Pith. sign in

REVIEW 3 major objections 5 minor 56 references

Seeing Through WiFi: Lightweight Human Pose Estimation with Dynamic Kernel Attention

T0 review · 3 major / 5 minor · reviewed 2026-07-12 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read A lightweight WiFi-only network estimates human poses by dynamically weighting channel and frequency kernels, matching or beating heavier camera-supervised models.

desk verdict Solid engineering paper: joint channel-frequency dynamic kernels give a real lightweight SOTA on two public WiFi-HPE sets, with the usual teacher-label and lab-CSI caveats. read the letter →

arxiv 2607.03196 v1 pith:NGBXZAXU submitted 2026-07-03 cs.CV cs.LG

Toan D. Gian , Van-Dinh Nguyen , Vo Phi Son , Nhan Thanh Nguyen , Dinh Thai Hoang , Diep N. Nguyen , Nguyen Cong Luong , Symeon Chatzinotas This is my paper
classification cs.CVcs.LG
keywords WiFiCSIhumanposeestimationdynamicconvolutionchannel-frequencyattentionlightweightneuralnetworkteacher-studentlearningPCK
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that ordinary WiFi channel-state information already contains enough multipath detail to recover 2-D human skeletons, provided the network is allowed to re-weight its convolutional kernels separately along the antenna (channel) and subcarrier (frequency) axes. The authors train a student network solely on raw CSI under the supervision of a fixed camera model, then discard the camera at test time. The resulting student, called WiLHPE, uses a handful of dynamic kernels whose attention maps are produced by simple squeeze-and-excitation blocks; Tree-Structured Parzen Estimator search finds the few critical hyper-parameters. On two public datasets the method reaches 86 % and 94 % of joints inside the standard PCK50 threshold while using only 1.8–3.5 million parameters—roughly an order of magnitude fewer than the previous best WiFi pose estimator. Even when strong Gaussian noise is added, accuracy stays near 80 %. The claim is that privacy-preserving, camera-free pose tracking becomes practical on resource-limited devices once the network learns to look at CSI in both of its native dimensions at once.

What carries the argument

CF-DyConv (channel-frequency dynamic convolution): for each short CSI frame the network computes separate attention scalars along the antenna axis and the subcarrier axis, multiplies them onto a small bank of basis kernels, and uses the resulting adaptive kernel for ordinary 2-D convolution.

What would settle it

Train and test the identical CF-DyConv architecture on a new indoor environment whose furniture, walls, and subject set differ from both MM-Fi and WiPose; if PCK50 falls below the previous best WiFi baseline while parameter count stays the same, the central claim fails.

Watch

Extended reading notes

Core claim

WiLHPE shows that a student network fed only raw multi-antenna CSI can recover human keypoints at state-of-the-art accuracy when its convolutional kernels are dynamically re-weighted by joint channel-frequency attention; the same lightweight architecture remains accurate under substantial additive noise and requires far fewer parameters than earlier WiFi pose models.

Load-bearing premise

The camera-derived keypoints used as training labels are assumed accurate enough that a CSI-only student can safely treat them as ground truth, and that the multipath statistics of the two laboratory setups transfer to real rooms.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes WiLHPE, a teacher-student framework for WiFi CSI-based 2D human pose estimation. A fixed vision teacher (HRNet-w48) supplies keypoint labels from RGB frames; the student processes raw CSI via a stack of CF-DyNet blocks whose core operator, CF-DyConv, generates dynamic kernels by multiplicative channel- and frequency-domain attention (Eqs. 3–7). Hyper-parameters (r, τ, n, architecture widths) are selected by TPE. On the public MM-Fi (P3-S1) and WiPose datasets the method reports 85.96 % / 94.27 % PCK50 with 1.78–3.49 M parameters, outperforming MetaFi++ and several earlier WiFi HPE baselines, and retains ~80 % PCK50 under AWGN (σ²=0.5) and moderate FGSM attacks.

