Pith. sign in

REVIEW 5 major objections 4 minor 32 references

USAD: End-to-End Human Activity Recognition via Diffusion Model with Spatiotemporal Attention

T0 review · 5 major / 4 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 claims that a single end-to-end HAR pipeline, USAD, can attack scarce labels, weak feature extraction, and embedded-device constraints at once by combining class-conditioned diffusion data augmentation, multi-branch…

desk verdict A competent engineering combination of known HAR components whose headline accuracy numbers disagree with its own tables and whose evaluation split is never stated. read the letter →

arxiv 2507.02827 v2 pith:4EB3NRNB submitted 2025-07-03 cs.CV cs.AI

Hang Xiao , Ying Yu , Jiarui Li , Zhifan Yang , Haotian Tang , Hanyu Liu , Chao Li This is my paper · ORCID
classification cs.CVcs.AI
keywords humanactivityrecognitiondiffusionmodeldataaugmentationspatiotemporalattentionmulti-scalenetworkadaptivelossfunctionclassimbalanceembeddeddeployment
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

The paper sets out to show that the three persistent bottlenecks in sensor-based human activity recognition—scarce labels for rare activities, weak high-level feature extraction, and poor fit on low-power devices—can be addressed in one end-to-end framework. The proposed USAD pipeline generates synthetic sensor sequences with a statistics-guided diffusion model, pretrains a classifier on those sequences, fine-tunes it on real data, and then relies on a multi-branch spatiotemporal attention network with an adaptively weighted composite loss. If the paper is right, the method raises accuracy on WISDM, PAMAP2, and OPPORTUNITY to 98.84%, 93.81%, and 80.92%, respectively, while keeping inference light enough for an embedded board. A sympathetic reader would take the central bet to be that synthetic pretraining is what closes the class-imbalance gap.

What carries the argument

The load-bearing object is the class-conditioned diffusion generator. It takes a noisy sequence $x_t$, a time embedding, a feature vector $f$ built from global statistics (mean, standard deviation, skewness) and Z-scores, and a label embedding $y$, and denoises with adaptive group normalization; sampling is guided by class prototypes $\mu_y = \mathbb{E}[f \mid y]$. The paper argues via a Wasserstein-distance bound, $|\mathbb{E}_{x\sim p_{\text{syn}}}\mathcal{L}(C_\phi(x)) - \mathbb{E}_{x\sim p_{\text{real}}}\mathcal{L}(C_\phi(x))| \le \mathcal{O}(W(p_{\text{syn}}, p_{\text{real}}))$, that if the synthetic distribution is close to the real one, pretraining the classifier $C_\phi$ on synthetic data cannot hurt much before fine-tuning. Around that generator sits a multi-branch network with residual cardinal groups, radix-style channel soft attention, spatial and temporal attention, and an adaptively reweighted sum of label-smoothing, focal, and cross-entropy losses.

What would settle it

Retrain USAD on a stated subject-independent split of WISDM or PAMAP2 with the diffusion augmentation removed and every other hyperparameter fixed; if accuracy does not fall outside run-to-run noise, the synthetic pretraining is not doing the claimed work. Separately, the abstract's OPPORTUNITY figure (80.92%) should be reproducible from the reported protocol; reproducing it, or finding the protocol that yields it, would settle the headline claim.

Watch

Extended reading notes

Core claim

The central claim is that a classifier pretrained on diffusion-generated sensor sequences conditioned on per-class statistical prototypes transfers to real data well enough to beat existing HAR models on three benchmarks, and that the gain is amplified by the network and loss design. Concretely, the paper reports USAD at 98.84% accuracy on WISDM, 93.81% on PAMAP2, and 80.92% on OPPORTUNITY, with better precision, recall, F1, G-mean, and AUC than the comparison methods under mild, moderate, and severe class imbalance. The paper also claims that the full pipeline runs within the inference time budget on a Raspberry Pi 5-class device while using less memory than most compared models, making the result usable outside the lab.

