Pith. sign in

REVIEW 4 major objections 5 minor 2 cited by

Scaling Human Activity Recognition: A Comparative Evaluation of Synthetic Data Generation and Augmentation Techniques

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Virtual IMU data generated from video or text consistently improves human activity recognition compared with real data alone or classical sensor-level augmentation, especially when real labeled data is scarce.

desk verdict Useful head-to-head benchmark, but the 'consistent' advantage claim collapses under volume mismatch and HAD-AW null results. read the letter →

arxiv 2506.07612 v2 pith:FUPBACZ5 submitted 2025-06-09 cs.CV

Zikang Leng , Archith Iyer , Thomas Plötz This is my paper · ORCID
classification cs.CV
keywords humanactivityrecognitionvirtualIMUdatacross-modalitytransferaugmentationIMUTubeIMUGPTsyntheticsensor
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

Human activity recognition from wearable sensors is limited by the cost of collecting labeled IMU data. This paper tries to establish that generating virtual IMU data from other modalities—video and text—is a more effective way to scale training data than traditional sensor-level augmentation. The authors directly compare the video-based IMUTube pipeline, the text-based IMUGPT pipeline, and classical perturbation methods on three benchmark HAR datasets with four models, building a 100-activity virtual dataset at 22 body locations. They report that virtual IMU data consistently improves macro F1 over real-only or augmented training, that the two virtual sources are complementary, and that the gains are largest when real data is limited to ten percent.

What carries the argument

The machinery that carries the argument is cross-modality transfer into virtual IMU signals. IMUTube starts from video: it filters low-quality clips, extracts 2D poses, lifts them to 3D, reconstructs global body motion, and simulates wearable sensor readings at 22 joints. IMUGPT starts from text: it uses a large language model to produce varied short descriptions of each activity, converts those descriptions into 3D motion with a text-to-motion transformer, and simulates the same sensor signals. The classical baseline perturbs real IMU windows by rotation about the z-axis, additive Gaussian noise, and sensor bias. The experimental comparison is built around selecting only virtual samples whose activity labels and sensor locations match the real datasets, and evaluating every configuration under full and ten percent real data.

What would settle it

Train the same four models on real augmented data and on virtual IMU data subsampled to exactly the same number of minutes, matching the augmented volumes of 85, 212, and 1,595 minutes reported in Table 1; the central claim is falsified if virtual data no longer outperforms augmentation under equal data volume.

Watch

Extended reading notes

Core claim

On its own terms, the claim is that incorporating virtual IMU data consistently enhances model performance compared to training on real data alone or with traditional data augmentation, particularly when real data is scarce. The paper constructs a 100-activity virtual IMU dataset from Kinetics-400 with simulated signals at 22 body locations, then adds activity-matched subsets of that data to three real HAR datasets. Across Random Forest, DeepConvLSTM, DeepConvLSTM with self-attention, and BiLSTM with attention, virtual data produces the largest macro F1 gains, and the two virtual sources together outperform either alone. The authors interpret this as showing that video-based generation supplies realistic motion while text-based generation supplies semantic diversity, and that classical augmentation of real signals cannot match the diversity contributed by cross-modality transfer.

Load-bearing premise

The comparison assumes that the amount of training data is not what drives the results, since the virtual-data training sets are two to twenty-five times larger than the augmented-real sets; if the volumes were equalized, the reported advantage of virtual data could shrink or disappear.

Editorial extensions

If this is right

  • If the central claim is correct, HAR practitioners with limited labeled data should prefer cross-modality virtual IMU data over perturbation-based augmentation when expanding training sets.
  • Video-generated virtual data should be chosen when realistic motion is the priority, while text-generated data should be added when diversity is needed; combining both gives the best average performance.
  • Virtual data is a viable partial remedy for class imbalance, since underrepresented activity classes show the largest per-class F1 improvements.
  • The benefit of virtual data grows as real data shrinks, so these methods are most valuable in low-data deployment scenarios.
  • Cost-conscious teams should start with the text-based pipeline, which generates ten seconds of virtual IMU data in about ten seconds, and add the video-based pipeline only when its roughly five minutes per ten seconds of output is acceptable.

