Pith. sign in

REVIEW 2 major objections 5 minor 57 references

Wave2Body translates mmWave radar point clouds into a pretrained discrete body-token space, rather than regressing coordinates directly, and reports lower cross-domain pose error with far less computation.

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 ↗

Radar-to-body token translation with a frozen, pose-pretrained body tokenizer beats direct coordinate regression on mmWave pose benchmarks and cuts FLOPs dramatically.

T0 review reviewed 2026-08-01 challenge →

load-bearing objection Solid three-stage token-translation system for mmWave HPE, with honest evaluations, but the headline claim that decoupling causes the gains is under-supported by a missing control. the 2 major comments →

arxiv 2607.18875 v1 pith:5GGQ3ZLP submitted 2026-07-21 cs.CV

Wave2Body: Rethinking mmWave Human Pose Estimation as Radar-to-Body Token Translation

Bo Liang , Chen Gong , Wei Gao , Chenren Xu This is my paper
classification cs.CV
keywords mmWave radarhuman pose estimationtoken translationVQ-VAEself-supervised learningcross-domain generalizationpoint cloudradar sensing
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.

The reading

Wave2Body argues that mmWave human pose estimation should be treated as translating radar point clouds into a pretrained discrete space of body tokens, rather than directly regressing joint coordinates. The system trains a self-supervised radar tokenizer, a pose-only compositional VQ-VAE body tokenizer, and a lightweight translator between them, using paired radar-pose data only for cross-modal mapping and root localization. On the M4Human and mmBody benchmarks it reduces cross-subject MPJPE by 9.7%, cross-action by 3.8%, and mmBody by 8.2%, while cutting training FLOPs by 31x and inference FLOPs by 89.8x versus a matched baseline. The finding suggests that decoupling sensor representation, body structure, and alignment improves generalization to unseen subjects and actions.

Core claim

The central claim is that mapping radar tokens into a frozen, anatomically structured discrete token space—rather than predicting raw coordinates—produces better cross-domain generalization at far lower computational cost. Wave2Body's mmWave tokenizer, pretrained by masked patch reconstruction, encodes radar point clouds into continuous tokens; a part-structured VQ-VAE trained on AMASS and benchmark poses defines a 24-slot discrete codebook across five kinematic regions; and a query-based transformer translates the radar tokens into body tokens that a frozen decoder turns into a pelvis-centered skeleton, with a separate head estimating the pelvis root. On M4Human Cross-Subject and Cross-Acti

What carries the argument

The central mechanism is a compositional body tokenizer built as a part-structured VQ-VAE: the skeleton is split into five kinematic regions, each with its own encoder, 96-entry codebook, and decoder, producing 24 discrete token slots that jointly define the output space. This frozen discrete space acts as the prediction interface, so the radar model only has to produce embeddings that quantize to the right body tokens; a query-based translator with 24 learnable body queries cross-attends to radar tokens and self-attends across parts, and a lightweight root-localization head recovers global position. The design lets each module be pretrained on its natural data source—unlabeled radar for the

Load-bearing premise

The whole pipeline inherits its ceiling from the frozen body VQ-VAE: if the discrete codebook cannot reconstruct fine-grained poses accurately—its reconstruction error is never reported—then even a perfect radar-to-token translator is capped by that error, and quantization may explain why gains are smaller on Cross-Action In-Place.

What would settle it

Compare the body VQ-VAE's own reconstruction MPJPE on held-out poses from the same distribution as the test set. If that reconstruction error is equal to or larger than Wave2Body's end-to-end MPJPE, then the discrete body-token space, not the radar-to-token translation, sets the accuracy ceiling, directly confirming or refuting the paper's stated limitation.

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

If this is right

  • Paired radar-pose data is only needed for translation and root localization, so expanding to new subjects or sensors could require far less paired annotation than coordinate regression.
  • The frozen body-token interface is modality-agnostic; other sensing modalities could reuse it by training a new translator, reusing anatomical priors.
  • The discrete slots give an interpretable intermediate: token recall correlates with articulation error, so one can debug which body parts are poorly aligned.
  • The 89.8x inference FLOP reduction and 8x higher throughput make real-time on-device radar pose estimation far more practical.
  • Because the body tokenizer is pretrained on AMASS, the output space encodes plausible human configurations, which may reduce anatomically implausible skeletons from ambiguous radar frames.

Where Pith is reading between the lines

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

  • If the body-token space is expressive enough, one could adapt the translator to a new radar domain with only unpaired radar and an independent pose prior, e.g., by aligning token distributions—an extension the paper does not test.
  • The discrete token output suggests a natural fit for temporal sequence modeling over tokens, where a transformer could learn motion smoothness and resolve single-frame ambiguities that the current fixed-window aggregation leaves open.
  • The 31x training-FLOP reduction implies that the expensive part of learning is moved to pretraining stages that reuse generic data, which could make radar pose systems feasible for edge hardware and small lab deployments with limited paired data.
Share X Bluesky LinkedIn Reddit HN

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

2 major / 5 minor

Summary. The paper proposes Wave2Body, a three-stage mmWave human pose estimation method. It first pretrains a self-supervised mmWave point-cloud tokenizer, then pretrains a part-structured VQ-VAE body tokenizer on pelvis-centered pose data (AMASS plus benchmark training poses), and finally trains a lightweight transformer that translates frozen radar tokens into frozen body-token embeddings, with a separate head for root localization. The method is evaluated on M4Human and mmBody under cross-subject and cross-action protocols, reporting lower MPJPE than compared baselines in most settings, as well as substantially lower training and inference FLOPs. The central claim is that decoupling radar representation learning, body modeling, and cross-modal alignment into independently pretrained token spaces improves cross-domain generalization relative to direct coordinate regression.

