Pith. sign in

REVIEW 5 major objections 6 minor 27 references

DeepConvContext: A Multi-Scale Approach to Timeseries Classification in Human Activity Recognition

T0 review · 5 major / 6 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 The paper claims that adding a second LSTM over a sequence of sliding windows gives human activity recognition the missing inter-window context, raising average F1 from 49.09% to 59.04% across six datasets.

desk verdict A plausible two-LSTM HAR architecture with broad benchmarking, but the unspecified 'unwindowing' rule and a mislabeled baseline make the headline numbers non-reproducible until fixed. read the letter →

arxiv 2505.20894 v2 pith:I2KFQZLV submitted 2025-05-27 cs.LG cs.HCeess.IV

Marius Bock , Juergen Gall , Michael Moeller , Kristof Van Laerhoven This is my paper
classification cs.LGcs.HCeess.IV
keywords humanactivityrecognitionmulti-scaletimeseriesclassificationinter-windowlearningDeepConvLSTMslidingwindowversusattentiontemporalactionlocalizationinertialsensordata
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 argues that sliding-window human activity recognition is held back because each window is classified independently, so the model never sees temporal context across windows. It proposes DeepConvContext, which first extracts within-window features with convolutions and an LSTM, then runs a second LSTM over the sequence of window summaries to learn across-window dependencies. On six benchmark datasets this multi-scale design improves average macro F1 from 49.09% for the classic DeepConvLSTM to 59.04%, and raises segment-level mean average precision from near zero to 18.23, indicating fewer fragmented activity predictions. It further claims that LSTMs outperform attention and transformer modules for this cross-window modeling, and that a bidirectional cross-window LSTM gives the highest F1 and mAP.

What carries the argument

The load-bearing mechanism is a two-stage multi-scale architecture. Stage one mirrors DeepConvLSTM: four convolutional layers with 64 filters of size 9x1 (5x1 for Opportunity) plus a one-layer LSTM with 128 hidden units, producing a per-patch temporal feature vector. A projection layer compresses each patch to a one-dimensional vector of size 1 x 128. Stage two is a second LSTM, also with 128 hidden units, that processes the ordered sequence of patch vectors as a new time series; its per-patch output goes through a dropout layer and a fully connected classifier. Because the model is trained and tested with the batch treated as a temporally ordered sequence and batch shuffling is disabled, the second LSTM can relate each window to earlier windows in the batch. This second LSTM is the inter-window learning mechanism, and a bidirectional variant additionally reads future windows.

What would settle it

Rerun the six-dataset benchmark with the unwindowing rule explicitly fixed to, say, center-window labeling for every model; if DeepConvContext no longer beats both DeepConvLSTM and Shallow DeepConvLSTM in average F1 and mAP under that rule, the paper's central claim fails.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that inter-window temporal context, not a larger window or an attention mechanism, is what sliding-window HAR models are missing. Across six datasets (WEAR, Wetlab, Hang-Time, RWHAR, Opportunity, and SBHAR), DeepConvContext achieves an average macro F1 of 59.04%, against 49.09% for DeepConvLSTM and 54.01% for Shallow DeepConvLSTM; its segment-level mAP of 18.23 is higher than Shallow's 16.79 and far above DeepConvLSTM's near-zero mAP. Ablations show that a one-layer LSTM outperforms two-layer LSTMs everywhere except Opportunity, and that LSTM-based cross-window modeling beats attention and transformer variants in average F1 (59.04% versus 56.17% and 53.50% for unidirectional variants). The bidirectional LSTM variant is the best overall, reaching 62.82% average F1.

Load-bearing premise

The reported F1 and mAP comparisons hinge on an unspecified 'unwindowing' rule that converts overlapping window predictions to per-sample labels; if that rule smooths or votes across windows, it could favor exactly the coherent-output model the paper proposes.