Significance. If the relative gains hold under the same teacher-label protocol used by prior work, the paper supplies a practically useful lightweight architecture for privacy-preserving, camera-free pose estimation on commodity WiFi. Strengths that should be credited include: (i) explicit multi-protocol / multi-setting tables on two public datasets, (ii) controlled robustness curves (AWGN + FGSM) that isolate the benefit of CF-DyConv over plain CNN, (iii) ablations of kernel count n, reduction ratio r and temperature τ, and (iv) a complexity argument showing the extra Mult-Adds remain modest. These elements make the engineering contribution reproducible and immediately usable by the wireless-sensing community.

major comments (3)
  1. Section IV-B and all reported metrics (Tables II, III, IV; Figs. 6–11) evaluate the student exclusively against 2-D keypoints produced by a fixed HRNet-w48 teacher. No teacher-error audit, cross-teacher ablation, or independent geometric ground truth (mocap / multi-view triangulation) is supplied. Consequently the absolute PCK/MPJPE numbers and the claimed superiority over MetaFi++ measure fidelity to the teacher rather than to true pose; any systematic teacher bias (occlusion, depth ambiguity, upper-body motion noted in §V-B1) is inherited by every figure. This is load-bearing for the SOTA claim and should be quantified or at least bounded.
  2. Abstract and Table IIa list average PCK50 = 85.96 % on MM-Fi, yet Table IV reports 85.26 % for the identical WiLHPE entry under P3-S1. The 0.7-point discrepancy is never explained; it undermines numerical trustworthiness of the central comparison.
  3. No multi-seed statistics, standard deviations or confidence intervals accompany any table or figure. Given that TPE is used for hyper-parameter search and that CSI multipath is environment-dependent, single-run point estimates leave open the possibility that the reported margins over MetaFi++ (≈2–4 % PCK) are within run-to-run variance.
minor comments (5)
  1. Section heading “IV. PROPOSEDWILPHE FRAMEWORK” contains a typographical error (missing space and “WiLPHE” vs. “WiLHPE”).
  2. Eq. (1) uses φ for both phase and the argument of the Dirac delta; the time-delay variable should be distinguished.
  3. Fig. 3 caption and surrounding text refer to “CF-DyNet blocks” while the figure itself shows only the attention paths of CF-DyConv; a clearer separation of the two would help readers.
  4. Several places write “A WGN” or “A WGN noise”; consistent “AWGN” is preferable.
  5. Table I lists “314” subcarriers for MM-Fi while the text and MetaFi++ literature usually cite 114; a brief clarification of the CSI extraction pipeline would remove ambiguity.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical NN architecture and held-out metrics against external teacher labels; no derivation reduces to inputs by construction.

full rationale

WiLHPE is an engineering paper that defines a student network (CF-DyNet with CF-DyConv attention over channel/frequency) trained by MSE (Eq. 10) to 2-D keypoints produced by a fixed external vision teacher (HRNet-w48). Test PCK/MPJPE numbers (Tables II–IV, Figs. 6–11) are computed on held-out CSI frames from public datasets (MM-Fi, WiPose) under that same supervision. Hyperparameters (r, τ, n, batch size, kernel size) are selected by TPE on validation performance (Algorithm 1, Section IV-B), which is ordinary model selection and does not force the reported test scores by construction. Complexity claims follow directly from counting Mult-Adds of the attention and aggregation stages (Section IV-A). Self-citations ([1], [48]–[52]) refer to related prior systems by overlapping authors but are not load-bearing for the numerical superiority claims; those rest on the new architecture and the tabulated comparisons. No equation equates a claimed accuracy or uniqueness result to a fitted constant, no uniqueness theorem is imported, and no known empirical pattern is merely renamed. The derivation chain is therefore self-contained and non-circular.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central performance claim rests on standard CSI multipath physics, the assumption that a vision teacher supplies usable labels, and a handful of architecture hyperparameters chosen by TPE. No new physical entities are postulated; the invented pieces are neural-network modules whose only evidence is the reported tables.