Reading between the lines

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

  • The paper's own Table 1 shows the virtual-data training sets are two to twenty-five times larger than the augmented-real sets, so some of the reported advantage may be a data-volume effect rather than a property of the generation method; subsampling virtual data to match augmented volumes would test this.
  • The experiments cover only activities that overlap with Kinetics-400, so the results may not transfer to fine-grained or hand-centric activities that video and text-to-motion models represent poorly.
  • A natural extension not tested in the paper is using virtual IMU data for pretraining followed by fine-tuning on real data, which could outperform the joint-training configuration used here.
  • Classical augmentation and virtual data are not mutually exclusive; combining augmentation on real data with virtual data may yield additional gains, though the paper does not test this combination.
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

4 major / 5 minor

Summary. The paper compares three strategies for expanding training data in sensor-based human activity recognition (HAR): classical sensor-level augmentation, video-based virtual IMU generation (IMUTube), and text-based virtual IMU generation (IMUGPT). Using subsets of Kinetics-400, the authors construct a large virtual IMU dataset at 22 body locations and evaluate the strategies on UTD-MHAD, PAMAP2, and HAD-AW with four classifiers, under both full and 10% real-data conditions. The reported results claim that virtual IMU data consistently improves performance over real-only and augmented training, especially when real data is scarce, and that IMUTube and IMUGPT provide complementary benefits. The paper also offers practical guidance on method selection and reports computational costs.

Significance. If the central claim were established, the paper would provide useful practical guidance for HAR practitioners facing labeled-data scarcity: it is the first head-to-head comparison of video-based and text-based virtual IMU generation against classical augmentation, and the promised public release of a 100-activity, 22-location virtual IMU dataset would be a valuable community resource. The evaluation spans multiple datasets and models and includes per-class analysis and cost considerations, which are strengths. However, the main comparative claim is currently undermined by an uncontrolled confound: training-set sizes differ substantially across configurations, and the reported 'consistent' improvements are contradicted by several results in the paper's own tables. These issues are load-bearing because they affect causal attribution of the observed gains to the data generation method.