Editorial extensions

If this is right

  • The gap between DeepConvLSTM and both inter-window models shows that letting the network see neighboring windows matters more than the depth of the recurrent stack; adding a second LSTM over window summaries is enough to capture it.
  • Because mAP measures segment coherence, the large mAP gain is a concrete promise that outputs will contain fewer rapid activity switches, which matters for continuous monitoring and real-time applications.
  • Making the cross-window LSTM bidirectional adds robustness for transitional activities such as sit-to-stand, at the cost of requiring future windows before a prediction is made.
  • Training and test batch size acts as a temporal context-length hyperparameter: average F1 rises from 48.76% at batch 25 to 59.67% at batch 200, so practitioners must set it with activity duration in mind.
  • DeepConvContext has roughly 2.5 times more parameters than the baseline but adds only about 3.5% GPU memory and 10% floating-point operations, so the context gain is not bought with a large compute penalty.

Reading between the lines

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

  • The same two-stage recipe should transfer to other HAR backbones, since it only requires compressing window-level features into one vector before a second sequence model; the paper expects this but does not test it.
  • The reported mAP comparison may be sensitive to the unwindowing rule; a rule that smooths overlapping window predictions would reward models already emitting coherent segment predictions, and specifying this rule is necessary before the mAP gain can be taken at face value.
  • The batch-size results suggest that context length in seconds, not just window size, is the hyperparameter that should be reported and tuned for each deployment, because the model cannot be run online until a buffer of b*(w-o) seconds has been collected.
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 / 6 minor

Summary. The paper proposes DeepConvContext, a multi-scale architecture for inertial human activity recognition that models both intra-window and inter-window temporal dependencies. The model processes sequences of sliding windows through a DeepConvLSTM-like feature extractor, then passes the per-window representations through a second LSTM (or an attention/Transformer alternative) to capture cross-window context. The authors evaluate the method on six HAR benchmark datasets using leave-one-subject-out cross-validation, reporting macro F1 and mean Average Precision (mAP), and compare against the classic DeepConvLSTM and the Shallow DeepConvLSTM. They also run ablations on LSTM depth and on replacing the inter-window LSTM with attention or Transformer modules. The central claim is that DeepConvContext achieves the highest average F1 (59.04%) and mAP (18.23) across the six datasets, and that inter-window learning produces more temporally coherent activity segments. The manuscript also discusses limitations related to online deployment and computational cost.

Significance. If the results hold, DeepConvContext offers a simple and reproducible way to extend window-based HAR models with inter-window context, and the systematic comparison of LSTM versus attention for inter-window modeling is a useful contribution to the ongoing discussion about temporal context in HAR. The authors provide code, describe their architectures in detail, and evaluate on six diverse public benchmarks, which is a strength. However, the significance is currently limited by several load-bearing empirical issues: the unwinding procedure that converts windowed predictions to per-sample predictions is not specified, the headline comparison uses an inconsistent baseline configuration, no measures of variability are reported, and the closest prior inter-window methods are not quantitatively compared. These issues must be resolved before the claimed improvements can be regarded as established.

