Pith. sign in

REVIEW 3 major objections 5 minor 54 references

SImpHAR: Advancing impedance-based human activity recognition using 3D simulation and text-to-motion models

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper argues that bio-impedance signals can be simulated from 3D motion and text and used to pretrain a HAR model that beats prior methods with little real labeled data.

desk verdict A genuinely new impedance-simulation pipeline with an overclaimed abstract and a possible calibration-leakage problem the paper never rules out. read the letter →

arxiv 2507.06405 v1 pith:BAKTXDSS submitted 2025-07-08 cs.CV cs.AI

Lala Shakti Swarup Ray , Mengxi Liu , Deepika Gurung , Bo Zhou , Sungho Suh , Paul Lukowicz This is my paper
classification cs.CVcs.AI
keywords bio-impedancesensinghumanactivityrecognitionwearablesensorssignalsimulationtext-to-motiongenerationcontrastivepretrainingdigitaltwindataaugmentation
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

Bio-impedance sensing can track fine-grained motion, but labeled impedance data are scarce and expensive to collect. SImpHAR proposes to generate synthetic impedance signals from 3D human meshes and text descriptions: geodesic shortest paths between electrodes, smoothed by soft-body deformation, stand in for the conductive path, and a neural module calibrates the simulation to real users. It then pretrains a contrastive encoder on synthetic impedance-text pairs and fine-tunes on small real labeled sets. On the new ImpAct dataset and the iMove and iEat benchmarks, the method reports consistent gains over prior work, up to 22.3% in accuracy and 21.8% in macro F1. If these results hold, simulation-driven augmentation could mitigate the data scarcity that has kept impedance-based HAR underused.

What carries the argument

The load-bearing mechanism is Pose2Imp, which converts a sequence of 3D body meshes into a time series of geodesic shortest paths between wrist electrode positions using Dijkstra's algorithm with Euclidean edge weights, deforms a soft body along those paths, and passes the result through a dual-LSTM encoder with a convolutional decoder that is trained with mean squared error on paired real impedance and pose data to personalize the signal. Text2Imp extends this by feeding activity descriptions to a diffusion-based text-to-motion generative model to produce pose sequences, which are then run through Pose2Imp. The training side is SImpHARNet, which uses a contrastive loss to align an impedance encoder with language-model text embeddings of synthetic impedance-text pairs, then fine-tunes the encoder on real labeled data with a freeze-then-unfreeze schedule to avoid catastrophic forgetting. Together these parts convert language and geometry into usable synthetic impedance training data without requiring label-aligned synthetic and real categories.

What would settle it

Compute the R2 of Pose2Imp's neural grounding on a held-out participant whose paired impedance-pose data were never used for calibration; if it falls to the 0.414 pose-only baseline rather than near the reported 0.843, the simulation is memorizing its calibration set rather than modeling impedance.

Watch

Extended reading notes

Core claim

The paper's central claim is that the impedance dynamics useful for activity recognition can be approximated by the changing length of the shortest conductive path between two electrodes on a 3D body surface, and that this geometric approximation, after neural calibration on paired real data, is realistic enough to support contrastive pretraining of a HAR encoder. From that premise, the paper claims that a two-stage pipeline—contrastive alignment of synthetic impedance with text embeddings, then fine-tuning on limited real data—consistently improves accuracy and macro F1 over state-of-the-art baselines on three impedance-based HAR datasets, with the largest gains in low-data regimes. The paper also introduces ImpAct, a dataset of synchronized impedance, video, and 3D pose across ten everyday activities, as a public benchmark for this modality.

Load-bearing premise

The central claim rests on the assumption that a geodesic shortest path between electrodes, deformed by soft-body physics and calibrated with paired real impedance-pose data, captures the impedance variation that distinguishes activities; the paper demonstrates this only for the participants and activities used in that calibration, not for unseen users or activities.