free parameters (4)
  • reduction ratio r = 16
    Bottleneck width of the excitation FC layers; selected by ablation/TPE (Table VII) and fixed at 16 for final results.
  • temperature τ = 30
    Softmax temperature controlling attention sparsity; ablated and set to 30 (Table VII).
  • number of kernels n = 3
    Count of basis kernels mixed by attention; ablated 1–4 and set to 3 (Table VI).
  • CF-DyNet depth M and channel widths n1/n2 = M=3, n1=64, n2=128
    Structural choices (M=3, n1=64, n2=128) fixed after TPE search; directly affect capacity and reported FLOPs.
assumptions (4)
  • domain assumption CSI multipath amplitude/phase across antennas and subcarriers is sufficiently altered by human joint positions to allow fine-grained pose recovery.
    Stated in Sections I and III-A; underpins the entire sensing premise.
  • domain assumption 2D keypoints extracted by a pre-trained vision model (HRNet-w48) are accurate enough to serve as ground-truth supervision for the CSI student.
    Section IV-B teacher-student setup; all reported metrics are measured against these labels.
  • ad hoc to paper Mean-squared error on keypoint coordinates is an adequate training objective for high-resolution CSI.
    Section IV-C explicitly prefers MSE over pose-adjacency losses used in prior work.
  • standard math Standard convolution, ReLU, softmax and global-average-pooling operators behave as usual.
    Used throughout Eqs. (2)–(7) without modification.
invented entities (2)
  • CF-DyConv (channel-frequency dynamic convolution)
    purpose: Generate input-dependent kernels by multiplying channel and frequency attention scalars onto a bank of basis kernels (Eq. 7).
    Core architectural novelty; evidence is only the ablation and end-to-end tables inside this paper.
  • WiLHPE / CF-DyNet student network
    purpose: End-to-end lightweight pipeline that maps raw CSI tensors to 2D keypoints under teacher supervision.
    System-level wrapper around CF-DyConv; no external validation beyond the two datasets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Seeing Through WiFi: Lightweight Human Pose Estimation with Dynamic Kernel Attention." pith.science (2026). https://pith.science/paper/NGBXZAXU

@misc{pith2026260703196,
  author       = {Pith},
  title        = {Pith review of: Seeing Through WiFi: Lightweight Human Pose Estimation with Dynamic Kernel Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NGBXZAXU}},
  note         = {Machine review of arXiv:2607.03196}
}
read the original abstract

WiFi-based human pose estimation (HPE) enables the detection and interpretation of human body positions and movements without the need for wearable devices while preserving individual privacy concerns. Implementing this solution requires enhancing model performance and maintaining efficiency, especially on resource-constrained devices. This paper introduces a novel framework, WiLHPE, for lightweight and efficient human pose estimation using WiFi CSI signals. Empowered by a camera-based model during training, WiLHPE processes raw WiFi signals directly to estimate human poses in the testing phase. It employs a novel neural network architecture to dynamically learn convolutional kernels and apply attention mechanisms across channel and frequency spaces. This innovative method diversifies the kernels to improve the recognition capabilities of WiFi signals without adding complexity, ensuring efficiency. Additionally, the Tree-Structured Parzen Estimator algorithm is employed to optimize the critical hyperparameters of the neural network efficiently, minimizing the time required for optimal hyperparameter search compared to heuristic methods. Results from experiments on both the MM-Fi and WiPose datasets highlight the superiority of WiLHPE over state-of-the-art approaches, achieving 85.96% and 94.27% at PCK50, respectively, with minimal computational overhead. Notably, WiLHPE performs impressively even under challenging conditions, maintaining around 80% at PCK50 under AWGN noise with an error variance of 0.5.

Figures

Figures reproduced from arXiv: 2607.03196 by the authors.

