Pith. sign in

REVIEW 3 major objections 5 minor 73 references

Ground Reaction Force Estimation via Time-aware Knowledge Distillation

T0 review · 3 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 A time-aware knowledge distillation framework trains a compact student network to estimate ground reaction forces from wearable insole sensor videos, matching or beating standard distillation baselines while using a fraction of the…

desk verdict The temporal extent map is a plausible idea, but the paper's own Table III contradicts its headline claim that TaKD outperforms current baselines. read the letter →

arxiv 2506.10265 v1 pith:S7UEM7NF submitted 2025-06-12 eess.SP cs.CVcs.HC

Eun Som Jeon , Sinjini Mitra , Jisoo Lee , Omik M. Save , Ankita Shukla , Hyunglae Lee , Pavan Turaga This is my paper · ORCID
classification eess.SPcs.CVcs.HC
keywords groundreactionforceestimationknowledgedistillationwearableinsolesensorstemporalextentmapssimilarity-preservinggaitanalysismodelcompressionleave-one-subject-outevaluation
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

Ground reaction force (GRF) is the force the ground exerts on the foot during walking, normally measured with expensive, non-portable instrumented treadmills. This paper claims that a compact student network can estimate GRF from low-cost insole pressure videos when it is distilled from a larger teacher with a loss that transfers both similarity and temporal structure. The proposed Time-aware Knowledge Distillation (TaKD) matches teacher-student similarity maps inside a mini-batch and temporal extent maps across time steps, and it is evaluated leave-one-subject-out at two window lengths and four walking speeds. Across three teacher architectures, the best variant TaKD$^{\dagger}$ ranks first or second in RMSE/MAE among the compared distillation methods in five of the six settings, while the student uses roughly a quarter to a sixth of the teachers' parameters. The paper concludes that temporal-aware distillation is a data- and resource-efficient route to wearable gait monitoring.

What carries the argument

The central object is the TaKD loss $L_{BT} = L_{gt} + \lambda_1(L_{bsm} + \kappa L_{tpm}) + \lambda_2(L_{bsi} + \kappa L_{tpi})$, where $L_{gt}$ is MSE against treadmill GRF, and the distillation terms compare teacher and student feature maps through two relational descriptors: the similarity map $G = F F^\top$ with $F \in \mathbb{R}^{b \times cthw}$ (relations among the $b$ samples in a mini-batch) and the temporal extent map $P = F F^\top$ with $F \in \mathbb{R}^{t \times bchw}$ (relations among the $t$ time steps). For 1D CNN features the same construction is used with $F \in \mathbb{R}^{c \times t}$ or $F \in \mathbb{R}^{t \times bc}$. These maps allow a student whose intermediate feature dimensions differ from the teacher's to still match relational structure, and the temporal map is the component that carries the sequential knowledge shared by video input and time-series target. Layer-selection ablations show that applying both maps at the compressed representation and at high-level layers near it (TaKD$^{\dagger}$) works better than adding more layers (TaKD$^{\ddagger}$).

What would settle it

Re-run the leave-one-subject-out comparison with the $[0,1]$ normalization computed only on the training folds (and applied to the test subject), and check whether TaKD's RMSE/MAE advantage over the strongest baseline (e.g., DIST) persists and remains significant at the 90% confidence level used in the paper.

Watch

Extended reading notes

Core claim

The central discovery is that transferring temporal extent, not just sample-wise similarity, is what makes the distilled student learn useful cross-modal representations for GRF estimation. Given a teacher trained on insole pressure videos ($16 \times 8$ frames over time) to predict treadmill GRF time series, TaKD supervises the student with the ground-truth loss and with feature-map matching losses: similarity maps $G = F F^\top$ over the mini-batch capture relations between samples, while temporal extent maps $P = F F^\top$ over the time axis capture how features evolve across the walking cycle. The student, a C3D encoder with a 1D convolutional decoder and far fewer parameters ($0.25$M vs $1.00$M$-$ $1.58$M teachers), is trained to align these maps at selected encoder/decoder layers, with bilinear interpolation when map sizes differ. The authors report that their best variant, TaKD$^{\dagger}$, which uses only high-level layers near the middle representation, gives the lowest or second-lowest RMSE/MAE in most teacher and window-size combinations, and that a two-stage learning strategy (pretraining the 1D decoder with a 1D encoder before joint training with the 3D encoder) further improves results. The paper frames this as evidence that sequential knowledge is a shared property between video input and time-series target, and that exploiting it in distillation yields accurate, lightweight GRF estimators.