Load-bearing premise

The load-bearing premise is that the synthetic sensor sequences made by the class-conditioned diffusion model are close enough to real sensor signals that a classifier pretrained on them ends up more accurate after fine-tuning than one trained on real data alone; the paper asserts the supporting guarantee rather than proving it.

Editorial extensions

If this is right

  • USAD reports 98.84% accuracy on WISDM, above every comparison in its Table 2, and 94.07% on full PAMAP2 data.
  • On half of the PAMAP2 data, USAD still reaches 89.29%, suggesting the synthetic pretraining offsets data loss.
  • Ablations attribute the final step to diffusion augmentation: 3.36% accuracy gain on PAMAP2 and 6.55% on OPPORTUNITY over the best non-augmented configuration.
  • Deployment tables show USAD with lower memory use than most compared models on WISDM and OPPORTUNITY, and per-segment inference within the 5%-window time budget on Raspberry Pi 5.

Reading between the lines

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

  • If the synthetic-pretraining stage really is the cause of the gains, the same class-prototype-conditioned diffusion recipe should transfer to other heavily imbalanced time-series tasks, such as fall detection, gesture recognition, or medical alarm monitoring; this is not tested in the paper.
  • The Wasserstein bound in Eq. (19) is asserted without proof; a rigorous version would give a principled guide for how many synthetic samples to generate per class, which the paper does not provide.
  • A direct stress test is strict leave-subjects-out evaluation: the paper does not state whether its splits are subject-dependent, so the margin over prior work may shrink under cross-subject conditions.
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

5 major / 4 minor

Summary. The paper proposes USAD, a human activity recognition (HAR) framework that combines three components: an unsupervised diffusion-based data augmentation stage, a multi-branch spatio-temporal attention network with cardinality/radix splits, and an adaptive composite loss that dynamically weights cross-entropy, focal, and label-smoothing terms. The authors claim that USAD achieves state-of-the-art accuracies of 98.84%, 93.81%, and 80.92% on WISDM, PAMAP2, and OPPORTUNITY, respectively, and that it can be deployed efficiently on an embedded Raspberry Pi 5 device. The body reports ablation studies, comparisons with existing HAR models, class-imbalance analyses, t-SNE visualizations, confusion matrices, and complexity measurements.

Significance. If the reported results are reliable, USAD would be a useful empirical contribution to sensor-based HAR: it addresses a real problem (class imbalance and scarce labels), reports a broad set of metrics (accuracy, precision, recall, F1, G-mean, AUC), and includes an embedded-device deployment study, which is relatively rare in the HAR literature. The paper also provides extensive ablations and per-class analyses. However, the significance is substantially weakened by internal numerical inconsistencies, an incompletely specified evaluation protocol, and a theoretical claim (Eq. 19) that is asserted without proof. No code is released, and no error bars or repeated-run statistics are provided, so the central empirical claim cannot currently be independently verified.

