Pith. sign in

REVIEW 3 major objections 6 minor 30 references

WiFi-JEPA: Self-supervised Learning for WiFi-CSI 3D Human Pose Estimation

T0 review · 3 major / 6 minor · reviewed 2026-07-14 · grok-4.5

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

Pith's one-line read WiFi-JEPA learns CSI-native pose features by predicting masked antenna-link views, beating prior WiFi 3D estimators and using simulated primitives as useful pre-training data.

desk verdict Solid CSI-native SSL paper with real ablations; headline SOTA vs older WiFi methods is partly confounded by decoder/protocol, but the controlled SSL and sim-object results still hold. read the letter →

arxiv 2607.11064 v1 pith:AE7AAOSU submitted 2026-07-13 cs.CV

Doeon Kim , Jungyoon Lee , Seongsin Kim , Seong-heum Kim This is my paper · ORCID
classification cs.CV
keywords WiFiCSI3Dhumanposeestimationself-supervisedlearningJEPAlinkmaskingray-tracingsimulationPerson-in-WiFi-3Dcross-domaingeneralization
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

WiFi Channel State Information can recover 3D human pose without cameras, but current estimators break under room or hardware shifts and depend on expensive camera labels. This paper argues that the right self-supervised objective is not to reconstruct raw CSI (which carries device noise) but to predict latent embeddings of entire masked transmitter–receiver links, forcing the model to use cross-link spatial correlations that encode 3D structure. Paired with a ray-tracing pipeline that turns randomized geometric primitives into unlabeled CSI, the method reaches new state-of-the-art single- and multi-person accuracy on Person-in-WiFi-3D while nearly halving leave-one-room error. Simulated frames from simple moving shapes prove roughly as useful for pre-training as real frames, and combining both improves further—whereas four standard vision self-supervised recipes actually hurt performance. A sympathetic reader cares because the result shows a practical path to scalable, privacy-preserving pose sensing that does not wait for more annotated rooms.

What carries the argument

Link masking on CSI-specific (C,T,L) tokens: entire Tx–Rx antenna links are masked so the model must predict their latent embeddings from the remaining links, turning multi-view spatial redundancy into the pretext task.

What would settle it

Pre-train identical encoders on equal volumes of real CSI versus sim-object CSI, fine-tune both on the same held-out real rooms and hardware, and check whether the simulation-only model still closes most of the gap to real-only pre-training; a large remaining gap would falsify the dynamics-diversity hypothesis.

Watch

Extended reading notes

Core claim

On Person-in-WiFi-3D, WiFi-JEPA with real-plus-simulated pre-training reaches 76.8 mm single-person and 93.5 mm multi-person MPJPE, outperforming prior WiFi-CSI baselines under the same protocol, while CSI-specific tokenization plus link masking improve over training from scratch and four vision-native SSL objectives degrade it; simulated CSI from geometric primitives supplies complementary pre-training value comparable to real CSI.

Load-bearing premise

The claim that diverse multipath dynamics from randomized non-human geometric primitives transfer well enough that roughly 90 000 simulated frames match the pre-training value of 90 000 real frames.

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 / 6 minor

Summary. The paper proposes WiFi-JEPA, a CSI-native self-supervised pre-training method for WiFi-CSI 3D human pose estimation. It tokenizes CSI as a factored (C,T,L) tensor, applies link masking (masking entire Tx–Rx links), and predicts latent embeddings under a JEPA objective rather than reconstructing raw CSI. A ray-tracing pipeline (sim-object) generates unlabeled CSI from randomized geometric primitives for pre-training. On Person-in-WiFi-3D, real+sim pre-training reports 76.8 mm single-person and 93.5 mm multi-person MPJPE, with ablations arguing that link masking, CSI-specific tokenization, and JEPA outperform four vision-native SSL objectives (which degrade below training from scratch) and that sim-object is complementary to real CSI.

