Pith. sign in

REVIEW 3 major objections 3 minor 31 references

Quiet-period Z-scores make WiFi CSI occupancy sensing transfer across rooms and chips, reaching F1 up to 0.99 zero-shot.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 11:12 UTC pith:RECGI6RS

load-bearing objection Clean central idea and the right cross-generation experiment, but the zero-shot claim overstates what the target empty-room bootstrap actually requires. the 3 major comments →

arxiv 2607.26665 v1 pith:RECGI6RS submitted 2026-07-29 cs.ET

OpenCSI: Self-Calibration Layer for Heterogeneous Mesh Wireless Sensor Networks

Karim Khamaisi , Simon Sigg , Bruno Rodrigues This is my paper
classification cs.ET
keywords WiFi CSI sensingoccupancy detectioncalibrationZ-score normalizationcross-environment transferzero-shot transferESP32indoor localization
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

OpenCSI claims that WiFi CSI occupancy sensing can be made portable by replacing per-session normalization with a per-link Z-score: each link's mean subcarrier amplitude is divided by its own quiet-period temporal standard deviation. Because chip gain and room geometry enter both numerator and denominator multiplicatively, the ratio cancels them, so a single empty-versus-occupied threshold trained in one room transfers zero-shot to other rooms, other ESP32 chip generations, and even across 802.11n-to-802.11ax PHY changes, with binary F1 up to 0.99 and no target-domain data. The method learns its baseline online from an empty-room bootstrap, reports a reliability/maturity tag so downstream logic can detect stale calibration and abstain, and requires only about five minutes of warmup per deployment. The paper deliberately scopes the claim to binary presence: distinguishing static from moving requires absolute magnitude, which the temporal-std denominator removes by design.

Core claim

The central claim is that Equation (1)'s per-link Z-score z = |A_bar - mu| / sigma, with mu and sigma learned online from quiet periods, is chip- and room-independent, so a model trained on one deployment holds a single empty-versus-occupied decision threshold zero-shot across nearly all transfer cells, reaching binary F1 up to 0.99 where standard normalization drops to 0.87 or fails outright, with no target-domain data or retraining. The paper tests this across three rooms, three ESP32 generations (S3, C3, C6), and a same-room 802.11n-to-802.11ax swap, and finds one direction (C3-to-S3') fails structurally due to the C3's higher noise floor compressing the Z-score scale. The argument is mec

What carries the argument

The carrying object is the Z-score in Eq (1): z = |A_bar - mu| / sigma, where A_bar is the mean amplitude across subcarriers for one directed link in one frame, and (mu, sigma) are that link's quiet-period mean and temporal standard deviation, updated online by an accumulator that never stores raw samples. The denominator is the load-bearing piece: it is learned during automatically detected quiet periods and absorbs the link's own gain and geometry multiplicatively, so the ratio becomes a dimensionless 'is something perturbing this link' signal. Around it sits a pipeline of quiet-gating (temporal variance plus spectral entropy), frequency-bucket and time-of-day baselines, a reliability scor

Load-bearing premise

The cancellation in Eq (1) assumes chip gain and room geometry enter the link's amplitude and its quiet-period standard deviation multiplicatively and identically; if additive noise or non-multiplicative AGC behavior dominates, the ratio is no longer deployment-invariant, and the paper observes exactly this compression on one chip generation.