major comments (5)
  1. [Abstract; Table 1; Table 2] The headline accuracies in the abstract do not match those in the body. The abstract reports WISDM 98.84%, PAMAP2 93.81%, and OPPORTUNITY 80.92%, while Table 1's final USAD row reports 98.28%, 94.07%, and 84.60%, and Table 2's USAD row reports WISDM 98.84% and PAMAP2 94.07%. At least one of these sets of numbers must be wrong, and the paper never explains which is authoritative. Because the central claim is that USAD 'significantly outperforms existing approaches' with these exact accuracies, this inconsistency is load-bearing and must be resolved before the results can be evaluated.
  2. [Appendix A.1] The evaluation protocol never states whether the train/test splits for WISDM, PAMAP2, and OPPORTUNITY are subject-dependent or subject-independent. The appendix lists only window size, batch size, learning rate, and number of epochs. These datasets are commonly benchmarked under both protocols, and the two protocols can differ by several percentage points. Without this information, the comparison against prior methods in Table 2, and the claim of superiority over existing approaches, is unverifiable. The paper also provides no error bars, standard deviations, or repeated-run statistics for any of the reported accuracy values.
  3. [Section 4.1, Table 1 (OPPORTUNITY column)] The text states that 'With the introduction of the ACL, the model's performance is further enhanced across all datasets,' and the abstract lists the adaptive multi-loss fusion as a contribution. However, Table 1's OPPORTUNITY column shows that the ACL row achieves 83.11% accuracy while the fixed-weight Combination 2 achieves 83.54%. The adaptive weighting scheme is therefore not uniformly better than a fixed weighting, contradicting the stated claim. This is not a minor discrepancy; it directly concerns a claimed contribution.
  4. [Section 3.1.4, Eq. (19)] Equation (19) is presented as theoretical support for the augmentation approach ('Theoretically, if the synthetic distribution p_syn approximates the real distribution p_real, the risk difference is bounded by...'), but it is an unproved assertion. No theorem statement, regularity conditions, or proof is supplied, and the function H(W(p_syn, p_real)) is undefined. As written, this is not a derivation or a bound; it is an informal statement that does not substantiate the claim that the generated synthetic data improve classifier performance.
  5. [Section 3.3; Appendix C] The adaptive loss weights and all loss hyperparameters (epsilon, gamma, alpha, beta, tau, initial weights, and weight bounds) are tuned in Appendix C using accuracy on the PAMAP2 test set, and the selected configuration is then reported as the final result. Selecting hyperparameters on the test set invalidates the reported accuracies as unbiased estimates of generalization. Additionally, the update rule in Eq. (27) is not internally consistent: if 'acc' is the accuracy fraction in [0,1], then 1/(acc+1e-8) is about 1, giving omega1 approximately 1 - tau, which is negative for tau > 1; if 'acc' is a percentage, the formula behaves differently. The paper should define the quantity and use a proper validation split.
minor comments (4)
  1. [Section 3.2, Eq. (21)] The soft attention weight formula has mismatched parentheses in the R>1 case; the denominator is not closed correctly. Also, the text refers to 'Dence' in Figure 2 where 'Dense' is meant.
  2. [Section 2, References] The in-text citations do not align with the reference list. For example, the text attributes ResNeSt to [1], but reference [1] is an attention-based BiLSTM paper; the text attributes ResNeXt to [2], but reference [2] is an 'In-sensor 24 classes HAR' paper; and the text's Att-CNN-BLSTM and AM-DLFC citations do not match the corresponding reference entries. The reference list and citation numbering need to be reconciled.
  3. [Section 4.4] The text says 'Table 4 shows the average results of ten rounds of 100 tests on WISDM and OPPO datasets,' but Table 4 reports parameter and memory values, not timing results from multiple test rounds. The relationship between the described procedure and the reported table is unclear. The dataset name 'OPPO' is also used inconsistently for OPPORTUNITY.
  4. [Appendix C.3] The text says 'as shown in Table 9, the optimal y was 0.05 in the independent optimization stage,' but Table 9 is the composite-weight experiment; the label-smoothing sensitivity results are in Table 10. The cross-references among Tables 9-12 are incorrect.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central accuracy claim is an externally benchmarked empirical result, and the only theoretical statement (Eq. 19) is unproved but not circular.

full rationale

The paper makes no derivation that reduces to its own inputs. The only theoretical statement, Eq. (19), asserts a Wasserstein transfer bound without proof; this is an unsupported assertion, a rigor and completeness issue, not a circular reduction, and it is not used to compute the reported accuracies. The augmentation pipeline (Eqs. 3-18) conditions synthetic samples on real-data class prototypes and labels, but the final accuracy is evaluated on held-out real test data after fine-tuning, so the synthetic-data pretraining is an ordinary empirical augmentation scheme rather than a fitted quantity renamed as a prediction. The adaptive loss weights in Eq. (27) are updated from the current epoch's accuracy as a feedback rule; they are not fit to the test set. Comparisons in Tables 2 and 3 are against external published baselines (CNN, LSTM, SE-Res2Net, MAG-Res2Net, ATFA, TCCSNet, ELK, DanHAR, etc.), so the central SOTA claim is externally benchmarked rather than secured by self-citation. Reference [25] (MAG-Res2Net) may share an author overlap with H. Liu, but it is used only as a comparison baseline and is not load-bearing for any derivation or uniqueness claim. The paper's genuine weaknesses, including the unproved Eq. (19), the unstated subject-dependent versus subject-independent evaluation protocol, and conflicting numbers across the abstract and body tables, are correctness and reproducibility risks rather than circularity. No circular step can be exhibited from the paper's own equations or citations.