Significance. If the causal claim holds, the paper makes a useful conceptual contribution: it changes the prediction interface for mmWave HPE from coordinate regression to translation between pretrained, reusable token spaces. The paper is strengthened by reporting three-seed means and standard deviations, profiling FLOPs with a documented protocol, providing token-recall analyses that go beyond aggregate MPJPE, and committing to public code and results. The efficiency gains are large and, if reproducible, practically relevant. However, the central generalization claim depends on a missing control that separates the effect of the decoupled architecture from the effect of the additional pose-only AMASS data.

major comments (2)
  1. [Sec. 4.4, Fig. 7; Sec. 4.2] The causal claim that decoupling improves generalization is not isolated from the extra AMASS data. The 'w/o body tokenizer' ablation removes both the compositional tokenizer and the AMASS pretraining, while every retained variant still uses the token-based translation pipeline. Thus the observed gains over WiCompass could in principle be due entirely to the additional pose-only data rather than to the decoupled interface. The missing control is a direct coordinate regressor (e.g., WiCompass) trained with the same AMASS pose-only data, for instance by pretraining its output decoder as a pose autoencoder or adding a pose-manifold regularizer. If that control matches the Cross-Subject and Cross-Action improvements, the paper's central claim is not supported; its claim would reduce to the far weaker statement that extra pose-only data helps. I request this control, or an explicit qualificat
  2. [Sec. 3.3, Eq. (5); Sec. 5] The frozen VQ-VAE body tokenizer defines the output space for all predicted poses, but its reconstruction fidelity is never reported. If the discrete codebook cannot accurately represent pelvis-centered poses, even a perfect radar-to-token translator cannot achieve low root-aligned MPJPE. The paper's Sec. 5 concession that 'quantization may limit fine-grained accuracy' does not quantify this ceiling. Please report VQ-VAE reconstruction MPJPE on held-out AMASS and benchmark pose-only sets, and compare that value with the observed end-to-end errors and with a continuous-decoder upper bound. This measurement is needed to substantiate the premise that the body-token space is a near-lossless target interface.
minor comments (5)
  1. [Throughout] Several formatting artifacts should be corrected: 'VQ-V AE' in Sec. 3.3, missing spaces in '31.00×fewer', and missing spacing around math in Sec. 3.1 and elsewhere. These do not affect content but reduce readability.
  2. [Fig. 3 caption] The legend/caption text 'Paired Same action / other subject Other action / same subject' is ambiguous. Please provide distinct, labeled legend entries for paired, same-action/other-subject, and other-action/same-subject conditions.
  3. [Fig. 6] The numeric labels in the figure are visually compressed and the relative magnitudes of '2.309 EFLOPs' and '74.495 PFLOPs' are not immediately clear. A small table or a clearer axis annotation would help the reader verify the 31.00x ratio.
  4. [Sec. 4.1] The paper states the only additional supervision is AMASS pose-only data, but it should also clarify whether the body tokenizer is trained separately per benchmark or jointly, and whether any AMASS subjects or motions overlap with the held-out M4Human/mmBody test subjects or action classes. This is relevant for interpreting the cross-domain gains.
  5. [Sec. 4.4] The four ablation variants are single-seed. Since the main results are reported with three seeds, it would be useful to know the seed-to-seed variability for at least the two most important ablations (w/o body tokenizer and AMASS-only body prior), as the observed differences in some action groups are small.

Circularity Check

0 steps flagged

No circular derivation: predictions are supervised against ground-truth poses/embeddings, and self-citations are not load-bearing.

full rationale

Wave2Body's derivation chain is not circular. Stage 1 (Sec. 3.2) pretrains the mmWave tokenizer by masked point-cloud reconstruction (Eq. 2) with no pose labels; Stage 2 (Sec. 3.3) trains the compositional VQ-VAE body tokenizer to reconstruct pelvis-centered poses from pose-only data (Eq. 5); Stage 3 (Sec. 3.4) trains the translator and root localizer with the embedding-alignment and pose-reconstruction losses in Eqs. 6 and 9, supervised by ground-truth poses and the frozen body encoder's embeddings. No fitted parameter is later reported as a prediction: the token-recall analysis (Sec. 4.2) is an evaluation metric computed on the test split, not a training target. The only overlap between training and evaluation is that the body tokenizer is pretrained on the benchmark training split's poses plus AMASS; test subjects/actions are held out, so there is no leakage. The self-cited WiCompass [32] serves as a re-evaluated comparison baseline, not as an input to Wave2Body's optimization; references [17,18,32] are contextual. The absence of a control that gives AMASS pose data to a direct regressor is a causal-attribution limitation (the AMASS prior may drive the gains), but it does not make any equation reduce to its inputs by construction. Hence no significant circularity; the score reflects only minor, non-load-bearing self-citations.