Significance. If the controlled claims hold, this is a useful contribution to WiFi sensing: it shows that axis-aware latent prediction and multi-link spatial pretext tasks can transfer to multi-person 3D pose, and that non-human multipath dynamics can provide scalable unlabeled pre-training. Strengths include same-decoder SSL comparisons (Table 6), masking/tokenization ablations (Tables 7–8), sim-object vs sim-human vs real (Table 5), person-count and extremity breakdowns, and leave-one-environment-out tests. The negative result that standard vision SSL objectives hurt CSI pose estimation is particularly informative for the community.

major comments (3)
  1. [Table 1; Sec. 2.2; Sec. 4.4; Sec. 5.2] Contribution (iii) and Table 1 claim SOTA “under the same evaluation protocol,” but the headline gaps vs DT-Pose/PiW3D/MetaFi++/HPE-Li confound the pre-training objective with decoder and protocol choices. Sec. 2.2 and 4.4 note DT-Pose uses a GCN–Transformer with skeleton topology constraints while WiFi-JEPA uses PETR; decoder architecture is controlled only in Sec. 5.4 (Table 6). The same PETR from-scratch baseline is already 102.4 mm (Tables 5–6), so a substantial fraction of the published-number gap may be head/protocol rather than link-masked JEPA. Please either reimplement the main prior methods with the PETR head (or re-run WiFi-JEPA with their decoders) under matched fine-tuning, or reframe Table 1/abstract claims to separate “best reported numbers” from “SSL contribution under matched architecture.”
  2. [Sec. 3; Sec. 5.3; Table 5] The sim-object complementarity claim (Sec. 3 hypothesis; Table 5) is interesting but currently rests on a single matched-frame comparison (90K sim-object 100.1 mm vs 90K real 97.1 mm; real+sim 93.5 mm) and one negative sim-human control. Because free parameters of the simulator (20 RT sub-frame passes, room/material/speed ranges, 1–4 primitives) are tuned to the PiW3D SIMO layout, it is unclear whether “dynamics diversity > geometric fidelity” generalizes beyond this hardware/frequency setup. At minimum, report variance over multiple sim seeds/randomization ranges and a transfer test where sim is generated with deliberately mismatched antenna geometry or carrier frequency, or qualify the claim as in-domain unlabeled augmentation rather than general sim-to-real pre-training.
  3. [Abstract; Sec. 1; Table 2; Limitations] Cross-domain results (Table 2) show a large relative gain (mean 324.2 vs 626.4 mm) but absolute errors remain ~3–4× same-environment performance, and PA-MPJPE gains are small relative to MPJPE (Tables 1, 6). The abstract/intro emphasize environment-shift robustness as a primary motivation. Please align claims with these numbers: state clearly that SSL reduces but does not solve domain shift, and add the few-shot DANN comparison protocol details (train/test splits, which labels are 10%, whether pre-training includes target CSI) so the 171.3 vs 208.6 mm result is reproducible.
minor comments (6)
  1. [Table 1] Table 1 marks several prior metrics with † from original papers and leaves multi-person blank for methods that never reported it; make explicit in the caption that those rows are not re-evaluated under your PETR pipeline.
  2. [Sec. 4.3; Eq. (1)] Eq. (1) uses Smooth L1 on layer-normalized target embeddings; briefly justify LN on targets (stability vs. collapsing scale) and whether unnormalized targets change downstream MPJPE.
  3. [Sec. 4.1; Fig. 3] Fig. 3 and Sec. 4.1: clarify whether amplitude and phase are stacked as 60 channels before the linear (1,1) embedding or processed with separate projections; the text says “linear embedding with patch size (1,1)” but does not specify input channel handling.
  4. [Fig. 6] Fig. 6 qualitative poses would be more informative with per-scene MPJPE and person-count labels; currently it is hard to judge failure modes.
  5. [Sec. 2.1] Related work: WirelessJEPA and AM-FM are cited as concurrent/adjacent; a short table contrasting input (raw IQ vs estimated CSI), masking, and downstream tasks would help readers place the contribution.
  6. [Abstract; Introduction] Minor typos/spacing: “signaltoreal,” “fourvision-nativeSSLobjectivesdegradeper-formance,” and similar concatenation artifacts in the abstract/intro should be cleaned.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: empirical SSL pre-train/fine-tune pipeline whose reported MPJPE is an external pose metric, not a quantity forced by the training objective or by self-citation.

full rationale

WiFi-JEPA is a standard self-supervised representation-learning paper. Pre-training (JEPA latent prediction under link masking on real and/or ray-traced CSI) uses only unlabeled CSI tensors; fine-tuning attaches a PETR decoder and optimizes supervised pose losses on PiW3D labels; evaluation reports MPJPE/PA-MPJPE/PCK against camera-derived 3D joints. None of these quantities is defined by, or algebraically equivalent to, the Smooth-L1 JEPA loss, the EMA target encoder, the 0.6 link-mask rate, or the sim-object generation procedure. The simulation pipeline produces unlabeled multipath trajectories from randomized geometric primitives and is never fitted to pose labels; Table 5 simply measures transfer of the resulting encoder. Ablations (Tables 6–8) retrain the identical backbone+decoder under alternative SSL objectives or masking schemes, so the relative gains are not tautological. Cross-paper SOTA numbers (Table 1) may be confounded by decoder architecture, but that is a protocol-fairness issue, not a circular reduction of the claimed result to its own inputs. No self-definitional equations, no fitted free parameters re-labeled as predictions, no load-bearing self-citations of uniqueness theorems, and no ansatz smuggled via prior work by the same authors appear in the derivation chain. The paper is therefore self-contained against external benchmarks.