Assumptions & free parameters 10 free parameters · 3 assumptions · 0 invented entities

The paper rests on several tuned hyperparameters whose values are only partially reported, plus unproved assumptions about synthetic data transfer and evaluation protocol. No invented entities are introduced.

free parameters (10)
  • Cardinality K = not specified
    Number of cardinal groups in the multi-branch network; not reported in experiments.
  • Radix R = not specified
    Number of splits per cardinal group; not reported in experiments.
  • Synthetic sample count M = not specified
    Size of generated dataset, Eq. (15); not reported.
  • Label smoothing coefficient epsilon = 0.1 (joint tuning), 0.05 (independent)
    Tuned on validation or test performance in Appendix C.
  • Focal loss gamma = 1
    Tuned; gamma=1 best in Tables 7 and 11.
  • Focal loss alpha = 0.25 (independent), 0.5 (joint)
    Tuned; shown in Tables 7 and 11.
  • CB loss beta = 0.9 (independent), 0.999 (joint)
    Tuned; shown in Tables 8 and 12.
  • Adaptive loss temperature tau = 1.0
    Tuned; Table 9 shows moderate temperature is best.
  • Initial loss weights [w0,w1,w2] = [0.33,0.33,0.34]
    Best uniform initialization in Table 9.
  • Weight bounds (min,max) = [0.1,0.8]
    Tuned; Table 9 shows these bounds prevent extreme weights.
assumptions (3)
  • domain assumption Synthetic distribution approximates real distribution (Eq. 19 bound)
    The Wasserstein bound is stated without proof in Section 3.1.4 and is used to justify classifier pretraining on synthetic data.
  • ad hoc to paper Adaptive loss update rule (Eq. 27) improves generalization
    The formula w1 = 2 - tau - 1/(acc+1e-8) is not derived; the OPPORTUNITY ablation shows fixed weights can outperform ACL.
  • domain assumption Diffusion model conditioned on labels can be called unsupervised
    The paper claims 'unsupervised' augmentation but Eqs. (9)-(11) condition on label y, making the unsupervised framing questionable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of USAD: End-to-End Human Activity Recognition via Diffusion Model with Spatiotemporal Attention." pith.science (2026). https://pith.science/paper/4EB3NRNB

@misc{pith2026250702827,
  author       = {Pith},
  title        = {Pith review of: USAD: End-to-End Human Activity Recognition via Diffusion Model with Spatiotemporal Attention},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4EB3NRNB}},
  note         = {Machine review of arXiv:2507.02827}
}
read the original abstract

The primary objective of human activity recognition (HAR) is to infer ongoing human actions from sensor data, a task that finds broad applications in health monitoring, safety protection, and sports analysis. Despite proliferating research, HAR still faces key challenges, including the scarcity of labeled samples for rare activities, insufficient extraction of high-level features, and suboptimal model performance on lightweight devices. To address these issues, this paper proposes a comprehensive optimization approach centered on multi-attention interaction mechanisms. First, an unsupervised, statistics-guided diffusion model is employed to perform data augmentation, thereby alleviating the problems of labeled data scarcity and severe class imbalance. Second, a multi-branch spatio-temporal interaction network is designed, which captures multi-scale features of sequential data through parallel residual branches with 3*3, 5*5, and 7*7 convolutional kernels. Simultaneously, temporal attention mechanisms are incorporated to identify critical time points, while spatial attention enhances inter-sensor interactions. A cross-branch feature fusion unit is further introduced to improve the overall feature representation capability. Finally, an adaptive multi-loss function fusion strategy is integrated, allowing for dynamic adjustment of loss weights and overall model optimization. Experimental results on three public datasets, WISDM, PAMAP2, and OPPORTUNITY, demonstrate that the proposed unsupervised data augmentation spatio-temporal attention diffusion network (USAD) achieves accuracies of 98.84%, 93.81%, and 80.92% respectively, significantly outperforming existing approaches. Furthermore, practical deployment on embedded devices verifies the efficiency and feasibility of the proposed method.