Axiom & Free-Parameter Ledger

4 free parameters · 6 axioms · 2 invented entities

The method is empirical ML; its central claim rests on architectural and data assumptions rather than on mathematical axioms. The list covers the hand-chosen hyperparameters and the domain assumptions about pose priors, pretraining transfer, and the representational sufficiency of the discrete body-token space.

free parameters (4)
  • Attribute-reconstruction loss weight gamma = not stated
    Eq. (2): hand-chosen weight for the radar-attribute reconstruction term in mmWave tokenizer pretraining.
  • Embedding/pose loss weights lambda_z, lambda_b = not stated
    Eq. (6): hand-chosen balance between pre-quantization embedding alignment and pelvis-centered pose reconstruction in translator training.
  • Localization loss weights lambda_r, lambda_y = not stated
    Eq. (9): hand-chosen weights for root-error and absolute-pose terms in the final joint phase.
  • Architecture hyperparameters (G=96, K=32, 128-D radar tokens, 60% masking, 24 body slots, five 96-entry codebooks, 256-D = as in Sec. 4.1
    Hand-chosen; no sensitivity analysis is reported for any of these values.
axioms (6)
  • domain assumption Pose annotations in M4Human and mmBody are accurate and expressed in the radar coordinate frame.
    Training targets and evaluation depend on this; Sec. 4.1.
  • domain assumption AMASS poses, together with benchmark training poses, are representative enough to train a body-token prior that transfers to radar benchmarks.
    Sec. 3.3 trains the body tokenizer on AMASS plus benchmark poses without paired radar.
  • domain assumption Masked point-cloud reconstruction (Point-MAE style) yields radar-token representations that preserve geometry and attributes useful for downstream pose.
    Sec. 3.2 / Eq. (2); no direct downstream-pose probe of the pretrained tokens is given beyond the full pipeline.
  • domain assumption A discrete compositional VQ-VAE with 96-entry per-region codebooks can represent human poses accurately enough for HPE.
    Sec. 3.3 / Eq. (5); this is the paper's weakest assumption and the reconstruction error is never reported.
  • domain assumption The straight-through estimator and standard VQ training objectives converge to useful codebooks.
    Standard practice, but no codebook-usage statistics are reported.
  • domain assumption SMPL 22-joint topology is a valid common skeleton for all datasets.
    Sec. 3.1 adopts J=22 following SMPL [35].
invented entities (2)
  • Radar-to-body token translation interface no independent evidence
    purpose: Couples a continuous radar-token source space to a frozen discrete body-token target space; the central design contribution.
    Empirical support is internal (token-recall analysis, MPJPE); there is no external falsifiable handle outside the paper.
  • Part-structured body token slots (five kinematic regions, 24 slots) no independent evidence
    purpose: Makes the output space compositional and interpretable by body region.
    Architectural construct; the region-level recall in Fig. 3 is internal evidence only.

reviewed 2026-08-01 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Wave2Body: Rethinking mmWave Human Pose Estimation as Radar-to-Body Token Translation." pith.science (2026). https://pith.science/paper/5GGQ3ZLP

@misc{pith2026260718875,
  author       = {Pith},
  title        = {Pith review of: Wave2Body: Rethinking mmWave Human Pose Estimation as Radar-to-Body Token Translation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5GGQ3ZLP}},
  note         = {Machine review of arXiv:2607.18875}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Millimeter-wave (mmWave) radar enables privacy-friendly human sensing, but its sparse point clouds are physical measurements of view-dependent electromagnetic reflections and only indirectly characterize body articulation. Recovering a complete 3D pose from such partial, geometry-dependent observations is therefore under-constrained. Existing methods directly regress joint coordinates from paired radar-pose data, relying on the same limited paired supervision to learn radar perception, human-body structure, and their alignment. This coupling can encourage dataset-specific shortcuts under ambiguous radar observations. We propose Wave2Body, a radar-to-body token translation framework that decouples these learning targets using a self-supervised mmWave tokenizer, a pretrained compositional body tokenizer that defines the output space, and a lightweight translator between them. Experiments on M4Human and mmBody show that Wave2Body achieves stronger cross-domain generalization than previous methods while incurring much lower computational costs for training and inference. All the code and experiment results are publicly available at https://github.com/Galaxywalk/Wave2Body.

Figures

Figures reproduced from arXiv: 2607.18875 by Bo Liang, Chen Gong, Chenren Xu, Wei Gao.