Assumptions & free parameters 6 free parameters · 5 assumptions · 2 invented entities

Load-bearing content is empirical method design plus standard SSL/wireless assumptions, not a formal derivation. Free parameters are training and simulation knobs chosen by the authors. Axioms are domain facts about CSI structure and JEPA-style latent prediction. Invented entities are the named method pieces (WiFi-JEPA, link masking recipe, sim-object pipeline) introduced to operationalize those assumptions.

free parameters (6)
  • link masking rate r
    Set to 0.6 (mask 5 of 9 links); defines the pretext task difficulty and is not derived from first principles.
  • EMA momentum schedule for target encoder
    Cosine schedule 0.996→1.0; standard JEPA hyperparameter chosen by authors.
  • pretrain/finetune learning rates and encoder LR scale 0.1×
    AdamW 5e-4 pretrain, 2e-5 finetune base with 0.1× encoder; fitted for stability/transfer, not predicted.
  • ViT depth/width (12-layer 512-d encoder; 8-layer predictor; PETR 5-layer 256-d)
    Architecture capacity choices that affect reported MPJPE.
  • 20 independent ray-tracing sub-frame passes per ~50 ms window
    Hand-chosen to inject temporal variation so temporal structure exists for masking; critical sim design knob.
  • sim scene randomization ranges (room 3–8 m, 1–4 primitives, speeds ≤3 m/s, materials)
    Domain-randomization ranges chosen to generate diversity; not identified from real PiW3D statistics alone.
assumptions (5)
  • domain assumption Raw CSI contains hardware-specific artifacts that make pixel/signal reconstruction a worse SSL target than latent prediction (JEPA-style).
    Stated in abstract/intro and used to justify Smooth-L1 on stop-grad LN target embeddings rather than MAE-style reconstruction (Sec. 1, 4.3).
  • domain assumption CSI factors as (C,T,L) with L independent spatial views of the same scene; cross-link correlation is informative of 3D structure.
    Underpins CSI-specific tokenization and link masking (Sec. 4.1–4.2).
  • ad hoc to paper Dynamics diversity of multipath over time/frequency/space can matter more for SSL transfer than scatterer geometric fidelity.
    Explicit hypothesis in Sec. 3; tested via sim-object vs sim-human but remains a modeling premise for the simulation contribution.
  • domain assumption PiW3D labels, PhaseFi phase denoising, and PETR-style matching losses are adequate supervision for evaluating 3D pose from CSI.
    Inherited evaluation stack from PiW3D/PETR used throughout Sec. 5.
  • standard math Standard transformer/JEPA training math (attention, EMA teacher, Smooth L1) is valid for this modality.
    Background ML machinery assumed without re-proof (Sec. 4.3).
invented entities (2)
  • WiFi-JEPA (CSI-native JEPA with link masking)
    purpose: Learn transferable CSI representations by predicting masked link latents for downstream 3D pose.
    Named method combining JEPA with CSI-specific tokenization/masking; evidence is internal ablations and PiW3D gains, not independent external theory.
  • sim-object ray-tracing pre-training corpus
    purpose: Provide scalable unlabeled CSI from randomized geometric primitives without pose labels.
    New data pipeline for this paper; transfer value shown only via the authors’ downstream MPJPE tables.

how reviews work

0 comments
Cite this review

Pith. "Pith review of WiFi-JEPA: Self-supervised Learning for WiFi-CSI 3D Human Pose Estimation." pith.science (2026). https://pith.science/paper/AE7AAOSU

@misc{pith2026260711064,
  author       = {Pith},
  title        = {Pith review of: WiFi-JEPA: Self-supervised Learning for WiFi-CSI 3D Human Pose Estimation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/AE7AAOSU}},
  note         = {Machine review of arXiv:2607.11064}
}
read the original abstract