Load-bearing premise

The results assume that scaling every variable with the global minimum and maximum of the whole dataset, including the held-out subject, does not smuggle that subject's measurement statistics into training; if it does, the reported accuracy gaps over the baselines could shrink or disappear.

Editorial extensions

If this is right

  • A student with $0.25$M parameters can estimate GRF with RMSE close to teachers of $1.00$M$-$ $1.58$M parameters, making real-time inference on portable hardware feasible without a treadmill.
  • Temporal extent knowledge consistently improves distilled students over similarity-only transfer (SP), and the best weight in the reported settings is small ($\kappa < 0.5$), meaning sequential structure helps most as a complement to similarity matching.
  • Using more intermediate layers for transfer can hurt: the variant with fewer selected layers (TaKD$^{\dagger}$) outperforms the variant with more layers (TaKD$^{\ddagger}$), so layer selection is itself a design decision.
  • A two-stage learning strategy that first trains the 1D decoder with a 1D encoder, then updates the 3D encoder jointly, improves RMSE across teacher types and is compatible with both AE and WAE teachers.
  • Temporal-aware distillation also improves expected calibration error on GRF estimates for some teacher-student pairs, suggesting the benefit is not only accuracy but reliability of confidence.

Reading between the lines

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

  • If the improvement transfers to more subjects and shoe sizes, the same distillation recipe could turn inexpensive insoles into at-home gait monitors, with downstream uses in rehabilitation and fall-risk screening; the paper itself notes that broader subject pools and footwear remain future work.
  • The similarity-plus-temporal map pair is a generic recipe for any regression task where input is video and target is a time series (for example, joint angles or heart rate from camera footage), so the method may be testable on other datasets without architectural changes.
  • The reported leave-one-subject-out evaluation scales all data, including the held-out subject, to $[0,1]$ with global min-max statistics; re-running the comparison with normalization fitted only on training subjects would show whether the accuracy gaps survive a stricter subject-independent protocol.
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 Time-aware Knowledge Distillation (TaKD), an extension of similarity-preserving knowledge distillation for regression from wearable insole pressure videos to ground reaction force time series. The method adds two matching losses on intermediate features: a batch similarity Gram matrix and a temporal Gram matrix computed by reshaping features with time as the first dimension. The authors built a synchronized insole/treadmill dataset for eight subjects, trained three teacher encoders (C3D, I3D, (2+1)D) and a smaller C3D student, and evaluated TaKD variants against KD, AT, SP, KD+SP, DIST, and SemCKD under leave-one-subject-out. The paper claims that TaKD outperforms current baselines and reduces model size.

Significance. The dataset collection and the temporal-extent map are potentially useful contributions to the wearable-sensor distillation literature, and the paper includes a broad set of ablations over teacher architectures, window lengths, autoencoder variants, and layer selections. However, the central empirical claim is not supported by the reported tables: DIST, a baseline, is better than the best TaKD variant in 11 of the 18 teacher–window–metric cells, and the statistical analysis never compares TaKD with DIST or SemCKD. The normalization protocol in Section IV-A1 also leaks held-out subject statistics into training. As written, the paper does not establish that TaKD is an improvement over current baselines; at best it documents a large engineering study and a new dataset.