Figure 1
Figure 1. Figure 1: Wave2Body reformulates mmWave HPE from coor￾dinate regression to radar-to-body token translation. A self￾supervised mmWave tokenizer produces radar tokens, which are translated and quantized into discrete body tokens before decod￾ing; the body encoder is used only to construct training targets. 13, 45, 51] has been made through stronger point-cloud backbones, signal augmentation, diffusion-based pose es￾ti… view at source ↗
Figure 2
Figure 2. Figure 2: Three-Stage Training of Wave2Body. Stage 1 pretrains a continuous mmWave tokenizer through masked radar-patch re￾construction. Stage 2 pretrains a discrete compositional body to￾kenizer, implemented as a VQ-VAE, using pose-only data from motion-capture and radar datasets. Stage 3 freezes the pretrained radar and body modules and trains the radar-to-body translator and localization head using paired radar–p… view at source ↗
Figure 3
Figure 3. Figure 3: Body-token accuracy. M4Human Cross-Subject token￾ID accuracy. Error bars show standard deviations across three seeds. actions; its 3.8% micro-average gain also trails its 9.7% Cross-Subject gain. Benefits thus concentrate on subject shift and dynamic held-out motions, suggesting a trade-off between a discrete body prior’s robustness and direct coor￾dinate regression’s fine-grained precision (Sec. 5). Error… view at source ↗
Figure 4
Figure 4. Figure 4: M4Human per-joint error. Error bars show three-seed standard deviations; dashes mark the mean. tion ρ [41]. On Cross-Subject, Top-1/3/5 mean correlations are ρ = 0.595, ρ = 0.742, and ρ = 0.774; Cross-Action Top-1 is ρ = 0.615. Under Schober et al.’s scale [39], the Top-1 correlation is moderate, while the Top-3/5 cor￾relations are strong; token-alignment mismatch correlates with predicted-pose error. Toge… view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results from easy to challenging cases. Representative M4Human and mmBody test samples are selected across the absolute-MPJPE spectrum. Gray points denote the input mmWave point cloud, red skeletons denote ground truth, and blue skeletons denote Wave2Body predictions. The value in each panel is the sample-level radar-frame absolute MPJPE. 10 30 100 300 1k 3k Training FLOPs (PFLOPs, log scale) W… view at source ↗
Figure 6
Figure 6. Figure 6: Cold-start training computation on M4Human. Wave2Body is decomposed into two tokenizer-pretraining phases and two paired-data phases. FLOPs include supported profiler operations in forward, loss, backward, and optimizer computation [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Component ablations studies. Lower MPJPE is better. Subject and Cross-Action, which evaluate held-out identi￾ties and held-out action classes, respectively. All tokenizer pretraining remains restricted to the training split. Dual-Tokenizer Ablation. The variant without the body tokenizer retains the frozen mmWave tokenizer and local￾ization head, but replaces body-token translation and decod￾ing with a poo… 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

57 extracted references · 4 linked inside Pith

  1. [1]

    Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35:23716–23736,

    Jean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech, Iain Barr, Yana Hasson, Karel Lenc, Arthur Men- sch, Katherine Millican, Malcolm Reynolds, et al. Flamingo: a visual language model for few-shot learning.Advances in neural information processing systems, 35:23716–23736,

  2. [2]

    Fast and scalable human pose estimation using mmwave point cloud

    Sizhe An and Umit Y Ogras. Fast and scalable human pose estimation using mmwave point cloud. InProceedings of the 59th ACM/IEEE Design Automation Conference, pages 889–894, 2022. 1, 2

    2022

  3. [3]

    mri: Multi-modal 3d human pose estimation dataset using mmwave, rgb-d, and inertial sensors.Advances in neural information processing systems, 35:27414–27426, 2022

    Sizhe An, Yin Li, and Umit Ogras. mri: Multi-modal 3d human pose estimation dataset using mmwave, rgb-d, and inertial sensors.Advances in neural information processing systems, 35:27414–27426, 2022. 2

    2022

  4. [4]

    Immfusion: Robust mmwave-rgb fusion for 3d human body reconstruction in all weather conditions.arXiv preprint arXiv:2210.01346, 2022

    Anjun Chen, Xiangyu Wang, Kun Shi, Shaohao Zhu, Bin Fang, Yingfeng Chen, Jiming Chen, Yuchi Huo, and Qi Ye. Immfusion: Robust mmwave-rgb fusion for 3d human body reconstruction in all weather conditions.arXiv preprint arXiv:2210.01346, 2022. 2

    Pith/arXiv arXiv 2022

  5. [5]

    mmbody benchmark: 3d body recon- struction dataset and analysis for millimeter wave radar

    Anjun Chen, Xiangyu Wang, Shaohao Zhu, Yanxu Li, Jim- ing Chen, and Qi Ye. mmbody benchmark: 3d body recon- struction dataset and analysis for millimeter wave radar. In Proceedings of the 30th ACM International Conference on Multimedia, pages 3501–3510, 2022. 2, 4

    2022

  6. [6]

    Cpformer: End-to-end multi- person human pose estimation from raw radar cubes with transformers.IEEE Sensors Journal, 2025

    Lin Chen and Guoli Wang. Cpformer: End-to-end multi- person human pose estimation from raw radar cubes with transformers.IEEE Sensors Journal, 2025. 2