Figure 1
Figure 1. Illustration of indoor WiFi-based human sensing: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of Dynamic Convolution. B. Overview of Dynamic Convolution Basic Concept. Dynamic convolution (DyConv) was intro￾duced in [44] to enhance the representational capacity of conventional convolution by dynamically adjusting the con￾volution kernel in response to input characteristics. Firstly, DyConv extracts attention weights based on the input and then performs a weighted summation of n basis kernels using t… view at source ↗
Figure 3
Figure 3. The proposed WiLHPE network: Empowered by the CV model, the student network predicts human pose from CSI data. [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Block diagram of the proposed channel-frequency dynamic convolution. [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: The architecture of CF-DyNet and decoder blocks [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: The PCKα performance with different protocols and settings on MM-Fi. TABLE III: The MPJPE and PA-MPJPE Results of WiLHPE with Different Protocols and Settings on MM-Fi (Best in bold and second best in underlined) P Setting 1 Packet Training Testing MPJPE↓ PA-MPJPE↓ 1 1…
Figure 7
Figure 7. Figure 7: Visualization of the human pose landmarks generated by the vision model (red) and WiFi model (blue) on the MM-Fi [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 8
Figure 8. Figure 8: The PCK scores of WiLHPE versus different levels of [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 11
Figure 11. Figure 11: The PCK scores of WiLHPE under FGSM-U-N versus [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 10
Figure 10. Figure 10: The PCK scores of WiLHPE under FGSM-U-N versus [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 2 linked inside Pith

  1. [1]

    WiLHPE: WiFi-enabled lightweight channel frequency dynamic convolution for HPE tasks,

    T. D. Gian, T.-H. Nguyen, N. T. Nguyen, and V .-D. Nguyen, “WiLHPE: WiFi-enabled lightweight channel frequency dynamic convolution for HPE tasks,” inInt. Conf. Commun. Elect. (ICCE), 2024, pp. 516–521

    2024

  2. [2]

    Human activity recognition: A survey,

    C. Jobanputra, J. Bavishi, and N. Doshi, “Human activity recognition: A survey,”Procedia Computer Science, vol. 155, pp. 698–703, 2019

    2019

  3. [3]

    Human activity recognition: A review,

    O. C. Ann and L. B. Theng, “Human activity recognition: A review,” inIEEE Int. Confe. Cont. Sys., Compu. and Engi. (ICCSCE), 2014, pp. 389–393

    2014

  4. [4]

    A survey on human activity recognition using wearable sensors,

    O. D. Lara and M. A. Labrador, “A survey on human activity recognition using wearable sensors,”IEEE Commun. Sur. Tut., vol. 15, no. 3, pp. 1192–1209, 2013

    2013

  5. [5]

    Human activity recognition using wearable sensors by deep convolutional neural networks,

    W. Jiang and Z. Yin, “Human activity recognition using wearable sensors by deep convolutional neural networks,” pp. 1307–1310, 2015

    2015

  6. [6]

    A review on video-based human activity recognition,

    S.-R. Ke, H. L. U. Thuc, Y .-J. Lee, J.-N. Hwang, J.-H. Yoo, and K.-H. Choi, “A review on video-based human activity recognition,”Computers, vol. 2, no. 2, pp. 88–131, 2013

    2013

  7. [7]

    A survey on video-based human action recognition: recent updates, datasets, challenges, and applications,

    P. Pareek and A. Thakkar, “A survey on video-based human action recognition: recent updates, datasets, challenges, and applications,”Artif. Intell. Rev., vol. 54, no. 3, pp. 2259–2322, mar 2021

    2021

  8. [8]

    Through-wall human mesh recovery using radio signals,

    M. Zhao, Y . Liu, A. Raghu, H. Zhao, T. Li, A. Torralba, and D. Katabi, “Through-wall human mesh recovery using radio signals,”Int. Conf. Comput. Vis., pp. 10 112–10 121, 2019

    2019

Show all 56 references
  1. [9]

    RF-based 3D skeletons,

    M. Zhaoet al., “RF-based 3D skeletons,”Proc. 2018 Conf. ACM Special Interest Group on Data Commun., 2018

    2018

  2. [10]

    Pose estimation at night in infrared images using a lightweight multi-stage attention network,

    Y . Zang, C.-N. Fan, Z. Zheng, and D. Yang, “Pose estimation at night in infrared images using a lightweight multi-stage attention network,” Signal, Image and Video Processing, vol. 15, pp. 1757 – 1765, 2021. 13