Editorial extensions

If this is right

  • SImpHARNet's fine-tuned model reaches 0.775 accuracy and 0.767 macro F1 on ImpAct, versus 0.675 and 0.648 for the strongest mixed-data baseline, so synthetic pretraining plus fine-tuning can beat naively mixing real and synthetic data.
  • On iEat, late-learning fine-tuning reaches 0.844 accuracy and 0.832 macro F1 despite the fact that simply adding synthetic SImp data to real data hurt performance (F1 dropped from 0.718 to 0.651), showing the decoupled design avoids negative transfer.
  • The method outperforms no-pretrain baselines with only 60–80% of the training data on iEat, indicating data-scarce regimes benefit most.
  • Because the pipeline generates impedance signals from text prompts or 3D meshes, it offers a way to synthesize training data for new activities without new hardware collection, limited by the motion diversity of the text-to-motion model.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the geodesic shortest-path proxy generalizes beyond upper-body wrist-to-wrist setups, the same simulator could generate training signals for chest, thigh, or other electrode placements from the same meshes, but this is untested in the paper.
  • The per-user neural grounding calibration means the simulator is best viewed as a cohort-conditional augmenter rather than a zero-calibration physics model; conditioning on user metadata such as height, weight, or body type is a natural testable extension the authors themselves flag.
  • A direct retrieval test—matching real impedance windows to their text descriptions in the pretrained shared space—would quantify whether the semantic alignment implied by the visualization of embedding clusters actually carries class-discriminative information.
  • The same contrastive design could be applied with other auxiliary modalities, such as IMU data or video, as anchors instead of text, potentially widening the use of synthetic data beyond impedance.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes SImpHAR, a framework that synthesizes bio-impedance signals from 3D body meshes and text descriptions via geodesic shortest-path estimation, soft-body deformation, and a learned neural grounding module (Pose2Imp), plus a text-to-motion stage (Text2Imp). It also introduces SImpHARNet, a two-stage contrastive-pretraining-plus-fine-tuning model, and presents ImpAct, a new multimodal impedance dataset. The authors claim consistent improvements over state-of-the-art HAR methods on ImpAct, iMove, and iEat, with accuracy and macro-F1 gains up to 22.3% and 21.8%, respectively, and emphasize benefits in low-data regimes.

Significance. If the central claims hold, this would be a valuable first step toward simulation-based data augmentation for bio-impedance HAR, a modality that currently lacks large labeled datasets. The paper's strengths include a concrete geometric simulation idea (geodesic shortest paths plus soft-body deformation), a two-stage learning scheme that avoids naive synthetic/real mixing, and an ablation study on training-set size and embedding dimension. However, the key claim of 'consistent improvements' is contradicted by the paper's own Table 1 on the iMove benchmark, and the simulation's neural grounding module depends on paired real impedance-pose data, creating a potential leakage risk in the leave-one-user-out evaluation. Because these issues directly affect the validity of the headline result, the work needs substantial clarification and re-evaluation before its contribution can be assessed.