WiFi Channel State Information (CSI) enables privacy-preserving human pose sensing in camera-denied environments, but existing WiFi-based pose estimators often fail under environment shifts and rely on costly camera-based annotation pipelines that limit scale. We propose WiFi-JEPA, a self-supervised framework that learns CSI-native representations by predicting masked latent embeddings instead of reconstructing raw CSI signals that may contain hardware-specific artifacts. WiFi-JEPA makes three contributions: (i) CSI-specific tokenization and link masking tailored to the CSI tensor over channel, time, and link (C,T,L); masking entire Tx-Rx antenna links forces the model to predict one spatial link view from others, capturing cross-link correlations informative of 3D spatial structure. (ii) A ray-tracing CSI simulation pipeline that generates diverse unlabeled CSI from randomized geometric primitives, providing scalable pre-training data without pose annotations. (iii) State-of-the-art results on Person-in-WiFi-3D: WiFi-JEPA outperforms prior WiFi-CSI baselines on both single- and multi-person 3D pose estimation under the same evaluation protocol. We also show that simulated CSI provides complementary pre-training signal to real CSI, and that four vision-native SSL objectives degrade performance below training from scratch, whereas WiFi-JEPA consistently improves downstream pose estimation.

Figures

Figures reproduced from arXiv: 2607.11064 by the authors.

Figure 1
Figure 1. Overall framework of WiFi-JEPA. Left: Pre-training data — sim-object and real CSI from PiW3D. Center : Generated CSI input and WiFi-JEPA. Right: GT and predicted 3D poses. WiFi channel state information (CSI) provides a compelling alternative. WiFi signals are ubiquitous indoors and can propagate through many common obstructions. Human motion modulates multipath propagation, which is cap￾tured as CSI—a factored, com… view at source ↗
Figure 2
Figure 2. sim-object pipeline. Top: Analogy to FractalDB and Dead Leaves—geometric primitives replace human models. Bottom: Stage 1 generates randomized scenes (CPU-only); Stage 2 runs Sionna RT with ×20 independent passes per frame. Inspired by the success of non-semantic pre-training in vision (Sec. 2.3), we randomize room geometry, object count, trajectory, and wall materials to cover a wide range of channel conditions. Hy… view at source ↗
Figure 3
Figure 3. (a) Flattening mixes temporal and link dimensions, causing patches to cross physical boundaries. (b) Our CSI-specific tokenization keeps (T, L) separate so each token corresponds to a specific spatio-temporal coordinate. The raw CSI data in the PiW3D dataset [28] is a tensor of shape Nrx ×Nant × T ×Nc = 3×3×20×30, where Nrx is the number of receivers, Nant is the number of antennas per receiver, T is the number of t… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Four masking strategies on the (T, L) token grid. The vertical axis corresponds to the 20 time steps, and the horizontal axis represents the 9 antenna links. We propose link masking, which masks entire columns of the (T, L) token grid, as shown in [PITH_FULL_IMAGE:fig…
Figure 5
Figure 5. Figure 5: WiFi-JEPA architecture. Phase 1 (top): self-supervised pre-training with link masking on the (T, L) token grid. Phase 2 (bottom): supervised fine-tuning with a PETR decoder for multi-person 3D pose estimation. We minimize the Smooth L1 loss between predicted and target…
Figure 6
Figure 6. Figure 6: Qualitative comparison of 3D pose estimation. Top row: PiW3D baseline [28]. Bottom row: WiFi-JEPA (ours). From left to right: 1-person, 2-person, and 3-person scenes. Challenging Scenarios Cross-Domain Generalization WiFi CSI amplitude profiles vary across environ￾ment…
Figure 7
Figure 7. Figure 7: Amplitude distribution shift across 3 environments [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

30 extracted references · 1 canonical work pages

  1. [1]

    arXiv preprint arXiv:1902.06435 (2019)

    Alkhateeb, A.: DeepMIMO: A generic deep learning dataset for millimeter wave and massive MIMO applications. arXiv preprint arXiv:1902.06435 (2019)

    arXiv 1902

  2. [2]

    arXiv preprint arXiv:2506.09985 (2025)

    Assran, M., Bardes, A., Fan, D., Garrido, Q., Howes, R., Komeili, M., Muckley, M., Rizvi, A., Roberts, C., Sinha, K., Zholus, A., Arnaud, S., Gejji, A., Martin, A., Hogan, F.R., Dugas, D., Bojanowski, P., Khalidov, V., Labatut, P., Massa, F., Szafraniec, M., Krishnakumar, K., Li, Y., Ma, X., Chandar, S., Meier, F., LeCun, Y., Rabbat, M., Ballas, N.: V-JEP...