major comments (3)
  1. [Abstract; §IV-B, Table III] The central claim that TaKD "outperforms current baselines" is contradicted by the paper's own results. In Table III, for Teacher1, the best TaKD variant (TaKD†) loses to DIST on all six RMSE/MAE/correlation comparisons for both W200 and W100. Across all teachers and windows, DIST beats TaKD† in 11 of 18 metric cells, and every winning margin is within the reported ±1 standard deviation (e.g., Teacher1 W200 RMSE: DIST 6.106±0.214 vs TaKD† 6.213±0.252). Since the method's main claim is superiority over baselines, this evidence fails to support it.
  2. [§IV-A1] The normalization step uses "the absolute difference between the maximum and minimum observed values in the dataset for each variable." Under leave-one-subject-out, the held-out subject's data are part of this dataset, so the global min/max statistics of the test subject are available to every training fold. This is a leakage of test distribution information and undermines the leave-one-subject-out generalization numbers; normalization must be recomputed per training fold using only that fold's training subjects.
  3. [§IV-A3, §V-C, Table IV] Hyperparameters λ1, λ2, κ are claimed to be "empirically determined" without describing a validation split, and the sensitivity analysis in Figure 10 is performed on a smaller dataset; it is therefore unclear whether the final hyperparameters were selected using held-out subjects or the test set itself. Additionally, Table IV's significance tests compare TaKD variants only against Student, KD, SP, and KD+SP; there is no test against DIST or SemCKD, so the paper provides no statistical evidence for the "outperforms current baselines" claim.
minor comments (5)
  1. [§VI] The sentence reporting TaKD* with an MAE of 6.88% bodyweight appears inconsistent with the Student MAEs (5.15% and 5.35%) quoted immediately before; clarify whether this is a typo and which window/teacher configuration is used.
  2. [Table IV caption] The caption says "Accuracy (%)" but the rows report RMSE; correct the caption.
  3. [§III-C2, Eq. (14)] The loss terms L_bsi and L_tpi are introduced without explicit definitions; define each term once rather than only inside the text.
  4. [§V-A, Table VIII] The "channel similarity map" notation Q = F F^T with F ∈ R^{c×bthw} is dimensionally unclear; specify the actual reshaping and the normalization.
  5. [§II-B, Eq. (3)] The discriminator loss appears to have a typo (log of an expression with no probability ratio); standard WAE-GAN losses should be written in the usual form.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: TaKD's losses are empirical training objectives and the claim is an experimental comparison, not a derivation equivalent to its inputs.

full rationale

The paper's central claim is that the proposed TaKD loss (Eq. 14) yields lower RMSE/MAE and higher correlation than baselines on insole-to-GRF estimation. This is established by training teachers and students and measuring against treadmill data in a leave-one-subject-out protocol; no fitted parameter is renamed as a prediction, and no quantity used in the loss is definitionally equal to the reported metric. The similarity-map term (Eq. 11) and temporal-extent term (Eq. 13) are transposes or reshapings of intermediate features; they are training losses, not evaluation metrics, so improving them is not the same as improving GRF accuracy by construction. The ablations (Tables VII, VIII, Figure 10) test actual RMSE, so the component contributions are empirically supported rather than assumed. The paper cites the authors' prior work [37] for the general observation that a better teacher does not guarantee a better student; this citation is not load-bearing for the claimed novelty and is corroborated by the paper's own WAE-vs-AE comparisons. There is no imported uniqueness theorem and no ansatz smuggled in via citation. Concerns raised by the skeptic, such as dataset-wide normalization in Section IV-A1 and the absence of DIST/SemCKD rows in the significance table, are data-leakage and statistical-support issues, not circularity; they do not make the derivation equivalent to its inputs.