Figures

Figures reproduced from arXiv: 2507.02827 by the authors.

Figure 1
Figure 1. Framework of our USAD method, including (A) Data synthesis, (B) Multi-branch network architecture, (C) Adaptive composite loss function, (D) Embedded device deployment. We construct a feature-conditioned vector by replicating the global features: 𝑓 = [ 𝜇, …, 𝜇 ⏟⏟⏟ 𝐿 , 𝜎,…, 𝜎 ⏟⏟⏟ 𝐿 , 𝛾,…, 𝛾 ⏟⏟⏟ 𝐿 , 𝑧1 ,…, 𝑧𝐿 ]⊤ ∈ ℝ 4𝐿 (3) To associate features with class labels, we compute label-specific prototype statistics: 𝜇𝑦 = 𝔼[… view at source ↗
Figure 2
Figure 2. Detailed structure of the proposed USAD model USAD employs a multi-branch convolutional architecture, focusing on the spatial and temporal properties at different time segments. By decomposing and reorganizing features, it optimizes the specialization and diversity of feature extraction as well as the expressiveness of the network. The generated data and features presented previously are directly inserted. The featu… view at source ↗
Figure 3
Figure 3. (a) shows the model performance comparison on the WISDM dataset, (b) on the PAMAP2 dataset, and (c) on the OPPORTUNITY dataset. The radar charts use Acc., Pre., Rec., F1, G - mean, and Area Under the Curve (AUC) as dimensions to present the performance differences of the USAD model compared with mainstream models such as CNN, LSTM, ResNet50, SERes2Net, MAG - Res2Net, ATFA, and TCCSNet on datasets with different degr… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Confusion matrices for WISDM, PAMAP2, and OPPORTUNITY. Xiao el al.: Preprint submitted to Elsevier Page 11 of 20 [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: T-SNE Visualization Comparison: Feature Distributions of CNN, ResNet50, TCCSNet, and USAD Models on the (a) WISDM, (b) PAMAP2, and (c) OPPORTUNITY Dataset. Results of additional comparison models are provided in the Supplementary Material. In contrast, traditional mode…
Figure 6
Figure 6. Figure 6: Inference delay of each model on the WISDM dataset [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 8
Figure 8. Figure 8: Training dynamics under label smoothing. (a) ̂𝑦 = 0.05: Optimal balance between validation accuracy (0.9760) and ECE (0.0109). (b) ̂𝑦 = 0.1: Higher ECE (0.0125) due to over-smoothing, despite comparable accuracy. (c) ̂𝑦 = 0.01: Lower accuracy (0.9653) and slower conver…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 30 canonical work pages

  1. [17]

    S. Zuo, V . F. Rey, S. Suh, S. Sigg, P. Lukowicz, Unsupervised statistical feature-guided diffusion model for sensor-based human activity recognition, arXiv preprint arXiv:2306.05285 (2023)

    work page Pith review arXiv 2023

  2. [1]

    Zhang, Y

    J. Zhang, Y . Liu, H. Yuan, Attention-based residual bilstm networks for human activity recognition, IEEE Access 11 (2023) 94173–94187

    work page 2023

  3. [2]

    In-sensor 24 classes HAR under 850 Bytes

    A. Benmessaoud, W. Kezai, F. Medjani, K. Bouaita, T. Kezai, et al., In-sensor 24 classes har under 850 bytes, arXiv preprint arXiv:2502.17472 (2025)

    work page Pith review arXiv 2025

  4. [3]