    arXiv 2025

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Assran, M., Duval, Q., Misra, I., Bojanowski, P., Vincent, P., Rabbat, M., LeCun, Y., Ballas, N.: Self-supervised learning from images with a joint-embedding pre- dictive architecture. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 15619–15629 (2023)

    2023

  4. [4]

    In: Advances in Neural Information Processing Systems (NeurIPS) (2021)

    Baradad, M., Wulff, J., Wang, T., Isola, P., Torralba, A.: Learning to see by looking at noise. In: Advances in Neural Information Processing Systems (NeurIPS) (2021)

    2021

  5. [5]

    arXiv preprint arXiv:2404.08471 (2024)

    Bardes, A., Garrido, Q., Ponce, J., Chen, X., Rabbat, M., LeCun, Y., Assran, M., Ballas, N.: Revisiting feature prediction for learning visual representations from video. arXiv preprint arXiv:2404.08471 (2024)

    arXiv 2024

  6. [6]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Caron, M., Touvron, H., Misra, I., Jégou, H., Mairal, J., Bojanowski, P., Joulin, A.: Emerging properties in self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 9650–9660 (2021)

    2021

  7. [7]

    In: International Conference on Machine Learning (ICML)

    Chen, T., Kornblith, S., Norouzi, M., Hinton, G.: A simple framework for con- trastive learning of visual representations. In: International Conference on Machine Learning (ICML). pp. 1597–1607 (2020)

    2020

  8. [8]

    In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision (ICCV)

    Chen, X., Xie, S., He, K.: An empirical study of training self-supervised vision transformers. In: Proceedings of the IEEE/CVF International Conference on Com- puter Vision (ICCV). pp. 9640–9649 (2021)

    2021

Show all 30 references
  1. [9]

    arXiv preprint arXiv:2501.09411 (2025)

    Chen, Y., Guo, J., Guo, S., Zhou, J., Tao, D.: Towards robust and realistic human pose estimation via WiFi signals. arXiv preprint arXiv:2501.09411 (2025)

    2025 arXiv

  2. [10]

    arXiv preprint arXiv:2601.20190 (2026)

    Chu, V., Mashaal, O., Abou-Zeid, H.: WirelessJEPA: A multi-antenna foun- dation model using spatio-temporal wireless latent predictions. arXiv preprint arXiv:2601.20190 (2026)

    2026

  3. [11]

    JMLR17(59), 1–35 (2016)

    Ganin, Y., Ustinova, E., Ajakan, H., Germain, P., Larochelle, H., Laviolette, F., Marchand, M., Lempitsky, V.: Domain-adversarial training of neural networks. JMLR17(59), 1–35 (2016)

    2016

  4. [12]

    In:EuropeanConferenceonComputerVision(ECCV).LectureNotesinComputer Science, vol

    Gian, T.D., Lai, T.D., Luong, T.V., Wong, K.S., Nguyen, V.D.: HPE-Li: WiFi- enabled lightweight dual selective kernel convolution for human pose estimation. In:EuropeanConferenceonComputerVision(ECCV).LectureNotesinComputer Science, vol. 15089, pp. 93–111. Springer (2024)

    2024

  5. [13]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Grill, J.B., Strub, F., Altché, F., Tallec, C., Richemond, P., Buchatskaya, E., Do- ersch, C., Avila Pires, B., Guo, Z., Gheshlaghi Azar, M., Piot, B., Kavukcuoglu, K., Munos, R., Valko, M.: Bootstrap your own latent – a new approach to self-supervised learning. In: Advances i...

    2020

  6. [14]

    He, K., Chen, X., Xie, S., Li, Y., Dollár, P., Girshick, R.: Masked autoencoders are scalablevisionlearners.In:ProceedingsoftheIEEE/CVFConferenceonComputer Vision and Pattern Recognition (CVPR). pp. 16000–16009 (2022) WiFi-JEPA 17

    2022

  7. [15]

    arXiv preprint arXiv:2303.11103 (2023)

    Hoydis, J., Ait Aoudia, F., Cammerer, S., Nimier-David, M., Binder, N., Marcus, G., Keller, A.: Sionna RT: Differentiable ray tracing for radio propagation model- ing. arXiv preprint arXiv:2303.11103 (2023)

    2023 arXiv

  8. [16]

    In: Proceedings of the 26th Annual International Conference on Mobile Computing and Networking (MobiCom)