What would settle it

Take a chip with a substantially different additive noise floor (or a gain stage that does not scale input linearly) and run the same empty/occupied protocol: if its quiet-period and occupied Z-score distributions do not align with the fixed threshold region (i.e., empty z approaches or overlaps occupied z), the invariance claim is falsified. The paper's own C3-to-S3' failure at F1=0.524 is a partial instance; a clean test would hold the room constant and swap only the silicon.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • A deployment needs one five-minute empty-room bootstrap; afterwards, models and thresholds transfer to new rooms, chip batches, and even PHY generations without any target-domain data.
  • Mesh sensing logic can consume one stable per-link Z-score per directed link instead of re-solving calibration per deployment, and can abstain when calibration maturity drops below a threshold.
  • Stale baselines are detectable on-device within seconds and recoverable with 30 seconds of fresh quiet observation, restoring F1 from 0.25 to 0.98.
  • The transfer is structurally limited to binary presence/sign-thresholding; tasks like static-vs-moving discrimination require absolute magnitude and should not be expected to transfer.
  • On hardware with higher additive noise floors, the Z-score compresses and transfer becomes asymmetric (one tested direction dropped to F1 0.524), so the invariance is not absolute.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A natural extension is a noise-floor correction: if the additive noise component of sigma could be estimated (e.g., from a no-target deep fade), the Z-score could be re-scaled to restore transfer on high-noise chips; this is my inference, not a paper claim.
  • The ratio-symmetry principle suggests the same quiet-period normalization could be applied to other per-link statistics, such as phase-difference variance, especially on multi-antenna platforms where antenna-ratio cancellation sharpens the phase signal.
  • Because the denominator intentionally destroys absolute magnitude, any task needing it (people counting, static-vs-moving, fine-grained activity) will need a second, separately calibrated magnitude channel; the paper's contribution is precisely to isolate the transferable sign channel.
  • The maturity/reliability machinery implies a self-healing mesh: a controller that watches the maturity tag could trigger re-bootstrap or re-anchoring automatically on drift, turning calibration into a monitored, recoverable resource.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 3 minor

Summary. The paper proposes OpenCSI, a preprocessing layer that converts per-link mean CSI amplitude into a dimensionless Z-score using per-link quiet-period mean and temporal standard deviation learned online via a Welford accumulator. It argues that multiplicative chip gain and room geometry cancel in the ratio, making a single empty-versus-occupied threshold transfer across rooms, chip generations, and PHY formats. The evaluation reports binary occupancy F1 on three ESP32 generations across three rooms, including a same-room 802.11n-to-802.11ax chip swap, and claims zero-shot transfer with no target-domain data or retraining. Source code and datasets are released.

Significance. If the claimed invariance holds, OpenCSI would be a practically valuable, lightweight calibration abstraction for low-cost ESP32 meshes. The open-source release and the reproducible evaluation pipeline are clear strengths, and the paper is transparent about the C3 failure direction and about the sign-vs-magnitude scope. However, the central 'zero-shot / no target data' framing is contradicted by the method's own required target empty-room bootstrap and by the baseline-swap ablation, and the headline comparison against warmup-normalized amplitude does not isolate the proposed mechanism. The core idea remains defensible, but the contribution as stated is overstated.

major comments (3)
  1. [Abstract, §III-C, §IV-E, Table I] The claim of zero-shot transfer with 'no target-domain data or retraining' is contradicted by the method itself. §III-C requires a 120 s empty-room bootstrap in every deployment to learn per-link μ and σ, and §IV-E shows that reusing a source baseline instead of the target baseline collapses A→B F1 from 0.98 to 0.25. Target quiet-period frames are therefore target-domain calibration data, not merely evaluation data. Table I's 'No tgt. data' checkmark is misleading. The contribution should be stated as 'no labeled target data and no retraining, but a target empty-room bootstrap is required.' This is load-bearing for the paper's headline.
  2. [§IV-A, Table II, §IV-E] The main comparison does not isolate the contribution of Eq. (1). Warmup-normalized amplitude is applied with source-session statistics zero-shot at the cross-environment boundary, while OpenCSI uses per-link target quiet-period statistics. The observed gain could therefore be due to target calibration rather than to the dimensionless Z-score. The §IV-E baseline-swap ablation (0.98→0.25) shows that target-baseline access is worth far more than most Table II gaps. Please add matched conditions: warmup-normalized with target warmup statistics, and OpenCSI with the source baseline as a Table II row or column, so the 'temporal-std denominator' mechanism is actually isolated.
  3. [§III-E, §IV-B] The ratio-cancellation argument in Eq. (1) relies on multiplicative gain and geometry affecting numerator and denominator identically. The paper itself notes that additive noise enters σ as σ_channel^2 + σ_n^2 and is not cancelled; the C3→B direction then fails (F1=0.524, below both warmup-normalized amplitude at 0.608 and the ablation at 0.690). This is not a minor outlier: it is the one source-hardware class with a substantially different noise floor. Moreover, Env C confounds chip (C3) with the smallest room, so the failure cannot be cleanly attributed to noise floor without an S3-in-Env-C control. Please either add that control or explicitly scope the transfer claim to deployments within the same noise-floor class and report the failing-cell rate.
minor comments (3)
  1. [§IV-A, Fig. 4] Eq. (4) uses τ=2.0 for the calibrated Z-score count feature, but Fig. 4 and §IV-B state the decision split is at z≈3. Reconcile these thresholds and clarify whether the central claim concerns the count-feature threshold, the RF classifier, or both.
  2. [§III-C, §V-A] The calibration-maturity plateau of ~0.65 is described as empirical to the three environments; it would be helpful to flag it explicitly as a deployment-specific hyperparameter rather than a principled constant.
  3. [Table II] The train→test direction is hard to read from the table layout. Consider labeling rows and columns with explicit source→target arrows or separate matrices per target environment.