major comments (4)
  1. [§3.4.2, Table 1, §3.5] The central comparison is confounded by training-set volume. Table 1 shows that on UTD-MHAD, Real+IMUGPT uses 544 minutes and Real+Augmentation uses 85 minutes (a 6.4x difference); on PAMAP2, Real+IMUGPT uses 272 vs 212 minutes for augmentation; on HAD-AW, Real+IMUGPT uses 943 minutes vs 255 minutes for IMUTube. Since model performance generally improves with more training data, the reported advantage of virtual data over augmentation can be explained by sample quantity rather than by any property of the generation method. The paper never runs volume-matched controls or learning curves over synthetic data quantity, so the abstract's causal claim ('virtual IMU data significantly improves performance... compared to... traditional data augmentation') is not supported by the current experimental design.
  2. [Abstract, §5, Table 3] The claim that virtual IMU data 'consistently enhances' performance is internally contradicted by Table 3. On HAD-AW with full data, Real Only achieves the best or near-best macro F1 for every model: e.g., BiLSTM with Attention Real Only is 89.49, while Real+IMUGPT is 82.70 and Real+IMUTube is 81.19; Random Forest Real Only is 55.68 vs 53.19 for Real+IMUGPT. On UTD-MHAD, DeepConvLSTM and DeepConvLSTM with Self-Attention are best with Real+Augmentation (57.33 and 58.26), exceeding both virtual configurations. The conclusion should be revised to acknowledge that virtual data can hurt performance on some datasets/models and that the benefit is not consistent.
  3. [§3.4.1, §3.5, Tables 3–4] No statistical significance testing is reported across the three seeds or across cross-validation folds. Many differences are smaller than the reported standard deviations: for example, Table 3 UTD-MHAD BiLSTM Real Only 78.57±0.18 vs Real+Augmentation 78.82±0.78, and HAD-AW Random Forest Real Only 55.68±0.22 vs Real+Augmentation 55.34±0.19. The text uses terms like 'significantly improves' and 'substantial improvements' without paired tests, effect sizes, or per-fold results. The authors should add appropriate statistical tests or at least report individual fold/seed results.
  4. [§3.3, Table 2] The evaluation is restricted to a subset of activities from each dataset that overlap with Kinetics-400 (e.g., 5 of 18 PAMAP2 activities, 7 of 27 UTD-MHAD activities), which changes the classification task relative to the original benchmarks. This is not necessarily a flaw, but the paper should state explicitly that conclusions apply only to these activity subsets and should justify that the selected classes are representative enough to support the general guidance offered in the Discussion.
minor comments (5)
  1. [§3.3, §3.5] There are inconsistent labels for the HAD-AW dataset: 'HAD-A W' appears in Sections 3.3 and 3.5, while 'HAD-AW' is used elsewhere.
  2. [Table 4] The capitalization 'IMUGPT+IMUTUBE' appears in Table 4 while the text uses 'IMUGPT + IMUTube'; please make these consistent.
  3. [§4] There is a typo in the Data Availability paragraph: 'underscoresthe' should be 'underscores the'.
  4. [§3.2, Eq. (3)] The bias augmentation samples b from U(−0.1, 0.1), but the random seed is not specified. For reproducibility, please state whether augmentation is applied with a fixed seed and describe the random sampling procedure.
  5. [§1, Contributions] The paper states the virtual dataset 'will be made publicly available upon acceptance.' Please clarify the intended license and access mechanism, as this is relevant to the reproducibility claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the comparison is empirical, anchored on external benchmark datasets; self-citations to IMUTube and IMUGPT supply the pipelines under test but do not by construction force the reported performance differences.

full rationale

This paper reports a comparative empirical evaluation rather than a derivation or first-principles prediction. The central claim — that virtual IMU data from IMUTube/IMUGPT improves HAR performance relative to real-only or classically augmented training — is supported by experiments on three public benchmark datasets (UTD-MHAD, PAMAP2, HAD-AW) with four models, so the comparison is anchored outside the authors' own fitted values. The authors' prior work is cited as the source of the two virtual pipelines ([17-22]), but those citations are not used to forbid alternatives or to justify an equation whose output equals its input; the pipelines are independently published tools being compared. The main threat to the paper's causal claim is experimental rather than circular: training-set sizes are not matched across conditions (Table 1), and the 'consistent' improvement is not borne out in every HAD-AW full-data row (Table 3). However, data-volume imbalance and contradictory results are validity questions, not cases where a prediction reduces by construction to its inputs. No fitted parameter is renamed as a prediction, no uniqueness theorem is imported from the authors, and no ansatz is smuggled in via citation. Accordingly, no circular step is exhibited, and the circularity score is 0.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

No new physical entities are introduced. The free parameters are the augmentation strengths, which are hand-picked. The axioms are the validity of the virtual-data pipelines and the representativeness of the activity subset; the former is inherited from prior work and the latter is a paper-specific choice.