    2021

  3. [11]

    Wireless sensing for human activity: A survey,

    J. Liu, H. Liu, Y . Chen, Y . Wang, and C. Wang, “Wireless sensing for human activity: A survey,”IEEE Commun. Sur. & Tut., vol. 22, no. 3, pp. 1629–1645, 2020

    2020

  4. [12]

    A survey on behavior recognition using WiFi channel state information,

    S. Yousefi, H. Narui, S. Dayal, S. Ermon, and S. Valaee, “A survey on behavior recognition using WiFi channel state information,”IEEE Commun. Mag., vol. 55, no. 10, pp. 98–104, 2017

    2017

  5. [13]

    Tool release: gathering 802.11n traces with channel state information,

    D. Halperin, W. Hu, A. Sheth, and D. Wetherall, “Tool release: gathering 802.11n traces with channel state information,” vol. 41, no. 1, p. 53, Jan. 2011

    2011

  6. [14]

    Human activity recognition across scenes and categories based on csi,

    Y . Zhang, X. Wang, Y . Wang, and H. Chen, “Human activity recognition across scenes and categories based on csi,”IEEE Trans. Mobile Comput,, vol. 21, no. 7, pp. 2411–2420, 2022

    2022

  7. [15]

    Towards 3D human pose construction using WiFi,

    W. Jianget al., “Towards 3D human pose construction using WiFi,” Proc. 26th Annual Inter. Conf. Mobile Comput. and Net., 2020

    2020

  8. [16]

    From signal to image: Capturing fine-grained human poses with commodity WiFi,

    L. Guo, Z. Lu, X. Wen, S. Zhou, and Z. Han, “From signal to image: Capturing fine-grained human poses with commodity WiFi,”IEEE Commun. Lett., vol. 24, no. 4, pp. 802–806, 2020

    2020

  9. [17]

    Can WiFi estimate person pose?

    F. Wang, S. Panev, Z. Dai, J. Han, and D. Huang, “Can WiFi estimate person pose?”Clinical Orthopaedics and Related Research(CORR), vol. abs/1904.00277, 2019

    1904 arXiv

  10. [18]

    MetaFi: Device-free pose estimation via commodity WiFi for metaverse avatar simulation,

    J. Yang, Y . Zhou, H. Huang, H. Zou, and L. Xie, “MetaFi: Device-free pose estimation via commodity WiFi for metaverse avatar simulation,” inProc. IEEE 8th World Int. of Things (WF-IoT), 2022, pp. 1–6

    2022

  11. [19]

    MetaFi++: WiFi-enabled transformer-based human pose estimation for metaverse avatar simulation,

    Y . Zhou, H. Huang, S. Yuan, H. Zou, L. Xie, and J. Yang, “MetaFi++: WiFi-enabled transformer-based human pose estimation for metaverse avatar simulation,”IEEE Internet of Things J., vol. 10, no. 16, pp. 14 128–14 136, 2023

    2023

  12. [20]

    MM-Fi: Multi-modal non-intrusive 4D human dataset for versatile wireless sensing,

    J. Yang, H. Huang, Y . Zhou, X. Chen, Y . Xu, S. Yuan, H. Zou, C. X. Lu, and L. Xie, “MM-Fi: Multi-modal non-intrusive 4D human dataset for versatile wireless sensing,” inProc. Thirty-seventh Conf. Neural Infor. Process. Sys. Data. and Bench. Track, 2023. [Online]. Available: ...

    2023

  13. [21]

    Algorithms for hyper- parameter optimization,

    J. Bergstra, R. Bardenet, Y . Bengio, and B. Kégl, “Algorithms for hyper- parameter optimization,” inProc. Neural Infor. Process. Sys., 2011. [Online]. Available: https://api.semanticscholar.org/CorpusID:11688126

    2011

  14. [22]

    Perunet: Deep signal channel attention in unet for WiFi-based human pose estimation,

    Y . Zhou, A. Zhu, C. Xu, F. Hu, and Y . Li, “Perunet: Deep signal channel attention in unet for WiFi-based human pose estimation,”IEEE Sensors J., vol. 22, no. 20, pp. 19 750–19 760, 2022

    2022

  15. [23]

    See through walls with WiFi!