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

The paper's central claim rests on several hand-chosen hyperparameters (lambda_1, lambda_2, kappa, fraction matrix coefficients) and on the assumption that temporal Gram matrices are complementary to batch similarity maps. No code, data, or formal verification is provided, so the empirical evidence is the only support for these choices.

free parameters (5)
  • lambda_1 = 0.01
    Loss weight on middle-representation similarity and temporal terms in Eq. 14; chosen empirically in Section IV-A3 and ablated in Section V-C.
  • lambda_2 = 0.1
    Loss weight on intermediate-layer similarity and temporal terms in Eq. 14; chosen empirically.
  • kappa = 0.1
    Relative weight of temporal extent maps versus similarity maps in Eqs. 11 and 13; set to 0.1 after sensitivity analysis in Section V-C.
  • fraction_matrix_boundary_coefficients = 0.33, 0.67
    Spatial filter weights assigned by visual inspection of insole pixel coverage in Section III-B; hand-chosen to mitigate boundary artifacts.
  • contrastive_loss_weight = 0.1
    Weight of cosine similarity contrastive loss in Strategy2 and Strategy3, introduced in Section IV-D.
assumptions (4)
  • domain assumption Insole pressure maps, arranged as video frames, contain enough information to regress vertical ground reaction force time series.
    Central premise of the system; no physiological or physical model is provided, only empirical correlation.
  • domain assumption Knowledge distillation via feature similarity maps transfers useful regression knowledge even when input (video) and target (time series) modalities differ.
    Motivated by prior KD literature and assumed throughout Section III-C.
  • ad hoc to paper The temporal Gram matrix P = F F^T with F in R^{t x bchw} captures complementary temporal knowledge beyond the batch similarity map of SP.
    Core novelty; the paper asserts complementarity in Section III-C2 and Figure 5 without a theoretical justification.
  • standard math Standard neural network optimization assumptions (Adam, ReLU, MSE loss, etc.) are appropriate for training and distillation.
    Implicit in the training pipeline detailed in Section IV-A3.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Ground Reaction Force Estimation via Time-aware Knowledge Distillation." pith.science (2026). https://pith.science/paper/S7UEM7NF

@misc{pith2026250610265,
  author       = {Pith},
  title        = {Pith review of: Ground Reaction Force Estimation via Time-aware Knowledge Distillation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S7UEM7NF}},
  note         = {Machine review of arXiv:2506.10265}
}
read the original abstract

Human gait analysis with wearable sensors has been widely used in various applications, such as daily life healthcare, rehabilitation, physical therapy, and clinical diagnostics and monitoring. In particular, ground reaction force (GRF) provides critical information about how the body interacts with the ground during locomotion. Although instrumented treadmills have been widely used as the gold standard for measuring GRF during walking, their lack of portability and high cost make them impractical for many applications. As an alternative, low-cost, portable, wearable insole sensors have been utilized to measure GRF; however, these sensors are susceptible to noise and disturbance and are less accurate than treadmill measurements. To address these challenges, we propose a Time-aware Knowledge Distillation framework for GRF estimation from insole sensor data. This framework leverages similarity and temporal features within a mini-batch during the knowledge distillation process, effectively capturing the complementary relationships between features and the sequential properties of the target and input data. The performance of the lightweight models distilled through this framework was evaluated by comparing GRF estimations from insole sensor data against measurements from an instrumented treadmill. Empirical results demonstrated that Time-aware Knowledge Distillation outperforms current baselines in GRF estimation from wearable sensor data.

Figures

Figures reproduced from arXiv: 2506.10265 by the authors.

Figure 1
Figure 1. Overview of Time-aware Knowledge Distillation, TaKD. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Arrangement of insole sensor and system architecture [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 2
Figure 2. Examples for insole sensors and shoes. In [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: An example of ground reaction force (GRF) data and [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: Illustration of maps representing similarity and tempo [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 6
Figure 6. Figure 6: Results of Student (blue) and TaKD† (orange) for all subjects. C. Analysis on Autoencoder Methods To analyze the effectiveness of different autoencoders on our method, we compare AE [48], VAE [49], and WAE [50] trained teachers in distillation. For WAE, the discriminat…
Figure 8
Figure 8. Figure 8: Illustration of strategies to train estimation models. [PITH_FULL_IMAGE:figures/full_fig_p010_8.png]
Figure 9
Figure 9. Figure 9: Activation maps representing similarity and temporal properties produced by the indicated layer of the network. Window [PITH_FULL_IMAGE:figures/full_fig_p011_9.png]
Figure 10
Figure 10. Figure 10: Results of TaKD† on various λ1 and κ. D. Model Reliability To investigate the model generalizability, we use expected calibration error (ECE) [70]–[72]. ECE is widely used to mea￾sure the miscalibration of a confidence measure, representing the reliability of the mode…
Figure 11
Figure 11. Figure 11: Illustration of the estimation results from various mod [PITH_FULL_IMAGE:figures/full_fig_p012_11.png]
Figure 12
Figure 12. Figure 12: Illustration of the estimation results from various [PITH_FULL_IMAGE:figures/full_fig_p013_12.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