Circularity Check

0 steps flagged

No circular derivation: the Z-score transfer is an empirical, held-out claim; the abstract's 'no target-domain data' wording overstates the required target empty-room bootstrap, and the only self-citation is peripheral.

full rationale

OpenCSI's central claim is that the per-link Z-score z=|A_bar-mu|/sigma (Eq. 1) cancels multiplicative chip/room effects, allowing a source-trained threshold to transfer. The per-link mu/sigma are learned from a target empty-room bootstrap, but the occupied deviations that the threshold must detect are not fitted; they are evaluated on held-out occupied and trailing-empty segments in target deployments. The cross-generation S3-C6 swap and cross-room evaluations are external, held-out tests, so the transfer claim has independent empirical content. The abstract's phrase 'with no target-domain data or retraining' is contradicted by the required 120 s bootstrap in each deployment (§III-C) and by the baseline-swap ablation showing that substituting the source baseline collapses A->B F1 from 0.98 to 0.25 (§IV-E). This is a claims-accuracy problem, not a circular reduction: the calibration statistics do not encode the occupied-class outcome. The limitation section honestly discloses the known-empty-room assumption (§V-D). The only self-citation, [18], concerns mesh airtime allocation and is not load-bearing for the transfer result. No equation reduces to its own input, and no prediction is statistically forced by fitted parameters.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

No new physical entities are postulated. OpenCSI introduces software-defined quantities (per-link Z-score, reliability score, calibration-maturity tag) that are defined and measured within the paper, not unexplained constructs. The free parameters and axioms above are the main external inputs the central claim depends on.

free parameters (5)
  • Feature threshold tau = 2.0 (calibrated), 1.5 (warmup-normalized)
    Eq. (4) uses fixed tau to count links exceeding the threshold; value is per feature type and affects the feature vector and F1.
  • Bootstrap/warmup duration = 120 s cold, 30 s recovery
    Section III-C sets the bootstrap window and §IV-E empirically determines recovery time; these durations control Welford-estimate stability.
  • w_phi SNR weighting parameters = center 15 dB, width 5
    Eq. (2) uses a sigmoid in SNR to weight phase; constants are hand-chosen, though phase is attenuated on single-antenna nodes.
  • Reliability saturation constants = not fully reported
    Eq. (3) defines reliability with several saturation constants (tau_age, N_min, f_min); values are not all specified and are used in drift analysis.
  • QuietGate variance/entropy thresholds = learned per deployment, values not reported
    Section III-C says the quiet gate rejects frames whose temporal variance or spectral entropy exceed learned thresholds; these are fitted to data and not specified.
axioms (5)
  • standard math Welford's online accumulator correctly maintains running mean and standard deviation.
    Used in §III-C for baseline learning; standard algorithm.
  • domain assumption The room is empty during the bootstrap window.
    §III-C: 'the room must be empty during this window or the occupant's perturbation is absorbed into the baseline.' Deployment requirement.
  • domain assumption Chip gain, AGC, and room geometry enter the per-link cross-subcarrier mean amplitude multiplicatively, so they cancel in Eq. (1).
    §III-E states the ratio cancels them; additive noise is acknowledged not to cancel and is observed to break transfer on C3.
  • domain assumption Quiet periods can be identified online via temporal variance and spectral entropy gates.
    §III-C assumes the gate reliably separates quiet frames from interference; otherwise baselines are corrupted.
  • domain assumption One 12-minute session per environment is representative of that deployment.
    §IV-A records one session per environment; no repeated sessions or day-to-day variance are reported.

pith-pipeline@v1.3.0-daily-deepseek · 12407 in / 18010 out tokens · 175379 ms · 2026-08-01T11:12:11.352042+00:00 · methodology

