose priors are increasingly treated as structured relational objects rather than purely Euclidean feature maps.

The significance of GraphPose-Fi lies in showing that a skeletal graph prior can be integrated directly into WiFi-based 3D HPE without auxiliary pose priors or explicit kinematic constraints beyond the graph itself. A plausible implication is that future RF pose estimators may benefit from stronger structural inductive biases, multi-sensor fusion with RGB or radar, and domain-generalization strategies specifically targeted at cross-environment RF shift.