73 extracted references · 59 canonical work pages

  1. [1]

    Bring gait lab to everyday life: Gait analysis in terms of activities of daily living,

    D. Chen, Y . Cai, X. Qian, R. Ansari, W. Xu, K.-C. Chu, and M.- C. Huang, “Bring gait lab to everyday life: Gait analysis in terms of activities of daily living,”IEEE Internet of Things Journal, vol. 7, no. 2, pp. 1298–1312, 2019. JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 14

    work page 2019

  2. [2]

    mhealth technologies toward active health information collection and tracking in daily life: A dynamic gait monitoring example,

    Y . Cai, X. Qian, H. Cao, J. Zheng, W. Xu, and M.-C. Huang, “mhealth technologies toward active health information collection and tracking in daily life: A dynamic gait monitoring example,”IEEE Internet of Things Journal, vol. 9, no. 16, pp. 15 077–15 088, 2022

    work page 2022

  3. [3]

    Sensor-based and vision-based human activity recognition: A compre- hensive survey,

    L. M. Dang, K. Min, H. Wang, M. J. Piran, C. H. Lee, and H. Moon, “Sensor-based and vision-based human activity recognition: A compre- hensive survey,”Pattern Recognition, vol. 108, p. 107561, 2020

    work page 2020

  4. [4]

    Design a compact wireless iot gait monitor wearable sensory system,

    M. Abdelhady, H. E. Elsheikh, and W. A. Ahmed, “Design a compact wireless iot gait monitor wearable sensory system,” inInternational Telecommunications Conference (ITC-Egypt). IEEE, 2021, pp. 1–4

    work page 2021

  5. [5]

    Wearable internet of things gait sensors for quantitative assessment of myers–briggs type indicator personality,

    Y . Zhao, H. Xing, X. Wang, Y . Tian, T. Sun, M. Chen, D. Y . Yeung, S. M. Ho, J. Wang, and W. J. Li, “Wearable internet of things gait sensors for quantitative assessment of myers–briggs type indicator personality,” Advanced Intelligent Systems, vol. 6, no. 3, p. 2300328, 2024

    work page 2024

  6. [6]

    A smart environment- adapting timed-up-and-go system powered by sensor-embedded insoles,

    Z. Yang, C. Song, F. Lin, J. Langan, and W. Xu, “A smart environment- adapting timed-up-and-go system powered by sensor-embedded insoles,” IEEE Internet of Things Journal, vol. 6, no. 2, pp. 1298–1305, 2018

    work page 2018

  7. [7]

    Early balance training with a computerized stabilometric platform in persons with mild hemiparesis in subacute stroke phase: A randomized controlled pilot study,

    S. Brunelli, N. Gentileschi, M. Iosa, F. R. Fusco, V . Grossi, S. Duri, C. Foti, and M. Traballesi, “Early balance training with a computerized stabilometric platform in persons with mild hemiparesis in subacute stroke phase: A randomized controlled pilot study,”Restorative Neu- rology and Neuroscience, vol. 38, no. 6, pp. 467–475, 2020

    work page 2020

  8. [8]

    R. Li, Y . Zhang, Y . Jiang, M. Wang, W. H. D. Ang, and Y . Lau, “Reha- bilitation training based on virtual reality for patients with parkinson’s disease in improving balance, quality of life, activities of daily living, and depressive symptoms: a systematic review and meta-regression analysis,”Clinical Rehabilitation, vol. 35, no. 8, pp. 1089–1102, 2021

    work page 2021

Show all 73 references
  1. [10]