    Jiang, W., Xue, H., Miao, C., Wang, S., Lin, S., Tian, C., Murali, S., Hu, H., Sun, Z., Su, L.: Towards 3D human pose construction using WiFi. In: Proceedings of the 26th Annual International Conference on Mobile Computing and Networking (MobiCom). pp. 1–14. ACM (2020)

    2020

  9. [17]

    In: Asian Confer- ence on Computer Vision (ACCV)

    Kataoka, H., Okayasu, K., Matsumoto, A., Yamagata, E., Yamada, R., Inoue, N., Nakamura, A., Satoh, Y.: Pre-training without natural images. In: Asian Confer- ence on Computer Vision (ACCV). pp. 583–600 (2020)

    2020

  10. [18]

    arXiv preprint arXiv:2512.04723 (2025)

    Liu, G., Hao, Y., Zou, Y.: CIG-MAE: Cross-modal information-guided masked au- toencoder for self-supervised WiFi sensing. arXiv preprint arXiv:2512.04723 (2025)

    2025

  11. [19]

    Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies6(2), 1–25 (2022)

    Ren, Y., Wang, Z., Wang, Y., Tan, S., Chen, Y., Yang, J.: GoPose: 3D human pose estimation using WiFi. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies6(2), 1–25 (2022)

    2022

  12. [20]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Shi, D., Wei, X., Li, L., Ren, Y., Tan, W.: End-to-end multi-person pose estimation with transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 11069–11078 (2022)

    2022

  13. [21]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Sun, K., Xiao, B., Liu, D., Wang, J.: Deep high-resolution representation learning for visual recognition. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 5693–5703 (2019)

    2019

  14. [22]

    In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)

    Tobin, J., Fong, R., Ray, A., Schneider, J., Zaremba, W., Abbeel, P.: Domain randomization for transferring deep neural networks from simulation to the real world. In: IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). pp. 23–30 (2017)

    2017

  15. [23]

    In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)

    Wang, F., Zhou, S., Panev, S., Han, J., Huang, D.: Person-in-WiFi: Fine-grained person perception using WiFi. In: Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV). pp. 5452–5461 (2019)

    2019

  16. [24]

    In: IEEE Global Communications Conference (GLOBECOM)

    Wang, X., Gao, L., Mao, S.: PhaseFi: Phase fingerprinting for indoor localiza- tion with a deep learning approach. In: IEEE Global Communications Conference (GLOBECOM). pp. 1–6 (2015)

    2015

  17. [25]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Xie, Z., Zhang, Z., Cao, Y., Lin, Y., Bao, J., Yao, Z., Dai, Q., Hu, H.: SimMIM: A simple framework for masked image modeling. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 9653–9663 (2022)

    2022

  18. [26]

    ACM Transactions on Sensor Networks21, 21:1–21:38 (2025).https://doi.org/10.1145/3715130

    Xu, K., Wang, J., Zhu, H., Zheng, D.: Evaluating self-supervised learning for WiFi CSI-based human activity recognition. ACM Transactions on Sensor Networks21, 21:1–21:38 (2025).https://doi.org/10.1145/3715130

    2025 doi

  19. [27]

    In: Advances in Neural Information Processing Systems (NeurIPS)

    Xu,Y.,Zhang,J.,Zhang,Q.,Tao,D.:ViTPose:Simplevisiontransformerbaselines for human pose estimation. In: Advances in Neural Information Processing Systems (NeurIPS). vol. 35, pp. 38571–38584 (2022)

    2022

  20. [28]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Yan, K., Wang, F., Qian, B., Ding, H., Han, J., Wei, X.: Person-in-WiFi 3D: End-to-end multi-person 3D pose estimation with Wi-Fi. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 969–978 (2024)

    2024

  21. [29]

    IEEE Internet of Things Journal10(16), 14128–14136 (2023)

    Zhou, Y., Huang, H., Yuan, S., Zou, H., Xie, L., Yang, J.: MetaFi++: WiFi-enabled transformer-based human pose estimation for metaverse avatar simulation. IEEE Internet of Things Journal10(16), 14128–14136 (2023)

    2023

  22. [30]

    arXiv preprint arXiv:2602.11200 (2026)

    Zhu, G., Hu, Y., Jayaweera, S., Gao, W., Wang, W.H., Zhang, J., Wang, B., Wu, C., Liu, K.J.R.: AM-FM: A foundation model for ambient intelligence through WiFi. arXiv preprint arXiv:2602.11200 (2026)

    2026

Pith tools

Reviewed July 14, 2026 · model on record in the stance chip above.