    F. Adib and D. Katabi, “See through walls with WiFi!” vol. 43, no. 4. New York, NY , USA: Asso. Comp. Mach., aug 2013, pp. 75–86

    2013

  16. [24]

    SpotFi: Decimeter level localization using WiFi,

    M. Kotaru, K. Joshi, D. Bharadia, and S. Katti, “SpotFi: Decimeter level localization using WiFi,” inProc. SIGCOMM Comput. Commun. Rev., vol. 45, no. 4. New York, NY , USA: Asso. Comp. Mach., aug 2015, pp. 269–282

    2015

  17. [25]

    Extracting gait velocity and stride length from surrounding radio signals,

    C.-Y . Hsu, Y . Liu, Z. Kabelac, R. Hristov, D. Katabi, and C. Liu, “Extracting gait velocity and stride length from surrounding radio signals,” inProc. CHI Conf. Hum. Fact. Compu. Sys., ser. CHI ’17. New York, NY , USA: Asso. Comp. Mach., 2017, pp. 2116–2126

    2017

  18. [26]

    Gait recognition using WiFi signals,

    W. Wanget al., “Gait recognition using WiFi signals,” inProc. ACM Int. Joint Conf. Perva. Ubi. Compu., ser. UbiComp ’16. New York, NY , USA: Asso. Comp. Mach., 2016, pp. 363–373

    2016

  19. [27]

    Smart homes that monitor breathing and heart rate,

    F. Adib, H. Mao, Z. Kabelac, D. Katabi, and R. C. Miller, “Smart homes that monitor breathing and heart rate,” inProc. 33rd Annual ACM Conf. Human Fact. Comput. Sys.New York, NY , USA: Asso. Comp. Mach., 2015, pp. 837–846

    2015

  20. [28]

    From fresnel diffraction model to fine-grained human respiration sens- ing with commodity Wi-Fi devices,

    F. Zhang, D. Zhang, J. Xiong, H. Wang, K. Niu, B. Jin, and Y . Wang, “From fresnel diffraction model to fine-grained human respiration sens- ing with commodity Wi-Fi devices,”Proc. ACM Interact. Mob. Wearable Ubiquitous Technol., vol. 2, no. 1, mar 2018

    2018

  21. [29]

    WFID: Passive device-free human identification using WiFi signal,

    F. Hong, X. Wang, Y . Yang, Y . Zong, Y . Zhang, and Z. Guo, “WFID: Passive device-free human identification using WiFi signal,” ser. MO- BIQUITOUS 2016. New York, NY , USA: Asso. Comp. Mach., 2016, pp. 47–56

    2016

  22. [30]

    WiWho: WiFi-based person identification in smart spaces,

    Y . Zeng, P. H. Pathak, and P. Mohapatra, “WiWho: WiFi-based person identification in smart spaces,” inProc. 15th ACM/IEEE Inter. Conf. Infor. Process. Sensor Netw. (IPSN), 2016, pp. 1–12

    2016

  23. [31]

    Enabling contactless detection of moving humans with dynamic speeds using CSI,

    K. Qian, C. Wu, Z. Yang, Y . Liu, F. He, and T. Xing, “Enabling contactless detection of moving humans with dynamic speeds using CSI,” vol. 17, no. 2, Jan. 2018

    2018

  24. [32]

    PADS: Passive detection of moving targets with dynamic speed using PHY layer information,

    K. Qian, C. Wu, Z. Yang, Y . Liu, and Z. Zhou, “PADS: Passive detection of moving targets with dynamic speed using PHY layer information,” in Proc. IEEE Inter. Conf. Para. Dist. Sys. (ICPADS), 2014, pp. 1–8

    2014

  25. [33]