    Rethinking running biomechanics: a critical review of ground reaction forces, tibial bone loading, and the role of wearable sensors,

    L. Xiang, Z. Gao, A. Wang, V . Shim, G. Fekete, Y . Gu, and J. Fernandez, “Rethinking running biomechanics: a critical review of ground reaction forces, tibial bone loading, and the role of wearable sensors,”Frontiers in Bioengineering and Biotechnology, vol. 12, p. 1377383, 2024

    2024

  2. [11]

    Musculoskeletal injury recovery assessment using gait analysis with ground reaction force sensor,

    J. Chakraborty, S. Upadhyay, and A. Nandy, “Musculoskeletal injury recovery assessment using gait analysis with ground reaction force sensor,”Medical Engineering & Physics, vol. 103, p. 103788, 2022

    2022

  3. [12]

    Comparison of ground reaction force and marker-based methods to estimate mediolateral center of mass displacement and margins of stability during walking,

    T. J. Buurke, L. van de Venis, R. den Otter, J. Nonnekes, and Keijsers, “Comparison of ground reaction force and marker-based methods to estimate mediolateral center of mass displacement and margins of stability during walking,”Journal of biomechanics, vol. 146, p. 111415, 2023

    2023

  4. [13]

    Artificial neural network-based ground reaction force estimation and learning for dynamic-legged robot systems,

    J. An and I. Lee, “Artificial neural network-based ground reaction force estimation and learning for dynamic-legged robot systems,”PeerJ Computer Science, vol. 9, p. e1720, 2023

    2023

  5. [14]

    The analysis of movement,

    J. Key, “The analysis of movement,” inBack Pain - A Movement Problem, J. Key, Ed. Elsevier, 2010, pp. 37–54

    2010

  6. [15]

    Ground reaction force and moment estimation through EMG sensing using long short-term memory network during posture coordination,

    S.-I. Sakamoto, Y . Hutabarat, D. Owaki, and M. Hayashibe, “Ground reaction force and moment estimation through EMG sensing using long short-term memory network during posture coordination,”Cyborg Bionic Syst., vol. 4, p. 0016, 2023

    2023

  7. [16]

    Estimating vertical ground reaction force during walking using a single inertial sensor,

    X. Jiang, C. Napier, B. Hannigan, J. J. Eng, and C. Menon, “Estimating vertical ground reaction force during walking using a single inertial sensor,”Sensors (Basel), vol. 20, no. 15, p. 4345, 2020

    2020

  8. [17]

    Comparison of vertical ground reaction forces during overground and treadmill running. a validation study,

    B. Kluitenberg, S. W. Bredeweg, S. Zijlstra, W. Zijlstra, and I. Buist, “Comparison of vertical ground reaction forces during overground and treadmill running. a validation study,”BMC musculoskeletal disorders, vol. 13, pp. 1–8, 2012

    2012

  9. [18]

    Developing a low-cost force treadmill via dynamic modeling,

    C.-Y . Hong, L.-Y . Guo, R. Song, M. L. Nagurka, J.-L. Sung, and C.- W. Yen, “Developing a low-cost force treadmill via dynamic modeling,” Journal of Healthcare Engineering, vol. 2017, no. 1, p. 9875471, 2017

    2017

  10. [19]

    Robust gait recognition based on deep cnns with camera and radar sensor fusion,

    Y . Shi, L. Du, X. Chen, X. Liao, Z. Yu, Z. Li, C. Wang, and S. Xue, “Robust gait recognition based on deep cnns with camera and radar sensor fusion,”IEEE Internet of Things Journal, vol. 10, no. 12, pp. 10 817–10 832, 2023

    2023

  11. [20]

    Validation of a wireless shoe insole for ground reaction force measurement,

    G. T. Burns, J. Deneweth Zendler, and R. F. Zernicke, “Validation of a wireless shoe insole for ground reaction force measurement,”Journal of sports sciences, vol. 37, no. 10, pp. 1129–1138, 2019