major comments (3)
  1. [Abstract, §4.3, Table 1] The claim of 'consistent improvements over state-of-the-art methods' is false as written. In Table 1, on iMove, SImpHARNet (fine-tuned) achieves 0.763 accuracy / 0.752 macro F1, while TCNNet achieves 0.778 / 0.771 and SR achieves 0.770 / 0.764. The claimed 22.3% and 21.8% gains appear to correspond to relative improvements over TCNNet on iEat only, but the abstract does not specify this baseline or restrict the claim to a subset of datasets. The authors should either revise the claim to be dataset-specific and baseline-specific or present evidence of improvement on iMove under a clearly defined comparison protocol.
  2. [§3.1, §4.2, §5] The evaluation protocol for the user-specific grounding module is under-specified and creates a leakage risk. Section 3.1 trains Pose2Imp's neural mapping with MSE on paired real impedance-pose data, and Section 5 admits that the module 'requires subject-specific impedance-pose data for calibration.' The paper uses leave-one-user-out evaluation (§4.2) but never states whether the grounding module for the held-out user is trained only on the nine training users or on all ten users including the held-out user. If the held-out user's real impedance-pose pairs are used to generate SImp signals for that user's fold, then SImpHARNet's contrastive pretraining has already seen the target user's impedance distribution, and the reported low-data gains could be due to leakage rather than to the simulation pipeline. The authors must specify the exact calibration data used in each fold and provide a leak-free ablation in which the held-out user's real impedance data are never used in Pose2Imp.
  3. [§3.1, Table 2] Table 2 reports R² values for predicting real impedance from 3D pose and simulated distance, but it is not clear whether these numbers are evaluated under the same leave-one-user-out protocol used for the downstream HAR experiments. If the R² values are computed on users seen during training, they do not support the claim that the simulator transfers to unseen users. The authors should report R² under exactly the same LOUO protocol, and ideally also show the downstream HAR result when synthetic data for a held-out user are generated without any calibration on that user. This is load-bearing because the paper's central claim is that simulation, not calibration on the target user, alleviates data scarcity.
minor comments (5)
  1. [§4.1] The dataset description says each participant 'performed a total of eight activities: six upper-body fitness exercises and three daily tasks,' but six plus three is nine; later the introduction says 'ten everyday activities.' The count should be corrected.
  2. [§4.3, Table 2] The abbreviation STN is used without definition, and the table header says 'Performance of STN' while the text refers to the neural mapping module. Please define STN explicitly (or replace it with the module's actual name) and render R² consistently.
  3. [§3.3] The InfoNCE loss in Eq. (1) is written with query/key notation but omits the temperature parameter that is standard in contrastive learning formulations; please state the temperature value used in the experiments.
  4. [§4.2] The synthetic data generation is described only qualitatively: the paper does not report the number of SImp windows per class, the number of PriorMDM samples used for iMove and iEat, or the exact text prompts for each target class. These details are needed to reproduce the pretraining stage.
  5. [§4.1, §6] The ImpAct dataset is introduced as a contribution but is not released, and the paper does not state a plan for releasing code or data. Given that the central claim depends on a new simulation pipeline and a new dataset, this limits reproducibility and verification.

Circularity Check

0 steps flagged · score 2.0 of 10

No load-bearing circularity; the simulation includes a data-driven calibration module, but downstream HAR results are not forced by construction.

full rationale

The paper's derivation chain does not contain a self-referential loop that makes any predicted quantity equivalent to its inputs by construction. The most circularity-adjacent element is Pose2Imp's neural grounding module (Section 3.1), which is trained on paired real-world impedance and pose data with an MSE loss. This means the 'simulated' SImp signals are not generated purely by the physics-based shortest-path and soft-body components; they are outputs of a learned mapping fit to real measurements. Section 5 explicitly acknowledges this: 'the neural grounding model requires subject-specific impedance-pose data for calibration, limiting the framework's ability to generalize across users without retraining.' That is an honest limitation and weakens the claim of a fully first-principles simulation pipeline, but it is not circular: the downstream HAR classifier must still learn a mapping from impedance signals to activity labels, and the paper's ablations (e.g., Figure 7) show that fine-tuning strategy and data fraction materially affect performance. The leave-one-user-out evaluation could in principle leak target-user information if the grounding module is trained on the held-out user's real impedance-pose data, but the paper never states that this occurs, and the text does not exhibit the specific reduction required to call it circular. Separately, the abstract's claim of 'consistent improvements over state-of-the-art methods' is contradicted by Table 1 on the iMove dataset, where TCNNet achieves 0.778 accuracy / 0.771 macro F1 versus SImpHARNet's 0.763 / 0.752; this is a factual inconsistency rather than a circularity. Given the explicit calibration requirement is disclosed and the central architecture has independent content, the circularity burden is low, though the simulation's novelty is partly empirical rather than purely physics-based.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a learned grounding module fitted to real impedance data, hand-chosen proxy prompts, and unstated soft-body parameters. The only genuinely new physical element is the use of geodesic shortest paths as an impedance proxy, which is itself an assumption. No new particles, forces, or conserved quantities are introduced.