major comments (5)
  1. [Section 3.4] The unwinding rule that maps windowed predictions to per-sample predictions is never defined, even though both F1 and mAP are computed on the resulting per-sample predictions. The choice of mapping—center-sample assignment, majority voting over overlapping windows, or segment stitching—can substantially affect both metrics, and mAP in particular rewards temporally coherent predicted segments. If the unwinding rule applies any temporal smoothing or majority filtering, it could systematically favor models that already emit coherent segment-level predictions, such as DeepConvContext, while partially correcting the fragmented outputs of the independent-window DeepConvLSTM. The paper must specify the exact rule, justify its fairness across models, and ideally test the sensitivity of the reported average F1 (59.04%) and mAP (18.23) to alternative unwinding choices.
  2. [Section 4.1 and Table 1] The headline comparison is internally inconsistent. Section 4.1 states that the Shallow DeepConvLSTM achieves an average F1 of 54.01%, but Table 1 lists 56.47% for the 1-layer Shallow model (the configuration used in Figure 2, whose per-dataset values match the 1-layer column) and 54.01% for the 2-layer variant. The correct comparison against the 1-layer Shallow baseline is therefore 59.04% versus 56.47%, not 59.04% versus 54.01%. This reduces the claimed margin from about 5 percentage points to about 2.6 percentage points and changes the per-dataset narrative: DeepConvContext is actually worse than the Shallow DeepConvLSTM on WEAR (77.39% vs. 78.66%) and Wetlab (38.21% vs. 38.68%). The text and abstract need to be corrected accordingly.
  3. [Section 4.1 and Table 1] No standard deviations, confidence intervals, or significance tests are reported for any of the headline results, even though all numbers are averages over only three random seeds. Given that the corrected average F1 gap between DeepConvContext and the 1-layer Shallow DeepConvLSTM is 2.57 percentage points, and that the per-dataset differences are often small (e.g., 0.53 points on Wetlab), the claim that DeepConvContext 'outperforms' the Shallow baseline is not statistically supported without variability information. The authors should report per-seed results or at least standard deviations, and perform a paired significance test across datasets.
  4. [Sections 2 and 5] The closest prior methods that extend temporal context across windows—CausalBatch [16] and the context-length study of Hiremath and Ploetz [9]—are discussed in the related work and mentioned in the limitations, but are never quantitatively compared. Since DeepConvContext is explicitly positioned against these inter-window approaches, and the abstract claims improvements 'over related approaches', the evaluation should include these baselines on the same benchmarks or provide a clear justification for their exclusion. A complexity-only comparison (Table 3) is not sufficient to support the performance claim.
  5. [Abstract and Section 5] The abstract claims that DeepConvContext produces 'more coherent activity segments even in online prediction scenarios', but Section 5 states that 'DeepConvContext cannot be directly applied in an online setting' because it requires buffering b×(w−o) seconds of data. This is a direct contradiction. The authors should either remove the online claim from the abstract or clarify what 'online' means in that context and reconcile it with the stated limitation.
minor comments (6)
  1. [Abstract] The code URL is inconsistent: the abstract lists github.com/mariusbock/deepconvcontext, while the full-text abstract lists github.com/mariusbock/context_har. Please unify the URLs.
  2. [Section 4.1] The text says 'the inter-window DeepConvLSTM struggles to produce coherent activity segments', but the DeepConvLSTM is the intra-window baseline; the Shallow DeepConvLSTM is the inter-window one. This terminology error makes the sentence confusing.
  3. [Table 4] The caption of Table 4 says 'One can see that on average LSTMs outperform both attention and Transformers', but the table reports results for varying batch sizes (25, 50, 200) and contains no attention or Transformer results. This appears to be a copy-paste error from Table 2.
  4. [Section 4.2] The claim that 'single-layer LSTMs tend to outperform their two-layer counterparts across all datasets, with the exception of the Opportunity dataset' is not fully accurate: for DeepConvContext on Opportunity, the 1-layer variant (39.08%) actually outperforms the 2-layer variant (33.39%), so the exception does not apply uniformly to all architectures.
  5. [Figure 2] The mAP values in Figure 2 are difficult to read because the bars are very small for the DeepConvLSTM and the numeric labels overlap. Consider using a separate panel or a log scale for mAP, and adding error bars based on the three seeds.
  6. [Table 3] The table header says 'MFLOPS' while the caption and text refer to 'GigaFLOPS' (GFLOPS). Please make the units consistent.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the central claim is an empirical benchmark comparison, and the self-citations are not load-bearing.

full rationale

DeepConvContext is an empirical architecture paper. The central claim—that the multi-scale model outperforms DeepConvLSTM and Shallow DeepConvLSTM in F1-score and mAP on six datasets—is established by running the proposed architecture against baselines, not by a derivation from assumptions. No equation in Sections 3.1–3.4 defines an output in terms of the reported result; LSTM hidden size, batch size, and training schedule are stated as experimental settings, not fitted to produce the headline numbers. The only self-referential elements are the choice of the Shallow DeepConvLSTM baseline from [4] and the WEAR and Hang-Time datasets from [5,10]; these function as comparison systems and benchmarks, and the outperformance claim also covers external datasets (Wetlab, SBHAR, RWHAR, Opportunity) and the external DeepConvLSTM baseline, so the conclusion does not reduce to a self-citation. The unspecified 'unwindowing' rule in Section 3.4 is a reproducibility gap and could affect the magnitude of the mAP gains, but it is not a circular reduction: the paper does not define the metric in terms of the model's output structure, nor does it fit the rule to the data. The mixed-layer Shallow baseline average (54.01% vs 56.47% for the 1-layer variant in Table 1) is an internal inconsistency, not circularity. No circular step can be exhibited with a specific equation-to-equation reduction.