free parameters (3)
  • Rotation angle theta in augmentation = pi/6
    Chosen by hand in Section 3.2 without justification. Controls the strength of the rotation augmentation baseline.
  • Gaussian noise sigma in augmentation = 0.05
    Chosen by hand in Section 3.2. Affects the noise augmentation strength; stronger noise could change the comparison.
  • Bias magnitude in augmentation = U(-0.1, 0.1)
    Chosen by hand in Section 3.2. Adds a constant offset; the range is arbitrary and affects the baseline.
assumptions (3)
  • domain assumption Simulated IMU signals from IMUSim are a valid proxy for real wearable sensor data at the same body locations.
    The entire evaluation rests on the virtual data being informative for downstream real-data classification. This premise is inherited from the IMUGPT and IMUTube pipelines, invoked in Sections 3.1.1 and 3.1.2.
  • ad hoc to paper The 100 selected Kinetics-400 classes, filtered to overlap with each benchmark's activities, are representative of real HAR deployment conditions.
    The activity selection is specific to this paper (Section 3.1) and limits the conclusions to the remaining activities, which are subsets of each benchmark (Table 2).
  • domain assumption Text-to-motion model T2M-GPT and pose-estimation models ViTPose and MixSTE produce realistic 3D human motion for the target activities.
    These are pretrained models from prior work; the paper relies on their output without validating against real motion capture. Sections 3.1.1 and 3.1.2.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Scaling Human Activity Recognition: A Comparative Evaluation of Synthetic Data Generation and Augmentation Techniques." pith.science (2026). https://pith.science/paper/FUPBACZ5

@misc{pith2026250607612,
  author       = {Pith},
  title        = {Pith review of: Scaling Human Activity Recognition: A Comparative Evaluation of Synthetic Data Generation and Augmentation Techniques},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FUPBACZ5}},
  note         = {Machine review of arXiv:2506.07612}
}
read the original abstract

Human activity recognition (HAR) is often limited by the scarcity of labeled datasets due to the high cost and complexity of real-world data collection. To mitigate this, recent work has explored generating virtual inertial measurement unit (IMU) data via cross-modality transfer. While video-based and language-based pipelines have each shown promise, they differ in assumptions and computational cost. Moreover, their effectiveness relative to traditional sensor-level data augmentation remains unclear. In this paper, we present a direct comparison between these two virtual IMU generation approaches against classical data augmentation techniques. We construct a large-scale virtual IMU dataset spanning 100 diverse activities from Kinetics-400 and simulate sensor signals at 22 body locations. The three data generation strategies are evaluated on benchmark HAR datasets (UTD-MHAD, PAMAP2, HAD-AW) using four popular models. Results show that virtual IMU data significantly improves performance over real or augmented data alone, particularly under limited-data conditions. We offer practical guidance on choosing data generation strategies and highlight the distinct advantages and disadvantages of each approach.

Figures

Figures reproduced from arXiv: 2506.07612 by the authors.

Figure 1
Figure 1. An overview of the three ways to scaling data. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Per-class macro F1 scores on the PAMAP2 dataset for the DeepConvLSTM with self-attention model under different [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Kinetic Mining in Context: Few-Shot Action Synthesis via Text-to-Motion Distillation

    cs.CV 2025-12 conditional novelty 7.0 of 10

    A CLIP-guided teacher-student pipeline distills a text-to-motion prior into a few-shot action-to-motion generator, improving HAR top-1 accuracy by 23.1 points on 3 NTU-120 classes.

  2. VSMP-IMU: Video-Grounded Semantic Motion Programs for Sensor-Aware Synthetic IMU Generation

    cs.CV 2026-08 conditional novelty 6.0 of 10

    A structured semantic motion program extracted from video enables controllable synthetic IMU generation that improves HAR accuracy, especially with little data and imbalanced classes.

Reference graph

Works this paper leans on

45 extracted references · 31 canonical work pages · cited by 2 Pith papers

  1. [1]

    REDUCELRONPLATEAU

    2023. REDUCELRONPLATEAU. https://pytorch.org/docs/stable/generated/ torch.optim.lr_scheduler.ReduceLROnPlateau.html (2024, Feb 1)

    work page 2023

  2. [2]

    Luay Alawneh, Tamam Alsarhan, Mohammad Al-Zinati, Mahmoud Al-Ayyoub, Yaser Jararweh, and Hongtao Lu. 2021. Enhancing human activity recognition us- ing deep learning and time series augmented data.Journal of Ambient Intelligence and Humanized Computing (2021), 1–16

    work page 2021

  3. [3]