0 comments
read the original abstract

WiFi CSI sensing models trained in one environment usually fail in another because standard per-session normalization bakes chip- and room-specific artifacts into feature space, requiring fresh calibration for every new room or radio. We propose OpenCSI, an abstraction layer that hides these artifacts by exposing each mesh link as a single dimensionless Z-score against its own quiet-period temporal standard deviation. The denominator is learned online from a short empty-room bootstrap and reported with a maturity tag, enabling downstream logic to detect drift and abstain when baselines become unreliable. We evaluate OpenCSI on binary occupancy across three distinct rooms and three ESP32 generations (S3, C3, C6, spanning 802.11n HT20 and 802.11ax HE20), including a same-room chip swap isolating hardware from geometry. A model trained on one deployment holds a single empty-versus-occupied decision threshold zero-shot across nearly all transfer cells, reaching binary F1 up to 0.99 where standard normalization drops to 0.87 or fails outright, with no target-domain data or retraining. The transfer is scoped to binary presence by construction, as distinguishing static from moving motion requires the absolute magnitude that temporal standard deviation removes. We release the source code and dataset to support reproducible cross-environment CSI research.

Figures

Figures reproduced from arXiv: 2607.26665 by Bruno Rodrigues, Karim Khamaisi, Simon Sigg.

Figure 1
Figure 1. Figure 1: OpenCSI normalises raw CSI from heterogeneous hardware [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: OpenCSI pipeline. Raw CSI from a heterogeneous ESP32 mesh flows through three phase bands (Acquire, Calibrate, Score) and [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Floorplans (left of each pair) and photographs (right) of the three deployment environments (hardware variants in Table II). [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Z-score distributions across deployments (one column per session). Top: standard session-normalized amplitude Z-score, per [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Cross-environment F1 and OpenCSI calibration maturity vs. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

31 extracted references · 2 linked inside Pith

  1. [1]

    Occupancy detection in non-residential buildings: A survey and novel privacy preserved occupancy monitoring solution,

    J. Ahmad, H. Larijani, R. Emmanuel, M. Mannion, and A. Javed, “Occupancy detection in non-residential buildings: A survey and novel privacy preserved occupancy monitoring solution,”Applied Sciences, Vol. 11, No. 24, p. 11856, 2021

    2021

  2. [2]

    Cross-domain WiFi sensing with channel state information: A survey,

    Z. Chen, L. Zhang, C. Jiang, J. Cao, and W . Cui, “Cross-domain WiFi sensing with channel state information: A survey,”ACM Computing Surveys, Vol. 56, No. 1, pp. 1–35, 2023

    2023

  3. [3]

    AX-CSI: Enabling CSI extraction on commercial 802.11ax Wi-Fi platforms,

    M. Cominelli, F . Gringoli, and F . Restuccia, “AX-CSI: Enabling CSI extraction on commercial 802.11ax Wi-Fi platforms,” inProc. 15th ACM Workshop on Wireless Network Testbeds, Experimental Evaluation & Characterization (WiNTECH). ACM, 2021, pp. 46–53

    2021

  4. [4]

    A review of smart building sensing system for better indoor environment control,

    B. Dong, V . Prakash, F . Feng, and Z. O’Neill, “ A review of smart building sensing system for better indoor environment control,”Energy and Buildings, Vol. 199, pp. 29–46, 2019

    2019

  5. [5]

    ESP-CSI: ESP32 Wi-Fi channel state information,

    Espressif Systems, “ESP-CSI: ESP32 Wi-Fi channel state information,” https://github.com/espressif/esp-csi, 2023

    2023

  6. [6]

    ESP-IDF v5.5 Wi-Fi CSI API: HE-LTF Acquisition on ESP32-C6,

    ——, “ESP-IDF v5.5 Wi-Fi CSI API: HE-LTF Acquisition on ESP32-C6,” https://docs.espressif.com/projects/esp-idf/en/v5.5/esp32c6/api-gui des/wifi.html, 2025

    2025

  7. [7]

    Free your CSI: A channel state information extraction platform for modern Wi-Fi chipsets,

    F . Gringoli, M. Schulz, J. Link, and M. Hollick, “Free your CSI: A channel state information extraction platform for modern Wi-Fi chipsets,” inProc. 13th International Workshop on Wireless Network Testbeds, Experimental Evaluation & Characterization (WiNTECH). ACM, 2019, pp. 21–28