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

The central empirical claim depends on standard hyperparameters and evaluation choices rather than on fitted physical constants. Key choices such as batch size (which defines the inter-window context length), window geometry, and the unwinding rule are not swept or reported in enough detail to establish their influence. No invented entities are introduced.

free parameters (5)
  • training/test batch size = 100
    Serves as the inter-window context length for Shallow DeepConvLSTM and DeepConvContext; the paper shows performance varies with batch size (Table 4), yet uses 100 for all main results.
  • window length and overlap = not specified in text
    Patch segmentation parameters are dataset-specific in the experiments but the exact values are not reported; the central comparison depends on them.
  • LSTM hidden size = 128
    Hidden units in both intra- and inter-patch LSTMs; chosen by the authors, not tuned.
  • convolutional filters and kernel size = 64 filters, 9x1 (5x1 for Opportunity)
    Architecture hyperparameters inherited from DeepConvLSTM; fixed for all experiments.
  • learning rate and schedule = 1e-4, x0.9 every 10 epochs
    Optimizer hyperparameters, chosen without reported tuning.
assumptions (4)
  • domain assumption Inertial sensor data can be segmented into equally sized, overlapping patches, and patch-level labels are sufficient for activity recognition.
    Section 3.1; the whole architecture operates on sliding windows and predicts one label per patch.
  • domain assumption Leave-one-subject-out cross-validation provides unbiased performance estimates for the compared architectures.
    Section 3.4; no significance testing is reported, so the LOSO protocol is assumed to justify the averages.
  • domain assumption The test-time use of batch size 100 preserves the temporal structure needed for inter-window learning without leaking information across subjects or sessions.
    Section 3.4; the model is trained without shuffling and tested with the same batch size, but the effect of this protocol on baseline fairness is not analyzed.
  • ad hoc to paper Windowed predictions can be converted to per-sample labels by an unspecified unwindowing rule that is fair across models.
    Section 3.4 states the conversion but not the rule; the mAP and F1 results depend on it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DeepConvContext: A Multi-Scale Approach to Timeseries Classification in Human Activity Recognition." pith.science (2026). https://pith.science/paper/I2KFQZLV

@misc{pith2026250520894,
  author       = {Pith},
  title        = {Pith review of: DeepConvContext: A Multi-Scale Approach to Timeseries Classification in Human Activity Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/I2KFQZLV}},
  note         = {Machine review of arXiv:2505.20894}
}
read the original abstract

Despite recognized limitations in modeling long-range temporal dependencies, Human Activity Recognition (HAR) has traditionally relied on a sliding window approach to segment labeled datasets. Deep learning models like the DeepConvLSTM typically classify each window independently, restricting learnable temporal context to within-window information and producing fragmented, temporally incoherent activity timelines. To address this constraint, we propose DeepConvContext, a multi-scale time series classification framework for HAR. Drawing inspiration from the vision-based Temporal Action Localization community, DeepConvContext models both intra- and inter-window temporal patterns separately by processing sequences of time-ordered windows. Across six widely-used HAR benchmarks, DeepConvContext achieves an average 5% improvement in F1-score and up to 18-point improvement in mAP over related approaches, while achieving latency and throughput comparable to prior methods that extend temporal context through hidden state propagation across batches. Our quantitative and qualitative analysis underline the importance of inter-window learning and show how it produces more coherent activity segments even in online prediction scenarios. Code to reproduce our experiments is publicly available via www.github.com/mariusbock/deepconvcontext.

Figures

Figures reproduced from arXiv: 2505.20894 by the authors.

Figure 1
Figure 1. illustrates the data flow of the proposed DeepConvContext architecture. This architecture follows a multi-scale approach, in￾corporating methods for extracting temporal features from inertial sensor time series, inspired by both the original [14] and Shallow DeepConvLSTM [4]. Given a multivariate time series from a set of inertial sensor axes, DeepConvContext begins by segmenting the sequence into equal-sized patche… view at source ↗
Figure 2
Figure 2. Average F1-score and mAP results of the DeepConvLSTM [ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Per-class confusion matrices of the DeepConvLSTM, Shallow DeepConvLSTM and DeepConvContext being applied [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

27 extracted references · 11 canonical work pages

  1. [16]