    WiFi CSI based passive human activity recognition using attention based BLSTM,

    Z. Chen, L. Zhang, C. Jiang, Z. Cao, and W. Cui, “WiFi CSI based passive human activity recognition using attention based BLSTM,”IEEE Trans. Mobile Comput., vol. 18, no. 11, pp. 2714–2724, 2019

    2019

  26. [34]

    Environment-robust device-free human activity recognition with channel-state-information enhancement and one-shot learning,

    Z. Shi, J. A. Zhang, R. Y . Xu, and Q. Cheng, “Environment-robust device-free human activity recognition with channel-state-information enhancement and one-shot learning,”IEEE Trans. Mobile Comput., vol. 21, no. 2, pp. 540–554, 2022

    2022

  27. [35]

    Openpose: Realtime multi-person 2D pose estimation using part affinity fields,

    Z. Cao, G. Hidalgo, T. Simon, S.-E. Wei, and Y . Sheikh, “Openpose: Realtime multi-person 2D pose estimation using part affinity fields,” IEEE Trans. Patt. Ana. and Mach. Int., vol. 43, pp. 172–186, 2018

    2018

  28. [36]

    Rmpe: Regional multi-person pose estimation,

    H.-S. Fang, S. Xie, Y .-W. Tai, and C. Lu, “Rmpe: Regional multi-person pose estimation,” inInt. Conf. Comput. Vis., 2017, pp. 2353–2362

    2017

  29. [37]

    HumanEva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion,

    L. Sigal, A. Balan, and M. J. Black, “HumanEva: Synchronized video and motion capture dataset and baseline algorithm for evaluation of articulated human motion,”Inter. J. Comp. Vis., vol. 87, no. 1, pp. 4–27, Mar. 2010

    2010

  30. [38]

    Microsoft kinect sensor and its effect,

    Z. Zhang, “Microsoft kinect sensor and its effect,”IEEE MultiMedia, vol. 19, no. 2, pp. 4–10, 2012

    2012

  31. [39]

    Human sensing using visible light communication,

    T. Li, C. An, T. Zhao, A. T. Campbell, and X. Zhou, “Human sensing using visible light communication,”Proc. Ann. Int. Conf. Mobi. Comp. Net., 2015. [Online]. Available: https://api.semanticscholar.org/ CorpusID:7473648

    2015

  32. [40]

    V oxNet: A 3D convolutional neural network for real-time object recognition,

    D. Maturana and S. Scherer, “V oxNet: A 3D convolutional neural network for real-time object recognition,” inProc. IEEE/RSJ Int. Confer. Intel. Rob. and Sys. (IROS), 2015, pp. 922–928

    2015

  33. [41]

    Capturing the human figure through a wall,

    F. Adib, C.-Y . Hsu, H. Mao, D. Katabi, and F. Durand, “Capturing the human figure through a wall,”ACM Trans. Graph., vol. 34, no. 6, 2015

    2015

  34. [42]

    Through-wall human pose estimation using radio signals,

    M. Zhao, T. Li, M. A. Alsheikh, Y . Tian, H. Zhao, A. Torralba, and D. Katabi, “Through-wall human pose estimation using radio signals,” IEEE Conf. Comput. Vis. Pattern Recog., pp. 7356–7365, 2018

    2018

  35. [43]

    Squeeze-and-excitation networks,

    J. Hu, L. Shen, and G. Sun, “Squeeze-and-excitation networks,” inIEEE Conf. Comput. Vis. Pattern Recog., 2018, pp. 7132–7141

    2018

  36. [44]

    Dynamic convolution: Attention over convolution kernels,

    Y . Chen, X. Dai, M. Liu, D. Chen, L. Yuan, and Z. Liu, “Dynamic convolution: Attention over convolution kernels,”IEEE Conf. Comput. Vis. Pattern Recog., pp. 11 027–11 036, 2019

    2019

  37. [45]

    Omni-dimensional dynamic convolution,

    C. Li, A. Zhou, and A. Yao, “Omni-dimensional dynamic convolution,” ArXiv, vol. abs/2209.07947, 2022

    2022 arXiv

  38. [46]

    Frequency dynamic convolution: Frequency-adaptive pattern recognition for sound event detection,