    Y . Wang, H. Xu, Y . Liu, M. Wang, Y . Wang, Y . Yang, S. Zhou, J. Zeng, J. Xu, S. Li, et al., A novel deep multifeature extraction framework based on attention mechanism using wearable sensor data for human activity recognition, IEEE Sensors Journal 23 (7) (2023) 7188–7198

    work page 2023

  5. [4]

    Zhang, L

    J. Zhang, L. Feng, Z. Liu, Y . Wu, Y . He, Y . Dong, D. Xu, Diverse intra-and inter-domain activity style fusion for cross-person generalization in activity recognition, in: Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2024, pp. 4213–4222

    work page 2024

  6. [5]

    Y . Zhao, J. Shao, X. Lin, T. Sun, J. Li, C. Lian, X. Lyu, B. Si, Z. Zhan, Cir-dfenet: Incorporating cross-modal image representation and dual-stream feature enhanced network for activity recognition, Expert Systems with Applications 266 (2025) 125912. Xiao el al.: Preprint submitted to Elsevier Page 15 of 20 USAD

    work page 2025

  7. [6]

    J. Wang, Y . Chen, Y . Gu, Y . Xiao, H. Pan, Sensorygans: An effective generative adversarial framework for sensor-based human activity recognition, in: 2018 International Joint Conference on Neural Networks (IJCNN), IEEE, 2018, pp. 1–8

    work page 2018

  8. [7]

    X. Li, J. Luo, R. Younes, 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 proceedings of the 2020 ACM international symposium on wearable computers, 2020, pp. 249–254

    work page 2020

Show all 32 references
  1. [8]

    J. Ho, A. Jain, P. Abbeel, Denoising diffusion probabilistic models, Advances in neural information processing systems 33 (2020) 6840–6851

    2020

  2. [9]

    S. Shao, V . Sanchez, A study on diffusion modelling for sensor-based human activity recognition, in: 2023 11th International Workshop on Biometrics and Forensics (IWBF), IEEE, 2023, pp. 1–7

    2023

  3. [10]

    Y . Zhou, H. Zhao, Y . Huang, T. Röddiger, M. Kurnaz, T. Riedel, M. Beigl, Autoaughar: automated data augmentation for sensor-based human activity recognition, Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies 8 (2) (2024) 1–27

    2024

  4. [11]

    Mekruksavanich, D

    S. Mekruksavanich, D. Tancharoen, A. Jitpattanakul, A hybrid deep neural network with attention mechanism for human activity recognition based on smartphone sensors, in: 2023 7th International Conference on Information Technology (InCIT), IEEE, 2023, pp. 153–157

    2023

  5. [12]

    X. Ye, K. I.-K. Wang, Deep generative domain adaptation with temporal attention for cross-user activity recognition, arXiv preprint arXiv:2403.17958 (2024)

    2024 arXiv

  6. [13]

    J. Zhao, Y . Liu, Q. Zhang, et al., Cnn - attbilstm mechanism: A ddos attack detection method based on attention mechanism and cnn - bilstm, IEEE Access 11 (2023) 136308 – 136317

    2023

  7. [14]

    Akter, S

    M. Akter, S. Ansary, M. A.-M. Khan, D. Kim, Human activity recognition using attention-mechanism-based deep learning feature combination, Sensors 23 (12) (2023) 5715

    2023

  8. [15]

    H. Feng, Q. Shen, R. Song, L. Shi, H. Xu, Atfa: Adversarial time–frequency attention network for sensor-based multimodal human activity recognition, Expert Systems with Applications 236 (2024) 121296. doi:https://doi.org/10.1016/j.eswa.2023.121296. URL https://www.sciencedirec...

    2024

  9. [16]

    E. Essa, I. R. Abdelmaksoud, Temporal-channel convolution with self-attention network for human activity recognition using wearable sensors, Knowledge-Based Systems 278 (2023) 110867. doi:https://doi.org/10.1016/j.knosys.2023.110867. URL https://www.sciencedirect.com/science/a...