    Lloyd Pellatt and Daniel Roggen. 2020. CausalBatch: Solving Complex- ity/Performance Tradeoffs for Deep Convolutional and Lstm Networks for Wear- able Activity Recognition. InACM International Joint Conference on Pervasive and Ubiquitous Computing and ACM International Symposium on Wearable Computers. doi:10.1145/3410530.3414365

    work page arXiv 2020

  2. [9]

    Shruthi Kashinath Hiremath and Thomas Ploetz. 2021. On the Role of Context Length for Feature Extraction and Sequence Modeling in Human Activity Recog- nition. In2021 International Symposium on Wearable Computers. ACM, Virtual USA, 13–17. doi:10.1145/3460421.3478825

    work page arXiv 2021

  3. [1]

    Ranasinghe

    Alireza Abedin, Mahsa Ehsanpour, Qinfeng Shi, Hamid Rezatofighi, and Damith C. Ranasinghe. 2021. Attend and Discriminate: Beyond the State- Of-The-Art for Human Activity Recognition Using Wearable Sensors.ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies5, 1 (2021), 1–22. doi:10.1145/3448083

    doi:10.1145/3448083 2021

  4. [2]

    Bengio, P

    Y. Bengio, P. Simard, and P. Frasconi. 1994. Learning Long-Term Dependencies with Gradient Descent Is Difficult.IEEE Transactions on Neural Networks5, 2 (1994), 157–166. doi:10.1109/72.279181

    doi:10.1109/72.279181 1994

  5. [3]

    Marius Bock, Alexander Hoelzemann, Michael Moeller, and Kristof Van Laer- hoven. 2022. Investigating (Re)Current State-of-the-Art in Human Activ- ity Recognition Datasets.Frontiers in Computer Science4 (2022), 924954. doi:10.3389/fcomp.2022.924954

    work page arXiv 2022

  6. [4]

    Marius Bock, Alexander Hölzemann, Michael Moeller, and Kristof Van Laerhoven

    work page

  7. [5]

    Marius Bock, Hilde Kuehne, Kristof Van Laerhoven, and Michael Moeller. 2024. WEAR: An Outdoor Sports Dataset for Wearable and Egocentric Activity Recog- nition.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies8, 4 (2024), 1–21. doi:10.1145/3699776

    doi:10.1145/3699776 2024

  8. [6]

    Marius Bock, Michael Moeller, and Kristof Van Laerhoven. 2024. Temporal Action Localization for Inertial-based Human Activity Recognition.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies8, 4 (2024). doi:10.1145/3699770 DeepConvContext: A Multi-Scale Approach to Timeseries Classification in Human Activity Recognition

    doi:10.1145/3699770 2024

Show all 27 references
  1. [7]

    Andreas Bulling, Ulf Blanke, and Bernt Schiele. 2014. A Tutorial on Human Activity Recognition Using Body-Worn Inertial Sensors.Comput. Surveys46, 3 (2014), 1–33. doi:10.1145/2499621

    2014 doi

  2. [10]

    Alexander Hoelzemann, Julia Lee Romero, Marius Bock, Kristof Van Laerhoven, and Qin Lv. 2023. Hang-Time HAR: A Benchmark Dataset for Basketball Activity Recognition Using Wrist-Worn Inertial Sensors.Sensors23, 13 (2023). doi:10. 3390/s23135879

    2023

  3. [11]

    Andrej Karpathy, Justin Johnson, and Fei-Fei Li. 2015. Visualizing and Under- standing Recurrent Networks.CoRRabs/1506.02078 (2015). doi:10.48550/arXiv. 1506.02078

    2015 doi arXiv

  4. [12]

    Alexander Kolesnikov, Alexey Dosovitskiy, Dirk Weissenborn, Georg Heigold, Jakob Uszkoreit, Lucas Beyer, Matthias Minderer, Mostafa Dehghani, Neil Houlsby, Sylvain Gelly, Thomas Unterthiner, and Xiaohua Zhai. 2021. An Image Is Worth 16x16 Words: Transformers for Image Recognit...