free parameters (4)
  • Pose2Imp neural mapping weights = trained via MSE on paired real impedance-pose data
    SImpHAR's simulated impedance is produced by a neural decoder learned from real measurements; this is a fitted component, not a physical first-principles prediction (Section 3.1).
  • Soft-body deformation parameters = not specified
    The paper states the soft body is 'deformed using internal strain dynamics' but gives no stiffness, damping, or solver parameters; these hand-chosen values affect the simulated path smoothing (Section 3.1).
  • Manually crafted text prompts per class = 3 prompts per activity class
    The contrastive pretraining pairs depend on hand-written proxy descriptions such as 'touching both ears' for 'wearing VR headset'; these are chosen by the authors and influence the semantic alignment (Section 3.2).
  • Architecture hyperparameters = window 60, embed size 256, lr 0.001, patience 10
    These choices were tuned on validation and affect the reported F1 curves in Figures 6 and 7; they are not derived from theory.
assumptions (5)
  • domain assumption Time-varying bio-impedance between two wrist electrodes is approximated by the geodesic shortest path between electrode positions on a 3D body mesh, plus soft-body smoothing.
    Section 3.1 motivates the whole simulation; if this proxy is wrong, SImp signals do not represent real impedance dynamics.
  • domain assumption The neural grounding module trained on paired real impedance-pose data can produce synthetic signals that improve a classifier evaluated on held-out users.
    Section 3.1 and the Limitations admit the grounding requires subject-specific calibration data; the cross-user transfer of this fitted module is assumed rather than demonstrated.
  • domain assumption Text-to-motion model PriorMDM, prompted with proxy descriptions, generates motions whose impedance dynamics are structurally similar enough to the target activities for contrastive pretraining.
    Section 3.2 uses proxy prompts because target motions like 'wearing VR glasses' are absent from HumanML3D/KIT-ML; similarity is only shown via t-SNE of text embeddings, not via motion or impedance similarity.
  • domain assumption OSX whole-body mesh recovery and joint2SMPL provide sufficiently accurate geometry for shortest-path computation.
    The simulation pipeline relies on these off-the-shelf pose and mesh estimators (Section 3.1); their errors propagate into the synthetic impedance signals.
  • domain assumption Leave-one-user-out evaluation with a random 9:1 train/validation split does not leak information through the Pose2Imp grounding model.
    The paper does not specify whether the grounding model is fitted on training users only; if test-user impedance pairs are used to calibrate the simulator, the evaluation is optimistic.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SImpHAR: Advancing impedance-based human activity recognition using 3D simulation and text-to-motion models." pith.science (2026). https://pith.science/paper/BAKTXDSS

@misc{pith2026250706405,
  author       = {Pith},
  title        = {Pith review of: SImpHAR: Advancing impedance-based human activity recognition using 3D simulation and text-to-motion models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BAKTXDSS}},
  note         = {Machine review of arXiv:2507.06405}
}
read the original abstract

Human Activity Recognition (HAR) with wearable sensors is essential for applications in healthcare, fitness, and human-computer interaction. Bio-impedance sensing offers unique advantages for fine-grained motion capture but remains underutilized due to the scarcity of labeled data. We introduce SImpHAR, a novel framework addressing this limitation through two core contributions. First, we propose a simulation pipeline that generates realistic bio-impedance signals from 3D human meshes using shortest-path estimation, soft-body physics, and text-to-motion generation serving as a digital twin for data augmentation. Second, we design a two-stage training strategy with decoupled approach that enables broader activity coverage without requiring label-aligned synthetic data. We evaluate SImpHAR on our collected ImpAct dataset and two public benchmarks, showing consistent improvements over state-of-the-art methods, with gains of up to 22.3% and 21.8%, in terms of accuracy and macro F1 score, respectively. Our results highlight the promise of simulation-driven augmentation and modular training for impedance-based HAR.

Figures

Figures reproduced from arXiv: 2507.06405 by the authors.