    2019

  12. [21]

    Functional data analyses of gait data measured using in-shoe sensors,

    J. Lee, G. Li, W. F. Christensen, G. Collins, M. Seeley, A. E. Bowden, D. T. Fullwood, and J. Goldsmith, “Functional data analyses of gait data measured using in-shoe sensors,”Statistics in biosciences, vol. 11, pp. 288–313, 2019

    2019

  13. [22]

    Center of pressure estimation by analyzing walking videos,

    J. Chen, Y . Qin, P. Lin, J. Li, Y . Xue, and H. Ma, “Center of pressure estimation by analyzing walking videos,” inIEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2024, pp. 3460–3464

    2024

  14. [23]

    Powered prosthesis locomo- tion on varying terrains: Model-dependent control with real-time force sensing,

    R. Gehlhar, J.-H. Yang, and A. D. Ames, “Powered prosthesis locomo- tion on varying terrains: Model-dependent control with real-time force sensing,”IEEE Robot. Autom. Lett., vol. 7, no. 2, pp. 5151–5158, 2022

    2022

  15. [24]

    Instrumented insole vs. force plate: A comparison of center of plantar pressure,

    P. S. Dyer and S. J. M. Bamberg, “Instrumented insole vs. force plate: A comparison of center of plantar pressure,” in2011 Annual International Conference of the IEEE Engineering in Medicine and Biology Society. IEEE, 2011, pp. 6805–6809

    2011

  16. [25]

    Variability of ground reaction forces during treadmill walking,

    K. Masani, M. Kouzaki, and T. Fukunaga, “Variability of ground reaction forces during treadmill walking,”Journal of applied physiology, vol. 92, no. 5, pp. 1885–1890, 2002

    2002

  17. [26]

    Reducing the dimensionality of data with neural networks,

    G. E. Hinton and R. R. Salakhutdinov, “Reducing the dimensionality of data with neural networks,”Science, vol. 313, no. 5786, pp. 504–507, 2006

    2006

  18. [27]

    Densely connected convolutional networks,

    G. Huang, Z. Liu, L. Van Der Maaten, and K. Q. Weinberger, “Densely connected convolutional networks,” inProceedings of the IEEE Confer- ence on Computer Vision and Pattern Recognition, 2017, pp. 4700–4708

    2017

  19. [28]

    Deep learning for time series classification: a review,

    H. I. Fawaz, G. Forestier, J. Weber, L. Idoumghar, and P.-A. Muller, “Deep learning for time series classification: a review,”Data Mining and Knowledge Discovery, vol. 33, no. 4, pp. 917–963, 2019

    2019

  20. [29]

    A survey of the recent architectures of deep convolutional neural networks,

    A. Khan, A. Sohail, U. Zahoora, and A. S. Qureshi, “A survey of the recent architectures of deep convolutional neural networks,”Artificial Intelligence Review, vol. 53, no. 8, pp. 5455–5516, 2020

    2020

  21. [30]

    Improving physical activity recognition using a new deep learning architecture and post-processing techniques,

    M. Gil-Mart ´ın, R. San-Segundo, F. Fernandez-Martinez, and J. Ferreiros- L´opez, “Improving physical activity recognition using a new deep learning architecture and post-processing techniques,”Engineering Ap- plications of Artificial Intelligence, vol. 92, p. 103679, 2020

    2020

  22. [31]

    Employing of ma- chine learning and wearable devices in healthcare system: tasks and challenges,

    H. S. Saad, J. F. Zaki, and M. M. Abdelsalam, “Employing of ma- chine learning and wearable devices in healthcare system: tasks and challenges,”Neural Computing and Applications, pp. 1–21, 2024

    2024

  23. [32]

    Wearabledl: Wear- able internet-of-things and deep learning for big data analytics—concept, literature, and future,

    A. R. Dargazany, P. Stegagno, and K. Mankodiya, “Wearabledl: Wear- able internet-of-things and deep learning for big data analytics—concept, literature, and future,”Mobile Information Systems, vol. 2018, no. 1, p. 8125126, 2018