    Sizhe An, Ganapati Bhat, Suat Gumussoy, and Umit Ogras. 2023. Transfer learning for human activity recognition using representational analysis of neural networks. ACM Transactions on Computing for Healthcare 4, 1 (2023), 1–21

    work page 2023

  4. [4]

    Chen Chen, Roozbeh Jafari, and Nasser Kehtarnavaz. 2015. UTD-MHAD: A multimodal dataset for human action recognition utilizing a depth camera and a wearable inertial sensor. In2015 IEEE International Conference on Image Processing (ICIP). 168–172. https://doi.org/10.1109/ICIP.2015.7350781

    work page arXiv 2015

  5. [5]

    Hui Chen, Charles Gouin-Vallerand, Kévin Bouchard, Sébastien Gaboury, Mélanie Couture, Nathalie Bier, and Sylvain Giroux. 2024. Enhancing human activity recognition in smart homes with self-supervised learning and self-attention. Sensors 24, 3 (2024), 884

    work page 2024

  6. [6]

    Hui Chen, Charles Gouin-Vallerand, Kévin Bouchard, Sébastien Gaboury, Hu- bert Kenfack Ngankam, Maxime Lussier, Mélanie Couture, Nathalie Bier, and Sylvain Giroux. 2024. Utilizing Self-Supervised Learning for Recognizing Human Activity in Older Adults through Labeling Applications in Real-World Smart Homes. In Proceedings of the 2024 International Confere...

    work page 2024

  7. [7]

    Wenqiang Chen, Shupei Lin, Elizabeth Thompson, and John Stankovic. 2021. SenseCollect: We Need Efficient Ways to Collect On-body Sensor-based Human Activity Data! Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 5, 3 (2021), 1–27

    work page 2021

  8. [8]

    Yuwen Chen, Kunhua Zhong, Ju Zhang, Qilong Sun, and Xueliang Zhao. 2016. LSTM networks for mobile human activity recognition. In 2016 International conference on artificial intelligence: technologies and applications . Atlantis Press, 50–53

    work page 2016

Show all 45 references
  1. [9]

    Ji Seok Choi and Jung Keun Lee. 2023. Effects of data augmentation on the nine- axis IMU-based orientation estimation accuracy of a recurrent neural network. Sensors 23, 17 (2023), 7458

    2023

  2. [10]

    Sourish Gunesh Dhekane and Thomas Ploetz. 2024. Transfer learning in human activity recognition: A survey. arXiv preprint arXiv:2401.10185 (2024)

    2024 arXiv

  3. [11]

    Iveta Dirgová Luptáková, Martin Kubovčík, and Jiří Pospíchal. 2022. Wearable Sensor-Based Human Activity Recognition with Transformer Model. Sensors 22, 5 (2022). https://doi.org/10.3390/s22051911

    2022 doi

  4. [12]

    Hassan Ismail Fawaz, Germain Forestier, Jonathan Weber, Lhassane Idoumghar, and Pierre-Alain Muller. 2018. Data augmentation using synthetic data for time series classification with deep residual networks. arXiv:1808.02455 [cs.CV] https://arxiv.org/abs/1808.02455

    2018 arXiv

  5. [13]

    N. Y. Hammerla, R. Kirkham, P. Andras, and T. Ploetz. 2013. On preserving statistical characteristics of accelerometry data using their empirical cumulative distribution. In Proceedings of the 2013 international symposium on wearable computers. 65–68

    2013

  6. [14]

    Harish Haresamudram, Irfan Essa, and Thomas Plötz. 2022. Assessing the State of Self-Supervised Human Activity Recognition Using Wearables. Proc. ACM Interact. Mob. Wearable Ubiquitous Technol. 6, 3 (2022). https://doi.org/10.1145/ 3550299

    2022

  7. [15]

    Yifan Hu. 2023. Bsdgan: Balancing sensor data generative adversarial networks for human activity recognition. In 2023 International Joint Conference on Neural Networks (IJCNN). IEEE, 1–8

    2023

  8. [16]

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, Mustafa Suleyman, and Andrew Zisserman. 2017. The Kinetics Human Action Video Dataset. arXiv:1705.06950 [cs.CV] https://arxiv.org...