    2023

  10. [18]

    M. Zeng, L. T. Nguyen, B. Yu, O. J. Mengshoel, J. Zhu, P. Wu, J. Zhang, Convolutional neural networks for human activity recognition using mobile sensors, in: 6th international conference on mobile computing, applications and services, IEEE, 2014, pp. 197–205

    2014

  11. [19]

    K. Xia, J. Huang, H. Wang, Lstm-cnn architecture for human activity recognition, IEEE Access 8 (2020) 56855–56866

    2020

  12. [20]

    N. Dua, S. N. Singh, V . B. Semwal, Multi-input cnn-gru based human activity recognition using wearable sensors, Computing 103 (7) (2021) 1461–1478

    2021

  13. [21]

    Gao, M.-M

    S.-H. Gao, M.-M. Cheng, K. Zhao, X.-Y . Zhang, M.-H. Yang, P. Torr, Res2net: A new multi-scale backbone architecture, IEEE transactions on pattern analysis and machine intelligence 43 (2) (2019) 652–662

    2019

  14. [22]

    Mekruksavanich, P

    S. Mekruksavanich, P. Jantawong, A. Jitpattanakul, A deep learning-based model for human activity recognition using biosensors embedded into a smart knee bandage, Procedia Computer Science 214 (2022) 621–627

    2022

  15. [23]

    C. Yang, M. Jiang, Z. Guo, Y . Liu, Gated res2net for multivariate time series analysis, in: 2020 International Joint Conference on Neural Networks (IJCNN), IEEE, 2020, pp. 1–7

    2020

  16. [24]

    Pramanik, R

    R. Pramanik, R. Sikdar, R. Sarkar, Transformer-based deep reverse attention network for multi-sensory human activity recognition, Engineering Applications of Artificial Intelligence 122 (2023) 106150

    2023

  17. [25]

    H. Liu, B. Zhao, C. Dai, B. Sun, A. Li, Z. Wang, Mag-res2net: A novel deep learning network for human activity recognition, Physiological Measurement 44 (11) (2023) 115007

    2023

  18. [26]

    Huang, L

    W. Huang, L. Zhang, H. Wu, F. Min, A. Song, Channel-equalization-har: A light-weight convolutional neural network for wearable sensor based human activity recognition, IEEE Transactions on Mobile Computing 22 (9) (2023) 5064–5077. doi:10.1109/TMC.2022.3174816

    2023

  19. [27]

    M. Yao, L. Zhang, D. Cheng, L. Qin, X. Liu, Z. Fu, H. Wu, A. Song, Revisiting large-kernel cnn design via structural re-parameterization for sensor-based human activity recognition, IEEE Sensors Journal 24 (8) (2024) 12863–12876. doi:10.1109/JSEN.2024.3371462

    2024

  20. [28]

    W. Gao, L. Zhang, Q. Teng, J. He, H. Wu, Danhar: Dual attention network for multimodal human activity recognition using wearable sensors, Applied Soft Computing 111 (2021) 107728. doi:https://doi.org/10.1016/j.asoc.2021.107728. URL https://www.sciencedirect.com/science/article...

    2021

  21. [29]

    Cheng, L

    D. Cheng, L. Zhang, C. Bu, H. Wu, A. Song, Learning hierarchical time series data augmentation invariances via contrastive supervision for human activity recognition, Knowledge-Based Systems 276 (2023) 110789

    2023

  22. [30]

    J. R. Kwapisz, G. M. Weiss, S. A. Moore, Activity recognition using cell phone accelerometers, ACM SigKDD Explorations Newsletter 12 (2) (2011) 74–82

    2011

  23. [31]

    Reiss, D

    A. Reiss, D. Stricker, Introducing a new benchmarked dataset for activity monitoring, in: 2012 16th international symposium on wearable computers, IEEE, 2012, pp. 108–109

    2012

  24. [32]

    softening

    R. Chavarriaga, H. Sagha, A. Calatroni, S. T. Digumarti, G. Tröster, J. d. R. Millán, D. Roggen, The opportunity challenge: A benchmark database for on-body sensor-based activity recognition, Pattern Recognition Letters 34 (15) (2013) 2033–2042. A. Experimental Design A.1. Dat...

    2013

Pith tools

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