    H. Nam, S.-H. Kim, B. Ko, and Y .-H. Park, “Frequency dynamic convolution: Frequency-adaptive pattern recognition for sound event detection,” inInterspeech, 2022

    2022

  39. [47]

    Selective kernel networks,

    X. Li, W. Wang, X. Hu, and J. Yang, “Selective kernel networks,”CVPR, pp. 510–519, 2019

    2019

  40. [48]

    Robust WiFi sensing-based human pose estimation using denoising autoencoder and CNN with dynamic subcarrier attention,

    X. Hoang Nguyen, V .-D. Nguyen, Q.-T. Luu, T. Dinh Gian, and O.- S. Shin, “Robust WiFi sensing-based human pose estimation using denoising autoencoder and CNN with dynamic subcarrier attention,” IEEE Internet of Things J., vol. 12, no. 11, pp. 17 066–17 079, 2025

    2025

  41. [49]

    Multi-modal human pose estimation: A Wi-Fi-driven approach with adaptive kernel selection,

    T. D. Gian, D. T. Tran, Q.-V . Pham, L.-N. Tran, and V .-D. Nguyen, “Multi-modal human pose estimation: A Wi-Fi-driven approach with adaptive kernel selection,”IEEE Transactions on Artificial Intelligence, pp. 1–14, 2025

    2025

  42. [50]

    WiLHPE: WiFi-enabled lightweight channel frequency dynamic convolution for HPE tasks,

    T. D. Gian, T.-H. Nguyen, N. T. Nguyen, and V .-D. Nguyen, “WiLHPE: WiFi-enabled lightweight channel frequency dynamic convolution for HPE tasks,” in2024 Tenth Int. Conf. on Comm. and Elect. (ICCE), 2024, pp. 516–521

    2024

  43. [51]

    HPE-Li: WiFi-enabled lightweight dual selective kernel convolution for human pose estimation,

    T. D. Gian, T. Dac Lai, T. Van Luong, K.-S. Wong, and V .-D. Nguyen, “HPE-Li: WiFi-enabled lightweight dual selective kernel convolution for human pose estimation,” inEuro. Conf. Comput. Vision (ECCV). Springer, 2024, pp. 93–111

    2024

  44. [52]

    TinySense: Effective CSI compression for scalable and accurate Wi-Fi sensing,

    T. D. Gian, D. T. Tran, V . Q. Pham, F. Restuccia, and V .-D. Nguyen, “TinySense: Effective CSI compression for scalable and accurate Wi-Fi sensing,”IEEE Int. Conf. on Pervasive Comput. Comm., 2026

    2026

  45. [53]

    Deep 3D human pose estimation: A review,

    J. Wang, S. Tan, X. Zhen, S. Xu, F. Zheng, Z. He, and L. Shao, “Deep 3D human pose estimation: A review,”Computer Vision and Image Understanding, vol. 210, p. 103225, 2021

    2021

  46. [54]

    From point to space: 3D moving human pose estimation using commodity WiFi,

    Y . Wang, L. Guo, Z. Lu, X. Wen, S. Zhou, and W. Meng, “From point to space: 3D moving human pose estimation using commodity WiFi,” IEEE Commun. Lett., vol. 25, no. 7, pp. 2235–2239, 2021

    2021

  47. [55]

    WiLDAR: WiFi signal-based lightweight deep learning model for human activity recognition,

    F. Deng, E. Jovanov, H. Song, W. Shi, Y . Zhang, and W. Xu, “WiLDAR: WiFi signal-based lightweight deep learning model for human activity recognition,”IEEE Internet of Things J., vol. 11, no. 2, pp. 2899–2908, 2024

    2024

  48. [56]

    On the robustness of 3D human pose estimation,

    Z. Chen, Y . Huang, and L. Wang, “On the robustness of 3D human pose estimation,” inProc. 25th Inter. Conf. Patt. Recog. (ICPR), 2021, pp. 5326–5332

    2021

Pith tools

Reviewed July 12, 2026 · model on record in the stance chip above.