    2017 arXiv

  9. [17]

    Hyeokhyen Kwon, Gregory D Abowd, and Thomas Plötz. 2019. Handling an- notation uncertainty in human activity recognition. In Proceedings of the 23rd International Symposium on Wearable Computers . 109–117

    2019

  10. [18]

    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

  11. [19]

    Hyeokhyen Kwon, Bingyao Wang, Gregory D Abowd, and Thomas Plötz. 2021. Approaching the Real-World: Supporting Activity Recognition Training with Virtual IMU Data. Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 5, 3 (2021), 1–32

    2021

  12. [20]

    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. (2024). arXiv:2402.01049 [cs.CV]

    2024 arXiv

  13. [22]

    Zikang Leng, Hyeokhyen Kwon, and Thomas Ploetz. 2023. Generating Virtual On-Body Accelerometer Data from Virtual Textual Descriptions for Human Activity Recognition. In Proceedings of the 2023 ACM International Symposium on Wearable Computers. Association for Computing Machine...

    2023

  14. [23]

    Xi’ang Li, Jinqi Luo, and Rabih Younes. 2020. ActivityGAN: Generative adversarial networks for data augmentation in sensor-based human activity recognition. In Adjunct proceedings of the 2020 ACM international joint conference on pervasive and ubiquitous computing and proceedi...

    2020

  15. [24]

    MinYen Lu, ChenHao Chen, Shigemi Ishida, Yugo Nakamura, and Yutaka Arakawa. 2022. A study on estimating the accurate head IMU motion from Video. Proceedings of the Symposium on Multimedia, Distributed, Cooperative, and Mobile (DICOMO) 2022 2022 (07 2022), 918–923. https://cir....

    2022

  16. [25]

    Sara Mohammed, Reda Elbasiony, and Walid Gomaa. 2018. An LSTM-based Descriptor for Human Activities Recognition using IMU Sensors. 504–511. https: //doi.org/10.5220/0006902405040511

    2018 doi

  17. [26]

    Sebastian Münzner, Philip Schmidt, Attila Reiss, Michael Hanselmann, Rainer Stiefelhagen, and Robert Dürichen. 2017. CNN-based sensor fusion techniques for multimodal human activity recognition. In Proceedings of the 2017 ACM International Symposium on Wearable Computers (Maui...

    2017

  18. [27]

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. 2018. Representation learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748 (2018)

    2018 arXiv

  19. [28]

    OpenAI, :, Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, Red Avila, Igor Babuschkin, Suchir Balaji, Valerie Balcom, et al. 2023. GPT-4 Technical Report. arXiv:2303.0...

    2023 arXiv

  20. [29]

    Francisco Javier Ordóñez and Daniel Roggen. 2016. Deep Convolutional and LSTM Recurrent Neural Networks for Multimodal Wearable Activity Recognition. Sensors (2016)

    2016

  21. [30]

    Thomas Plötz. 2023. If only we had more data!: Sensor-Based Human Activity Recognition in Challenging Scenarios. In 2023 IEEE International Conference on Pervasive Computing and Communications Workshops and other Affiliated Events (PerCom Workshops). 565–570. https://doi.org/1...

    2023

  22. [31]

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. 2021. Learning Transferable Visual Models From Natural Language Supervision. In Proceedings...

    2021

  23. [32]

    Attila Reiss and Didier Stricker. 2012. Introducing a New Benchmarked Dataset for Activity Monitoring (ISWC ’12). IEEE Computer Society. https://doi.org/10. 1109/ISWC.2012.13

    2012

  24. [33]

    Vitor Fortes Rey, Peter Hevesi, Onorina Kovalenko, and Paul Lukowicz. 2019. Let There Be IMU Data: Generating Training Data for Wearable, Motion Sensor Based Activity Recognition from Monocular RGB Videos. In Adjunct Proceedings of the 2019 ACM International Joint Conference o...