    2025

  7. [7]

    Dilated point spatio-temporal mesh transformer for mmwave radar-based human mesh re- construction

    Lin Chen and Guoli Wang. Dilated point spatio-temporal mesh transformer for mmwave radar-based human mesh re- construction. In2025 International Joint Conference on Neural Networks (IJCNN), pages 1–8. IEEE, 2025

    2025

  8. [8]

    Lin Chen, Cong Li, Shuxin Zhong, Jun Chen, Yufei Wen, Haotian Song, and Kaishun Wu. Sensing life in stillness: Unified dynamic and static human mesh reconstruction with mmwave radar.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, 10(1):1–25, 2026

    2026

  9. [9]

    Mvdoppler-pose: Multi-modal multi-view mmwave sensing for long-distance self-occluded human walking pose estima- tion

    Jaeho Choi, Soheil Hor, Shubo Yang, and Amin Arbabian. Mvdoppler-pose: Multi-modal multi-view mmwave sensing for long-distance self-occluded human walking pose estima- tion. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 27750–27759, 2025. 2

    2025

  10. [10]

    Multi-modal align- ment using representation codebook

    Jiali Duan, Liqun Chen, Son Tran, Jinyu Yang, Yi Xu, Belinda Zeng, and Trishul Chilimbi. Multi-modal align- ment using representation codebook. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15651–15660, 2022. 6

    2022

  11. [11]

    Tokenhmr: Advancing human mesh re- covery with a tokenized pose representation

    Sai Kumar Dwivedi, Yu Sun, Priyanka Patel, Yao Feng, and Michael J Black. Tokenhmr: Advancing human mesh re- covery with a tokenized pose representation. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 1323–1333, 2024. 2

    2024

  12. [12]

    Point 4d trans- former networks for spatio-temporal modeling in point cloud videos

    Hehe Fan, Yi Yang, and Mohan Kankanhalli. Point 4d trans- former networks for spatio-temporal modeling in point cloud videos. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 14204–14213,

  13. [13]

    Dif- fusion model is a good pose estimator from 3d rf-vision

    Junqiao Fan, Jianfei Yang, Yuecong Xu, and Lihua Xie. Dif- fusion model is a good pose estimator from 3d rf-vision. InEuropean Conference on Computer Vision, pages 1–18. Springer, 2024. 1, 2, 5

    2024

  14. [14]

    M4human: A large-scale multimodal mmwave radar benchmark for human mesh reconstruction

    Junqiao Fan, Yunjiao Zhou, Yizhuo Yang, Xinyuan Cui, Jiarui Zhang, Lihua Xie, Jianfei Yang, Chris Xiaoxuan Lu, and Fangqiang Ding. M4human: A large-scale multimodal mmwave radar benchmark for human mesh reconstruction. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 42836–42846, 2026. 1, 2, 4, 5

    2026

  15. [15]

    Vq-hps: Hu- man pose and shape estimation in a vector-quantized latent space

    Gu ´enol´e Fiche, Simon Leglaive, Xavier Alameda-Pineda, Antonio Agudo, and Francesc Moreno-Noguer. Vq-hps: Hu- man pose and shape estimation in a vector-quantized latent space. InEuropean Conference on Computer Vision, pages 471–490. Springer, 2024. 2

    2024

  16. [16]

    Human pose as compositional tokens

    Zigang Geng, Chunyu Wang, Yixuan Wei, Ze Liu, Houqiang Li, and Han Hu. Human pose as compositional tokens. In Proceedings of the IEEE/CVF Conference on Computer Vi- sion and Pattern Recognition, pages 660–671, 2023. 2

    2023

  17. [17]

    Data can speak for itself: Quality-guided utilization of wireless syn- thetic data

    Chen Gong, Bo Liang, Wei Gao, and Chenren Xu. Data can speak for itself: Quality-guided utilization of wireless syn- thetic data. InProceedings of the 23rd Annual International Conference on Mobile Systems, Applications and Services, pages 209–222, 2025. 2

    2025

  18. [18]

    Towards generalizable wireless sensing models via pre-training on multi-source datasets

    Chen Gong, Bo Liang, Qihao Zhu, Wei Gao, Yin Chen, Jin Nakazawa, and Chenren Xu. Towards generalizable wireless sensing models via pre-training on multi-source datasets. In Proceedings of the 2026 ACM/IEEE International Confer- ence on Embedded Artificial Intelligence and Sensing Sys- tems, pages 488–502, 2026. 2

    2026

  19. [19]

    Millimeter wave radar-based human activity recognition for healthcare monitoring robot

    Zhanzhong Gu, Xiangjian He, Gengfa Fang, Chengpei Xu, Feng Xia, and Wenjing Jia. Millimeter wave radar-based human activity recognition for healthcare monitoring robot. arXiv preprint arXiv:2405.01882, 2024. 1

    Pith/arXiv arXiv 2024

  20. [20]

    Dghmesh: A large- scale dual-radar mmwave dataset and generalization-focused benchmark for human mesh reconstruction.arXiv preprint arXiv:2604.22827, 2026