    2019

  8. [8]

    Tool release: Gather- ing 802.11n traces with channel state information,

    D. Halperin, W . Hu, A. Sheth, and D. Wetherall, “Tool release: Gather- ing 802.11n traces with channel state information,” ACM SIGCOMM Computer Communication Review, p. 53, 2011

    2011

  9. [9]

    IEEE 802.11bf: WLAN sensing,

    IEEE Standards Association, “IEEE 802.11bf: WLAN sensing,” IEEE 802.11 Task Group bf, 2024, https://www.ieee802.org/11/Reports/t gbf_update.htm

    2024

  10. [10]

    Scale what counts, mask what matters: Evaluating foundation models for zero-shot cross- domain Wi-Fi sensing,

    C. Jiang, Y. Yan, Y. Wang, C. T . Chou, and W . Hu, “Scale what counts, mask what matters: Evaluating foundation models for zero-shot cross- domain Wi-Fi sensing,”arXiv preprint arXiv:2511.18792, 2025

    arXiv 2025

  11. [11]

    Towards environment independent device free human activity recognition,

    W . Jiang, C. Miao, F . Ma, S. Yao, Y. Wang, Y. Yuan, H. Xue, C. Song, X. Ma, D. Koutsonikolas, W . Xu, and L. Su, “Towards environment independent device free human activity recognition,” inProc. 24th Annual International Conference on Mobile Computing and Networking (MobiCom). ACM, 2018, pp. 289–304

    2018

  12. [12]

    PicoScenes: A versatile Wi-Fi channel state information toolkit for IEEE 802.11n/ac/ax sensing applications,

    Z. Jiang, T . H. Luan, X. Ren, D. Lv, H. Hao, J. Wang, K. Zhao, W . Xi, Y. Xu, and R. Li, “PicoScenes: A versatile Wi-Fi channel state information toolkit for IEEE 802.11n/ac/ax sensing applications,”ACM Transactions on Sensor Networks, Vol. 18, No. 2, pp. 1–36, 2022

    2022

  13. [13]

    Wifi sensing with channel state information: A survey,

    Y. Ma, G. Zhou, and S. Wang, “Wifi sensing with channel state information: A survey,”ACM Computing Surveys, Vol. 52, No. 3, pp. 1–36, 2019

    2019

  14. [14]

    Big Brother is Watching You: Non-Intrusive ZigBee User Profiling,

    K. O. E. Müller, D. Datsomor, D. Schumm, B. Rodrigues, and B. Stiller, “Big Brother is Watching You: Non-Intrusive ZigBee User Profiling,” in Proc. 20th International Conference on Network and Service Manage- ment (CNSM). IEEE, 2024, pp. 1–7

    2024

  15. [15]

    Esp32 based low-power and low-cost wireless sensor network,

    A. Perdomo-Campos, I. Vega-González, and J. Ramírez-Beltrán, “Esp32 based low-power and low-cost wireless sensor network,” inThe confer- ence on Latin America Control Congress. Springer, 2020, pp. 275–285

    2020

  16. [16]

    Optimal preprocessing of WiFi CSI for sensing applications,

    V . V . Ratnam, H. Chen, H. H. Chang, A. Sehgal, and J. C. Zhang, “Optimal preprocessing of WiFi CSI for sensing applications,”IEEE Transactions on Wireless Communications, Vol. 23, No. 9, pp. 10 820– 10 833, 2024, arXiv:2307.12126

    Pith/arXiv arXiv 2024

  17. [17]

    BluePIL: a Bluetooth-based PassIve Localization Method,

    B. Rodrigues, C. Halter, M. Franco, E. J. Scheid, C. Killer, and B. Stiller, “BluePIL: a Bluetooth-based PassIve Localization Method,” inProc. IFIP/IEEE International Symposium on Integrated Network Manage- ment (IM). IEEE, 2021, pp. 28–36

    2021

  18. [18]

    Less is More: The Dilution Effect in Multi-Link Wireless Sensing,

    B. Rodrigues and K. Khamaisi, “Less is More: The Dilution Effect in Multi-Link Wireless Sensing,” 2026

    2026

  19. [19]

    Towards deep learning- based occupancy detection via wifi sensing in unconstrained environ- ments,