    2019 doi

  25. [34]

    Connor Shorten and Taghi M Khoshgoftaar. 2019. A survey on image data augmentation for deep learning. Journal of big data 6, 1 (2019), 1–48

    2019

  26. [35]

    Singh, Madan Kumar Sharma, Aimé Lay-Ekuakille, Deepak Gangwar, and Sukrit Gupta

    Satya P. Singh, Madan Kumar Sharma, Aimé Lay-Ekuakille, Deepak Gangwar, and Sukrit Gupta. 2021. Deep ConvLSTM With Self-Attention for Human Activity Decoding Using Wearable Sensors. IEEE Sensors Journal 21, 6 (2021), 8575–8582. https://doi.org/10.1109/JSEN.2020.3045135

    2021

  27. [36]

    Elnaz Soleimani and Ehsan Nazerfard. 2021. Cross-subject transfer learning in human activity recognition systems using generative adversarial networks. Neurocomputing 426 (2021), 26–34

    2021

  28. [37]

    Lena Uhlenberg and Oliver Amft. 2022. Comparison of Surface Models and Skeletal Models for Inertial Sensor Data Synthesis. In 2022 IEEE-EMBS Interna- tional Conference on Wearable and Implantable Body Sensor Networks (BSN) . 1–5. https://doi.org/10.1109/BSN56160.2022.9928504

    2022

  29. [38]

    Fanyi Xiao, Ling Pei, Lei Chu, Danping Zou, Wenxian Yu, Yifan Zhu, and Tao Li

  30. [39]

    Yufei Xu, Jing Zhang, Qiming Zhang, and Dacheng Tao. 2022. ViTPose: Simple Vision Transformer Baselines for Human Pose Estimation. In Advances in Neural Information Processing Systems

    2022

  31. [40]

    Yamane and Y

    K. Yamane and Y. Nakamura. 2003. Natural motion animation through constrain- ing and deconstraining at will. IEEE Transactions on Visualization and Computer Graphics 9, 3 (2003), 352–360. https://doi.org/10.1109/TVCG.2003.1207443

    2003 arXiv

  32. [41]

    A. D. Young, M. J. Ling, and D. K. Arvind. 2011. IMUSim: A simulation envi- ronment for inertial sensing algorithm design and evaluation. In Proceedings of the 10th ACM/IEEE International Conference on Information Processing in Sensor Networks. 199–210

    2011

  33. [42]

    Junjie Zhang, Yuanhao Liu, and Hua Yuan. 2023. Attention-Based Residual BiLSTM Networks for Human Activity Recognition.IEEE Access 11 (2023), 94173– 94187. https://doi.org/10.1109/ACCESS.2023.3310269

    2023

  34. [43]

    Jinlu Zhang, Zhigang Tu, Jianyu Yang, Yujin Chen, and Junsong Yuan. 2022. MixSTE: Seq2seq Mixed Spatio-Temporal Encoder for 3D Human Pose Estimation in Video. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). 13232–13242

    2022

  35. [44]

    Jianrong Zhang, Yangsong Zhang, Xiaodong Cun, Shaoli Huang, Yong Zhang, Hongwei Zhao, Hongtao Lu, and Xi Shen. 2023. T2M-GPT: Generating Human Motion from Textual Descriptions with Discrete Representations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Patte...

    2023

  36. [45]

    Wang Zhao, Shaohui Liu, Yezhi Shu, and Yong-Jin Liu. 2020. Towards Better Generalization: Joint Depth-Pose Learning without PoseNet. In Proceedings of IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    2020

  37. [2021]

    In Spatial Data and Intelligence

    A Deep Learning Method for Complex Human Activity Recognition Using Virtual Wearable Sensors. In Spatial Data and Intelligence . Springer International Publishing. https://doi.org/10.1007/978-3-030-69873-7_19

    doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.