    Rongxiao Guo and Qingchao Chen. Dghmesh: A large- scale dual-radar mmwave dataset and generalization-focused benchmark for human mesh reconstruction.arXiv preprint arXiv:2604.22827, 2026. 2

    Pith/arXiv arXiv 2026

  21. [21]

    A two-stage motion-aware framework for mmwave-based hu- man mesh recovery.arXiv e-prints, pages arXiv–2605, 2026

    Hoang Hai Pham, Shuntian Zheng, Jiaqi Li, and Yu Guan. A two-stage motion-aware framework for mmwave-based hu- man mesh recovery.arXiv e-prints, pages arXiv–2605, 2026. 2

    2026

  22. [22]

    Rt-pose: A 4d radar tensor-based 3d human pose estimation and localization benchmark

    Yuan-Hao Ho, Jen-Hao Cheng, Sheng Yao Kuan, Zhongyu Jiang, Wenhao Chai, Hsiang-Wei Huang, Chih-Lung Lin, and Jenq-Neng Hwang. Rt-pose: A 4d radar tensor-based 3d human pose estimation and localization benchmark. In European Conference on Computer Vision, pages 107–125. Springer, 2024. 2

    2024

  23. [23]

    mmwave radar for sit-to-stand analysis: A compar- ative study with wearables and kinect.IEEE Transactions on Biomedical Engineering, 72(9):2623–2634, 2025

    Shuting Hu, Peggy Ackun, Xiang Zhang, Siyang Cao, Jen- nifer Barton, Melvin G Hector, Mindy J Fain, and Nima Too- sizadeh. mmwave radar for sit-to-stand analysis: A compar- ative study with wearables and kinect.IEEE Transactions on Biomedical Engineering, 72(9):2623–2634, 2025. 1

    2025

  24. [24]

    Waveman: mmwave- 9 based room-scale human interaction perception for hu- manoid robots.arXiv preprint arXiv:2601.07454, 2026

    Yuxuan Hu, Kuangji Zuo, Boyu Ma, Shihao Li, Zhaoyang Xia, Feng Xu, and Jianfei Yang. Waveman: mmwave- 9 based room-scale human interaction perception for hu- manoid robots.arXiv preprint arXiv:2601.07454, 2026. 1

    arXiv 2026

  25. [25]

    Tran- shupr: Cross-view fusion transformer for human pose esti- mation using mmwave radar

    Niraj Prakash Kini, Ruey-Horng Shiue, Ryan Chandra, Wen- Hsiao Peng, Ching-Wen Ma, and Jenq-Neng Hwang. Tran- shupr: Cross-view fusion transformer for human pose esti- mation using mmwave radar. InBMVC, 2024. 2

    2024

  26. [26]

    mil- limamba: Specular-aware human pose estimation via dual mmwave radar with multi-frame mamba fusion

    Niraj Prakash Kini, Shiau-Rung Tsai, Guan-Hsun Lin, Wen- Hsiao Peng, Ching-Wen Ma, and Jenq-Neng Hwang. mil- limamba: Specular-aware human pose estimation via dual mmwave radar with multi-frame mamba fusion. InProceed- ings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pages 1481–1490, 2026

    2026

  27. [27]

    Twr- pose3d: Through-wall radar based 3d human pose estima- tion with dual-stage temporal and view fusion

    Ailing Kong, Hengyu Gu, Xueya Yan, and Jian Guo. Twr- pose3d: Through-wall radar based 3d human pose estima- tion with dual-stage temporal and view fusion. In2026 38th Chinese Control and Decision Conference (CCDC), pages 4617–4622. IEEE, 2026. 2

    2026

  28. [28]

    m3track: mmwave-based multi-user 3d posture tracking

    Hao Kong, Xiangyu Xu, Jiadi Yu, Qilin Chen, Chenguang Ma, Yingying Chen, Yi-Chao Chen, and Linghe Kong. m3track: mmwave-based multi-user 3d posture tracking. In Proceedings of the 20th Annual International Conference on Mobile Systems, Applications and Services, pages 491–503,

  29. [29]

    Stacked cross attention for image-text matching

    Kuang-Huei Lee, Xi Chen, Gang Hua, Houdong Hu, and Xi- aodong He. Stacked cross attention for image-text matching. InProceedings of the European conference on computer vi- sion (ECCV), pages 201–216, 2018. 6

    2018

  30. [30]

    Hupr: A benchmark for human pose estimation using millimeter wave radar

    Shih-Po Lee, Niraj Prakash Kini, Wen-Hsiao Peng, Ching- Wen Ma, and Jenq-Neng Hwang. Hupr: A benchmark for human pose estimation using millimeter wave radar. InPro- ceedings of the IEEE/CVF Winter Conference on Applica- tions of Computer Vision, pages 5715–5724, 2023. 2

    2023

  31. [31]

    Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. InIn- ternational conference on machine learning, pages 19730– 19742. PMLR, 2023. 2

    2023

  32. [32]

    Wicompass: Oracle-driven data scal- ing for mmwave human pose estimation.arXiv preprint arXiv:2602.18726, 2026