Figure 1
Figure 1. Synthesized impedance signal (SImp) vs. real [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Direct correlation between impedance magnitude, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 4
Figure 4. SImpHARNet consists of 2 steps (1) pretraining with [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Text embeddings of target classes (orange) and pre [PITH_FULL_IMAGE:figures/full_fig_p004_5.png]
Figure 6
Figure 6. Figure 6: SImpHARNet trained with varying embedding size [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Change in F1 score for classification models trained [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

54 extracted references · 40 canonical work pages

  1. [1]

    Sizhen Bian, Mengxi Liu, Bo Zhou, and Paul Lukowicz. 2022. The state-of-the-art sensing techniques in human activity recognition: A survey.Sensors22, 12 (2022), 4596

    work page 2022

  2. [2]

    Sizhen Bian, Vitor F Rey, Peter Hevesi, and Paul Lukowicz. 2019. Passive capacitive based approach for full body gym workout recognition and counting. In2019 IEEE International Conference on Pervasive Computing and Communications (PerCom. IEEE, 1–10

    work page 2019

  3. [3]

    Haofeng Chen, Gang Ma, Peng Wang, and Xiaojie Wang. 2021. A bio-impedance analysis method based on human hand anatomy for hand gesture recognition. IEEE Transactions on Instrumentation and Measurement70 (2021), 1–10

    work page 2021

  4. [4]

    Jingyuan Cheng, Mathias Sundholm, Bo Zhou, Matthias Kreil, and Paul Lukow- icz. 2014. Recognizing Subtle User Activities and Person Identity with Cheap Resistive Pressure Sensing Carpet. In2014 International Conference on Intelligent Environments. 148–153. doi:10.1109/IE.2014.29

    work page doi:10.1109/ie.2014.29 2014

  5. [5]

    Mathias Ciliberto, Vitor Fortes Rey, Alberto Calatroni, Paul Lukowicz, and Daniel Roggen. 2021. Opportunity++: A multimodal dataset for video-and wearable, ob- ject and ambient sensors-based human activity recognition.Frontiers in Computer Science3 (2021), 792065

    work page 2021

  6. [6]

    L Minh Dang, Kyungbok Min, Hanxiang Wang, Md Jalil Piran, Cheol Hee Lee, and Hyeonjoon Moon. 2020. Sensor-based and vision-based human activity recognition: A comprehensive survey.Pattern Recognition108 (2020), 107561

    work page 2020

  7. [7]

    Smith, Hao Xue, and Flora D

    Shohreh Deldari, Daniel V. Smith, Hao Xue, and Flora D. Salim. 2021. Time Series Change Point Detection with Self-Supervised Contrastive Predictive Coding. In Proceedings of the Web Conference 2021 (WWW ’21). ACM. doi:10.1145/3442381. 3449903

    doi:10.1145/3442381 2021

  8. [8]

    Vitor Fortes Rey, Kamalveer Kaur Garewal, and Paul Lukowicz. 2021. Translating videos into synthetic training data for wearable sensor-based activity recognition systems using residual deep convolutional networks.Applied Sciences11, 7 (2021), 3094

    work page 2021

Show all 54 references
  1. [9]

    Vitor Fortes Rey, Sungho Suh, and Paul Lukowicz. 2022. Learning from the best: contrastive representations learning across sensor locations for wearable activity recognition. InProceedings of the 2022 ACM International Symposium on Wearable Computers. 28–32

    2022

  2. [10]

    Rafael Gonzalez-Landaeta, Oscar Casas, and Ramon Pallas-Areny. 2008. Heart rate detection from plantar bioimpedance measurements.IEEE transactions on Biomedical Engineering55, 3 (2008), 1163–1167

    2008

  3. [11]

    Chuan Guo, Shihao Zou, Xinxin Zuo, Sen Wang, Wei Ji, Xingyu Li, and Li Cheng

  4. [12]

    Yujiao Hao, Xijian Lou, Boyu Wang, and Rong Zheng. 2022. Cromosim: A deep learning-based cross-modality inertial measurement simulator.IEEE Transactions on Mobile Computing(2022)

    2022

  5. [13]