    2021 doi arXiv

  5. [13]

    Murahari and Thomas Plötz

    Vishvak S. Murahari and Thomas Plötz. 2018. On Attention Models for Human Activity Recognition. InACM International Symposium on Wearable Computers. doi:10.1145/3267242.3267287

    2018

  6. [14]

    Francisco Javier Ordóñez and Daniel Roggen. 2016. Deep Convolutional and LSTM Recurrent Neural Networks for Multimodal Wearable Activity Recognition. MDPI Sensors16, 1 (2016). doi:10.3390/s16010115

    2016 doi

  7. [15]

    Razvan Pascanu, Tomas Mikolov, and Yoshua Bengio. 2013. On the Difficulty of Training Recurrent Neural Networks. InProceedings of the 30th International Con- ference on Machine Learning (Proceedings of Machine Learning Research, Vol. 28), Sanjoy Dasgupta and David McAllester (...

    2013

  8. [17]

    Reyes-Ortiz, Luca Oneto, Albert Samà, Xavier Parra, and Davide An- guita

    Jorge-L. Reyes-Ortiz, Luca Oneto, Albert Samà, Xavier Parra, and Davide An- guita. 2016. Transition-Aware Human Activity Recognition Using Smartphones. Neurocomputing171 (2016). doi:10.1016/j.neucom.2015.07.085

    2016 doi

  9. [18]

    Daniel Roggen, Alberto Calatroni, Mirco Rossi, Thomas Holleczek, Kilian Förster, Gerhard Tröster, Paul Lukowicz, David Bannach, Gerald Pirkl, Alois Ferscha, Jakob Doppler, Clemens Holzmann, Marc Kurz, Gerald Holl, Ricardo Chavar- riaga, Hesam Sagha, Hamidreza Bayati, Marco Cre...

  10. [19]

    Scholl, Matthias Wille, and Kristof Van Laerhoven

    Philipp M. Scholl, Matthias Wille, and Kristof Van Laerhoven. 2015. Wearables in the Wet Lab: A Laboratory System for Capturing and Guiding Experiments. InACM International Joint Conference on Pervasive and Ubiquitous Computing. doi:10.1145/2750858.2807547

    2015

  11. [20]

    Shuai Shao, Yu Guan, and Victor Sanchez. 2024. Beyond Isolated Frames: Enhanc- ing Sensor-Based Human Activity Recognition through Intra- and Inter-Frame Attention. doi:10.48550/arXiv.2405.19349 arXiv:2405.19349 [eess]

    2024 doi arXiv

  12. [21]

    Timo Sztyler and Heiner Stuckenschmidt. 2016. On-Body Localization of Wearable Devices: An Investigation of Position-Aware Activity Recognition. InIEEE International Conference on Pervasive Computing and Communications. doi:10.1109/PERCOM.2016.7456521

    2016

  13. [22]

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. InAdvances in Neural Information Processing Systems. doi:10.48550/ arXiv.1706.03762

    2017 doi arXiv

  14. [23]

    Rui Xi, Mengshu Hou, Mingsheng Fu, Hong Qu, and Daibo Liu. 2018. Deep Dilated Convolution on Multimodality Time Series for Human Activity Recog- nition. InIEEE International Joint Conference on Neural Networks. doi:10.1109/ IJCNN.2018.8489540

    2018

  15. [24]

    Cheng Xu, Duo Chai, Jie He, Xiaotong Zhang, and Shihong Duan. 2019. InnoHAR: A Deep Neural Network for Complex Human Activity Recognition.IEEE Access 7 (2019). doi:10.1109/ACCESS.2018.2890675

    2019

  16. [25]

    Chen-Lin Zhang, Jianxin Wu, and Yin Li. 2022. Actionformer: Localizing Mo- ments of Actions With Transformers. InEuropean Conference on Computer Vision. doi:10.1007/978-3-031-19772-7_29

    2022 doi

  17. [26]

    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. InACM International Symposium on Wearable Computers. doi:10.1145/3544794.3558467

    2022

  18. [2010]

    InIEEE Seventh International Conference on Networked Sensing Systems

    Collecting Complex Activity Datasets in Highly Rich Networked Sensor Environments. InIEEE Seventh International Conference on Networked Sensing Systems. doi:10.1109/INSS.2010.5573462

    2010

  19. [2021]

    InProceedings of the 2021 ACM International Symposium on Wearable Computers

    Improving Deep Learning for HAR with Shallow LSTMs. InProceedings of the 2021 ACM International Symposium on Wearable Computers. doi:10.1145/ 3460421.3480419

    2021

Pith tools

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