    Bo Liang, Chen Gong, Haobo Wang, Qirui Liu, Rungui Zhou, Fengzhi Shao, Yubo Wang, Wei Gao, Kaichen Zhou, Guolong Cui, et al. Wicompass: Oracle-driven data scal- ing for mmwave human pose estimation.arXiv preprint arXiv:2602.18726, 2026. 1, 2, 5, 8

    arXiv 2026

  33. [33]

    Masked dis- crimination for self-supervised learning on point clouds

    Haotian Liu, Mu Cai, and Yong Jae Lee. Masked dis- crimination for self-supervised learning on point clouds. In European Conference on Computer Vision, pages 657–675. Springer, 2022. 2

    2022

  34. [34]

    Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning.Advances in neural information processing systems, 36:34892–34916, 2023. 2

    2023

  35. [35]

    Smpl: A skinned multi- person linear model

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J Black. Smpl: A skinned multi- person linear model. InSeminal Graphics Papers: Pushing the Boundaries, Volume 2, pages 851–866. 2023. 2

    2023

  36. [36]

    Amass: Archive of motion capture as surface shapes

    Naureen Mahmood, Nima Ghorbani, Nikolaus F Troje, Ger- ard Pons-Moll, and Michael J Black. Amass: Archive of motion capture as surface shapes. InProceedings of the IEEE/CVF international conference on computer vision, pages 5442–5451, 2019. 5

    2019

  37. [37]

    Masked autoencoders for 3d point cloud self- supervised learning.World Scientific Annual Review of Arti- ficial Intelligence, 1:2440001, 2023

    Yatian Pang, Eng Hock Francis Tay, Li Yuan, and Zhenghua Chen. Masked autoencoders for 3d point cloud self- supervised learning.World Scientific Annual Review of Arti- ficial Intelligence, 1:2440001, 2023. 2, 3

    2023

  38. [38]

    Expanding mmwave datasets for hu- man pose estimation with unlabeled data and lidar datasets

    Zhuoxuan Peng, Boan Zhu, Xingjian Zhang, Wenying Li, and S-H Gary Chan. Expanding mmwave datasets for hu- man pose estimation with unlabeled data and lidar datasets. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 21221–21230, 2026. 2, 8

    2026

  39. [39]

    Correlation coefficients: appropriate use and interpretation

    Patrick Schober, Christa Boer, and Lothar A Schwarte. Correlation coefficients: appropriate use and interpretation. Anesthesia & analgesia, 126(5):1763–1768, 2018. 6

    2018

  40. [40]

    mm-pose: Real-time human skeletal posture estimation using mmwave radars and cnns.IEEE sensors journal, 20 (17):10032–10044, 2020

    Arindam Sengupta, Feng Jin, Renyuan Zhang, and Siyang Cao. mm-pose: Real-time human skeletal posture estimation using mmwave radars and cnns.IEEE sensors journal, 20 (17):10032–10044, 2020. 1, 2

    2020

  41. [41]

    The proof and measurement of associa- tion between two things

    Charles Spearman. The proof and measurement of associa- tion between two things. 1961. 6

    1961

  42. [42]

    mmwave- flow: Unified enhancement and generation of mmwave hu- man point clouds

    Chang Su, Beihong Jin, Qiwen Shi, and Zhi Wang. mmwave- flow: Unified enhancement and generation of mmwave hu- man point clouds. InProceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 31366–31376, 2026. 2

    2026

  43. [43]

    Poseg- raphnet: Pose prior and graph structure for 3d human pose estimation using mmwave radar.Measurement, page 118851, 2025

    Yuanzhi Su, Huiying Cynthia Hou, and Chun Zhao. Poseg- raphnet: Pose prior and graph structure for 3d human pose estimation using mmwave radar.Measurement, page 118851, 2025. 2

    2025

  44. [44]

    Neural discrete representation learning.Advances in neural information pro- cessing systems, 30, 2017

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning.Advances in neural information pro- cessing systems, 30, 2017. 2, 3

    2017

  45. [45]

    mmjoints: Expanding joint representations beyond (x, y, z) in mmwave-based 3d pose estimation.arXiv preprint arXiv:2510.08970, 2025

    Zhenyu Wang, Mahathir Monjur, and Shahriar Nirjon. mmjoints: Expanding joint representations beyond (x, y, z) in mmwave-based 3d pose estimation.arXiv preprint arXiv:2510.08970, 2025. 1, 2

    arXiv 2025

  46. [46]

    Maepose: Self-supervised spa- tiotemporal learning for human pose estimation on mmwave video.arXiv preprint arXiv:2605.00242, 2026

    Xijia Wei, Yuan Fang, Kevin Chetty, Youngjun Cho, and Nadia Bianchi-Berthouze. Maepose: Self-supervised spa- tiotemporal learning for human pose estimation on mmwave video.arXiv preprint arXiv:2605.00242, 2026. 2

    Pith/arXiv arXiv 2026

  47. [47]

    mmhpe: Robust multiscale 3-d human pose estimation using a single mmwave radar.IEEE Internet of Things Journal, 12 (1):1032–1046, 2024