    Mir Kanon Ara Jannat, Md Shafiqul Islam, Sung-Hyun Yang, and Hui Liu. 2023. Efficient Wi-Fi-based human activity recognition using adaptive antenna elimi- nation.IEEE Access(2023)

    2023

  6. [14]

    Hyeokhyen Kwon, Catherine Tong, Harish Haresamudram, Yan Gao, Gregory D Abowd, Nicholas D Lane, and Thomas Ploetz. 2020. Imutube: Automatic extrac- tion of virtual on-body accelerometry from video for human activity recognition. Proceedings of the ACM on Interactive, Mobile, W...

    2020

  7. [15]

    Paula Lago, Moe Matsuki, Kohei Adachi, and Sozo Inoue. 2021. Using addi- tional training sensors to improve single-sensor complex activity recognition. InProceedings of the 2021 ACM International Symposium on Wearable Computers (Virtual, USA)(ISWC ’21). Association for Computi...

    2021

  8. [16]

    Paula Lago, Tsuyoshi Okita, Shingo Takeda, and Sozo Inoue. 2018. Improving Sensor-based Activity Recognition Using Motion Capture as Additional Infor- mation. InProceedings of the 2018 ACM International Joint Conference and 2018 International Symposium on Pervasive and Ubiquit...

    2018

  9. [17]

    Woojae Lee and SeongHwan Cho. 2015. Integrated all electrical pulse wave velocity and respiration sensors using bio-impedance.IEEE Journal of Solid-State Circuits50, 3 (2015), 776–785

    2015

  10. [18]

    Zikang Leng, Amitrajit Bhattacharjee, Hrudhai Rajasekhar, Lizhe Zhang, Eliza- beth Bruda, Hyeokhyen Kwon, and Thomas Plötz. 2024. IMUGPT 2.0: Language- Based Cross Modality Transfer for Sensor-Based Human Activity Recognition. arXiv preprint arXiv:2402.01049(2024)

    2024 arXiv

  11. [19]

    Jing Lin, Ailing Zeng, Haoqian Wang, Lei Zhang, and Yu Li. 2023. One-stage 3d whole-body mesh recovery with component aware transformer. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 21159– 21168

    2023

  12. [20]

    Mengxi Liu, Hymalai Bello, Bo Zhou, Paul Lukowicz, and Jakob Karolus. 2024. iFace: Hand-Over-Face Gesture Recognition Leveraging Impedance Sensing. In Proceedings of the Augmented Humans International Conference 2024. 131–137

    2024

  13. [21]

    Mengxi Liu, Vitor Fortes Rey, Yu Zhang, Lala Shakti Swarup Ray, Bo Zhou, and Paul Lukowicz. 2024. iMove: Exploring Bio-impedance Sensing for Fitness Activity Recognition.arXiv preprint arXiv:2402.09445(2024)

    2024 arXiv

  14. [22]

    Mengxi Liu, Yu Zhang, Bo Zhou, Sizhen Bian, Agnes Grünerbl, and Paul Lukowicz

  15. [23]

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

    2023

  16. [24]

    James R Matthie. 2008. Bioimpedance measurements of human body composition: critical analysis and outlook.Expert review of medical devices5, 2 (2008), 239–261

    2008

  17. [25]

    Seungwhan Moon, Andrea Madotto, Zhaojiang Lin, Alireza Dirafzoon, Aparajita Saraf, Amy Bearman, and Babak Damavandi. 2022. Imu2clip: Multimodal con- trastive learning for imu motion sensors from egocentric videos and text.arXiv preprint arXiv:2210.14395(2022)

    2022 arXiv

  18. [26]

    Yazeed K Musallam, Nasser I AlFassam, Ghulam Muhammad, Syed Umar Amin, Mansour Alsulaiman, Wadood Abdul, Hamdi Altaheri, Mohamed A Bencherif, and Mohammed Algabri. 2021. Electroencephalography-based motor imagery classification using temporal convolutional network fusion.Biome...