    C. Turetta, G. Skenderi, L. Capogrosso, F . Demrozi, P . H. Kindt, A. Mas- rur, F . Fummi, M. Cristani, and G. Pravadelli, “Towards deep learning- based occupancy detection via wifi sensing in unconstrained environ- ments,” in2023 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 2023, pp. 1–6

    2023

  20. [20]

    Device-free human activity recognition using commercial WiFi devices,

    W . Wang, A. X. Liu, M. Shahzad, K. Ling, and S. Lu, “Device-free human activity recognition using commercial WiFi devices,”IEEE Journal on Selected Areas in Communications, Vol. 35, No. 5, pp. 1118–1131, 2017

    2017

  21. [21]

    Placement matters: Understanding the effects of device placement for WiFi sensing,

    X. Wang, K. Niu, J. Xiong, B. Qian, Z. Yao, T . Lou, and D. Zhang, “Placement matters: Understanding the effects of device placement for WiFi sensing,”Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, Vol. 6, No. 1, pp. 1–25, 2022

    2022

  22. [22]

    Note on a method for calculating corrected sums of squares and products,

    B. P . Welford, “Note on a method for calculating corrected sums of squares and products,”Technometrics, Vol. 4, No. 3, pp. 419–420, 1962

    1962

  23. [23]

    Radio Tomographic Imaging with Wireless Networks,

    J. Wilson and N. Patwari, “Radio Tomographic Imaging with Wireless Networks,”IEEE Transactions on Mobile Computing, Vol. 9, No. 5, pp. 621–632, 2010

    2010

  24. [24]

    SenseFi: A library and benchmark on deep-learning-empowered WiFi human sensing,

    J. Yang, X. Chen, D. Wang, H. Zou, C. X. Lu, S. Sun, L. Xie, and Y. Lu, “SenseFi: A library and benchmark on deep-learning-empowered WiFi human sensing,”Patterns, Vol. 4, No. 3, p. 100703, 2023

    2023

  25. [25]

    EfficientFi: Toward large-scale lightweight WiFi sensing via CSI compression,

    J. Yang, X. Chen, H. Zou, D. Wang, Q. Xu, and L. Xie, “EfficientFi: Toward large-scale lightweight WiFi sensing via CSI compression,”IEEE Internet of Things Journal, Vol. 9, No. 15, pp. 13 086–13 095, 2022

    2022

  26. [26]

    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 Communications Magazine, Vol. 55, No. 10, pp. 98–104, 2017

    2017

  27. [27]

    Boosting WiFi sensing performance via CSI ratio,

    D. Zhang, Y. Zeng, D. Wu, and J. Xiong, “Boosting WiFi sensing performance via CSI ratio,”IEEE Pervasive Computing, Vol. 20, No. 1, pp. 62–70, 2021

    2021

  28. [28]

    CrossSense: Towards cross-site and large-scale WiFi sensing,

    J. Zhang, Z. Tang, M. Li, D. Fang, P . Nurmi, and Z. Wang, “CrossSense: Towards cross-site and large-scale WiFi sensing,” inProc. 24th Annual International Conference on Mobile Computing and Networking (Mo- biCom). ACM, 2018, pp. 305–320

    2018

  29. [29]

    Widar3.0: Zero-effort cross-domain gesture recognition with WiFi,

    Y. Zhang, Y. Zheng, K. Qian, G. Zhang, Y. Liu, C. Wu, and Z. Yang, “Widar3.0: Zero-effort cross-domain gesture recognition with WiFi,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 44, No. 11, pp. 8671–8688, 2022

    2022

  30. [30]

    DGSense: A Do- main Generalization Framework for Wireless Sensing,

    R. Zhou, Y. Cheng, S. Li, H. Zhang, and C. Liu, “DGSense: A Do- main Generalization Framework for Wireless Sensing,”arXiv preprint arXiv:2502.08155, 2025

    Pith/arXiv arXiv 2025

  31. [31]

    CSI- Bench: A large-scale in-the-wild dataset for multi-task WiFi sensing,

    G. Zhu, Y. Hu, W . Gao, W .-H. Wang, B. Wang, and K. J. R. Liu, “CSI- Bench: A large-scale in-the-wild dataset for multi-task WiFi sensing,” inAdvances in Neural Information Processing Systems (NeurIPS), 2025, arXiv:2505.21866

    arXiv 2025