    Yingxiao Wu, Zhongmin Jiang, Haocheng Ni, Changlin Mao, Zhiyuan Zhou, Wenxiang Wang, and Jianping Han. mmhpe: Robust multiscale 3-d human pose estimation using a single mmwave radar.IEEE Internet of Things Journal, 12 (1):1032–1046, 2024

    2024

  48. [48]

    mmpoint: Dense human point cloud generation from mmwave

    Qian Xie, Qianyi Deng, Ta Ying Cheng, Peijun Zhao, Amir Patel, Niki Trigoni, and Andrew Markham. mmpoint: Dense human point cloud generation from mmwave. InBMVC, pages 194–196, 2023. 2

    2023

  49. [49]

    mmmesh: To- wards 3d real-time dynamic human mesh construction using millimeter-wave

    Hongfei Xue, Yan Ju, Chenglin Miao, Yijiang Wang, Shiyang Wang, Aidong Zhang, and Lu Su. mmmesh: To- wards 3d real-time dynamic human mesh construction using millimeter-wave. InProceedings of the 19th annual interna- tional conference on mobile systems, applications, and ser- vices, pages 269–282, 2021. 1, 2, 5

    2021

  50. [50]

    M4esh: mmwave-based 3d human mesh construction for multiple subjects

    Hongfei Xue, Qiming Cao, Yan Ju, Haochen Hu, Haoyu Wang, Aidong Zhang, and Lu Su. M4esh: mmwave-based 3d human mesh construction for multiple subjects. InPro- 10 ceedings of the 20th ACM Conference on Embedded Net- worked Sensor Systems, pages 391–406, 2022. 2

    2022

  51. [51]

    Towards generalized mmwave-based human pose estimation through signal aug- mentation

    Hongfei Xue, Qiming Cao, Chenglin Miao, Yan Ju, Haochen Hu, Aidong Zhang, and Lu Su. Towards generalized mmwave-based human pose estimation through signal aug- mentation. InProceedings of the 29th Annual International Conference on Mobile Computing and Networking, pages 1– 15, 2023. 1

    2023

  52. [52]

    Hongliu Yang, Zizhou Fan, Yueyang Wang, Jie Xiong, Duo Zhang, Xusheng Zhang, Junzhe Wang, Fusang Zhang, and Daqing Zhang. Bridging the resolution gap: Cost-effective human point cloud generation via low-bandwidth mmwave radar.Proceedings of the ACM on Interactive, Mobile, Wear- able and Ubiquitous Technologies, 10(2):1–26, 2026. 2

    2026

  53. [53]

    Mm-fi: Multi-modal non-intrusive 4d human dataset for versatile wireless sensing.Advances in Neural Information Processing Systems, 36:18756–18768, 2023

    Jianfei Yang, He Huang, Yunjiao Zhou, Xinyan Chen, Yue- cong Xu, Shenghai Yuan, Han Zou, Chris Xiaoxuan Lu, and Lihua Xie. Mm-fi: Multi-modal non-intrusive 4d human dataset for versatile wireless sensing.Advances in Neural Information Processing Systems, 36:18756–18768, 2023. 2

    2023

  54. [54]

    Point-bert: Pre-training 3d point cloud transformers with masked point modeling

    Xumin Yu, Lulu Tang, Yongming Rao, Tiejun Huang, Jie Zhou, and Jiwen Lu. Point-bert: Pre-training 3d point cloud transformers with masked point modeling. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19313–19322, 2022. 2

    2022

  55. [55]

    Duo Zhang, Xusheng Zhang, Shengjie Li, Yaxiong Xie, Yang Li, Xuanzhi Wang, and Daqing Zhang. Lt-fall: The design and implementation of a life-threatening fall detec- tion and alarming system.Proceedings of the ACM on In- teractive, Mobile, Wearable and Ubiquitous Technologies, 7 (1):1–24, 2023. 1

    2023

  56. [56]

    Mupose: Breaking the scalability barrier of mmwave multi-user pose estimation in the wild

    Duo Zhang, Zhehui Yin, Xusheng Zhang, Junzhe Wang, Hongliu Yang, Zhiyun Yao, Zizhou Fan, Wenwei Li, and Daqing Zhang. Mupose: Breaking the scalability barrier of mmwave multi-user pose estimation in the wild. InProceed- ings of the 24th Annual International Conference on Mobile Systems, Applications and Services, pages 622–636, 2026. 8

    2026

  57. [57]

    Umimo: Universal unsupervised learning for mmwave radar sensing with mimo array synthesis.IEEE Transactions on Mobile Computing, 2025

    Haoyu Zhang, Dongheng Zhang, Ruiyuan Song, Zhi Wu, Jinbo Chen, Liang Fang, Zhi Lu, Yang Hu, Hui Lin, and Yan Chen. Umimo: Universal unsupervised learning for mmwave radar sensing with mimo array synthesis.IEEE Transactions on Mobile Computing, 2025. 2 11

    2025

This paper was first reviewed by deepseek-v4-flash on August 1, 2026.