    2021

  19. [27]

    Duc-Anh Nguyen, Cuong Pham, and Nhien-An Le-Khac. 2023. Virtual Fusion with Contrastive Learning for Single Sensor-based Activity Recognition.arXiv preprint arXiv:2312.02185(2023)

    2023 arXiv

  20. [28]

    Matthias Plappert, Christian Mandery, and Tamim Asfour. 2016. The kit motion- language dataset.Big data4, 4 (2016), 236–252

    2016

  21. [29]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. 2021. Learning transferable visual models from natural language supervision. InInternational conference on machine learnin...

    2021

  22. [30]

    Lala Ray, Daniel Geißler, Bo Zhou, Paul Lukowicz, and Berit Greinke. 2024. Origami single-end capacitive sensing for continuous shape estimation of mor- phing structures.Scientific Reports14, 1 (2024), 17448

    2024

  23. [31]

    Lala Shakti Swarup Ray, Daniel Geißler, Mengxi Liu, Bo Zhou, Sungho Suh, and Paul Lukowicz. 2024. ALS-HAR: Harnessing Wearable Ambient Light Sensors to Enhance IMU-based HAR.arXiv preprint arXiv:2408.09527(2024)

    2024 arXiv

  24. [32]

    Lala Shakti Swarup Ray, Daniel Geißler, Bo Zhou, Paul Lukowicz, and Berit Greinke. 2023. Capafoldable: self-tracking foldable smart textiles with capacitive sensing.arXiv preprint arXiv:2307.05370(2023)

    2023 arXiv

  25. [33]

    Lala Shakti Swarup Ray, Vitor Fortes Rey, Bo Zhou, Sungho Suh, and Paul Lukow- icz. 2023. PressureTransferNet: Human Attribute Guided Dynamic Ground Pressure Profile Transfer using 3D simulated Pressure Maps.arXiv preprint arXiv:2308.00538(2023)

    2023 arXiv

  26. [34]

    Lala Shakti Swarup Ray, Bo Zhou, Sungho Suh, Lars Krupp, Vitor Fortes Rey, and Paul Lukowicz. 2024. Text me the data: Generating Ground Pressure Sequence from Textual Descriptions for HAR. In2024 IEEE International Conference on Per- vasive Computing and Communications Worksho...

    2024

  27. [35]

    Lala Shakti Swarup Ray, Bo Zhou, Sungho Suh, and Paul Lukowicz. 2023. Pressim: An end-to-end framework for dynamic ground pressure profile generation from monocular videos using physics-based 3d simulation. In2023 IEEE International Conference on Pervasive Computing and Commun...

    2023

  28. [36]

    Vitor Fortes Rey, Lala Shakti Swarup Ray, Xia Qingxin, Kaishun Wu, and Paul Lukowicz. 2024. Enhancing Inertial Hand based HAR through Joint Represen- tation of Language, Pose and Synthetic IMUs.arXiv preprint arXiv:2406.01316 (2024)

    2024 arXiv

  29. [37]

    Matthias Ring, Clemens Lohmueller, Manfred Rauh, Joachim Mester, and Bjo- ern M Eskofier. 2015. A temperature-based bioimpedance correction for water loss estimation during sports.IEEE journal of biomedical and health informatics 20, 6 (2015), 1477–1484

    2015

  30. [38]

    Kaan Sel, Deen Osman, and Roozbeh Jafari. 2021. Non-invasive cardiac and respiratory activity assessment from various human body locations using bioimpedance.IEEE open journal of engineering in medicine and biology2 (2021), 210–217

    2021

  31. [39]

    Yonatan Shafir, Guy Tevet, Roy Kapon, and Amit H Bermano. 2023. Human motion diffusion as a generative prior.arXiv preprint arXiv:2303.01418(2023)

    2023 arXiv

  32. [40]

    David Strömbäck, Sangxia Huang, and Valentin Radu. 2020. Mm-fit: Multimodal deep learning for automatic exercise logging across sensing devices.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies4, 4 (2020), 1–22

    2020

  33. [41]

    Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A Smith, Luke Zettlemoyer, and Tao Yu. 2022. One embedder, any task: Instruction-finetuned text embeddings.arXiv preprint arXiv:2212.09741 (2022)

    2022 arXiv

  34. [42]

    Sungho Suh, Vitor Fortes Rey, Sizhen Bian, Yu-Chi Huang, Jože M Rožanec, Hooman Tavakoli Ghinani, Bo Zhou, and Paul Lukowicz. 2023. Worker Activity Conference acronym ’XX, June 03–05, 2018, Woodstock, NY Ray et al. Recognition in Manufacturing Line Using Near-body Electric Fie...

    2023

  35. [43]

    Chi Ian Tang, Ignacio Perez-Pozuelo, Dimitris Spathis, and Cecilia Mascolo. 2020. Exploring contrastive learning in human activity recognition for healthcare. arXiv preprint arXiv:2011.11542(2020)

    2020 arXiv

  36. [44]

    Laurens Van der Maaten and Geoffrey Hinton. 2008. Visualizing data using t-SNE. Journal of machine learning research9, 11 (2008)

    2008

  37. [45]

    Huiyuan Yang, Han Yu, Kusha Sridhar, Thomas Vaessen, Inez Myin-Germeys, and Akane Sano. 2022. More to Less (M2L): Enhanced health recognition in the wild with reduced modality of wearable sensors. In2022 44th Annual International Conference of the IEEE Engineering in Medicine ...

    2022

  38. [46]

    Naoya Yoshimura, Jaime Morales, Takuya Maekawa, and Takahiro Hara. 2024. Openpack: A large-scale dataset for recognizing packaging works in iot-enabled logistic environments. In2024 IEEE International Conference on Pervasive Com- puting and Communications (PerCom). IEEE, 90–97

    2024

  39. [47]

    Alexander D Young, Martin J Ling, and Damal K Arvind. 2011. IMUSim: A simulation environment for inertial sensing algorithm design and evaluation. InProceedings of the 10th ACM/IEEE International Conference on Information Processing in Sensor Networks. IEEE, 199–210

    2011

  40. [48]

    Bo Zhou, Mathias Sundholm, Jingyuan Cheng, Heber Cruz, and Paul Lukowicz

  41. [49]

    Yexu Zhou, Haibin Zhao, Yiran Huang, Till Riedel, Michael Hefenbrock, and Michael Beigl. 2022. Tinyhar: A lightweight deep learning model designed for human activity recognition. InProceedings of the 2022 ACM International Symposium on Wearable Computers. 89–93

    2022

  42. [50]

    Henrik Zimmer, Marcel Campen, and Leif Kobbelt. 2013. Efficient computation of shortest path-concavity for 3D meshes. InProceedings of the IEEE Conference on Computer Vision and Pattern Recognition. 2155–2162

    2013

  43. [51]

    Parham Zolfaghari, Vitor Fortes Rey, Lala Ray, Hyun Kim, Sungho Suh, and Paul Lukowicz. 2024. Sensor Data Augmentation from Skeleton Pose Sequences for Improving Human Activity Recognition. In2024 International Conference on Activity and Behavior Computing (ABC). IEEE, 1–8

    2024

  44. [2016]

    In2016 IEEE International Conference on Pervasive Computing and Communications (PerCom)

    Never skip leg day: A novel wearable approach to monitoring gym leg exercises. In2016 IEEE International Conference on Pervasive Computing and Communications (PerCom). IEEE, 1–9

  45. [2022]

    InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition

    Generating diverse and natural 3d human motions from text. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 5152– 5161

  46. [2023]

    InAdjunct Proceedings of the 2023 ACM International Joint Conference on Pervasive and Ubiquitous Computing & the 2023 ACM International Symposium on Wearable Computing

    iEat: Human-food interaction with bio-impedance sensing. InAdjunct Proceedings of the 2023 ACM International Joint Conference on Pervasive and Ubiquitous Computing & the 2023 ACM International Symposium on Wearable Computing. 207–207

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.