    2018

  24. [33]

    A survey on multimodal wearable sensor-based human action recognition,

    J. Ni, H. Tang, S. T. Haque, Y . Yan, and A. H. Ngu, “A survey on multimodal wearable sensor-based human action recognition,”arXiv preprint arXiv:2404.15349, 2024

    2024 arXiv

  25. [34]

    A survey of convolutional neural networks: Analysis, applications, and prospects,

    Z. Li, F. Liu, W. Yang, S. Peng, and J. Zhou, “A survey of convolutional neural networks: Analysis, applications, and prospects,”IEEE Transac- tions on Neural Networks and Learning Systems, vol. 33, no. 12, pp. 6999–7019, 2022

    2022

  26. [35]

    A critical review of recurrent neural networks for sequence learning,

    Z. C. Lipton, “A critical review of recurrent neural networks for sequence learning,”ArXiv, vol. abs/1506.00019, 2015

    2015 arXiv

  27. [36]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,” inProceedings of the NeurIPS Deep Learning and Representation Learning Workshop, vol. 2, no. 7, 2015

    2015

  28. [37]

    Role of data augmentation strategies in knowledge distillation for wearable sensor data,

    E. S. Jeon, A. Som, A. Shukla, K. Hasanaj, M. P. Buman, and P. Turaga, “Role of data augmentation strategies in knowledge distillation for wearable sensor data,”IEEE Internet of Things Journal, vol. 9, no. 14, pp. 12 848–12 860, 2022

    2022

  29. [38]

    Knowledge distillation: A survey,

    J. Gou, B. Yu, S. J. Maybank, and D. Tao, “Knowledge distillation: A survey,”International Journal of Computer Vision, vol. 129, no. 6, pp. 1789–1819, 2021

    2021

  30. [39]

    A survey on heterogeneous transfer learning,

    O. Day and T. M. Khoshgoftaar, “A survey on heterogeneous transfer learning,”Journal of Big Data, vol. 4, pp. 1–42, 2017

    2017

  31. [40]

    Improved fine-tuning by better leveraging pre-training data,

    Z. Liu, Y . Xu, Y . Xu, Q. Qian, H. Li, X. Ji, A. Chan, and R. Jin, “Improved fine-tuning by better leveraging pre-training data,”Advances in Neural Information Processing Systems, vol. 35, pp. 32 568–32 581, 2022

    2022

  32. [41]

    Improved three-dimensional inception networks for hyper- spectral remote sensing image classification,

    X. Zhang, “Improved three-dimensional inception networks for hyper- spectral remote sensing image classification,”IEEE Access, vol. 11, pp. 32 648–32 658, 2023

    2023

  33. [42]

    Adaptive multiple classifiers fusion for inertial sensor based human activity recognition,

    Y . Tian, X. Wang, W. Chen, Z. Liu, and L. Li, “Adaptive multiple classifiers fusion for inertial sensor based human activity recognition,” Cluster Computing, vol. 22, no. 4, pp. 8141–8154, 2019

    2019

  34. [43]

    Center of pressure estimation by analyzing walking videos,

    J. Chen, Y . Qin, P. Lin, J. Li, Y . Xue, and H. Ma, “Center of pressure estimation by analyzing walking videos,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2024, pp. 3460–3464

    2024

  35. [44]

    Interpreting video features: A comparison of 3d convolutional networks and convo- lutional lstm networks,

    J. M ¨antt¨ari, S. Broom ´e, J. Folkesson, and H. Kjellstr ¨om, “Interpreting video features: A comparison of 3d convolutional networks and convo- lutional lstm networks,” inComputer Vision – ACCV 2020. Springer International Publishing, 2021, pp. 411–426

    2020

  36. [45]

    Human activity recognition based on multiple inertial sensors through feature-based knowledge distillation paradigm,

    M. Mardanpour, M. Sepahvand, F. Abdali-Mohammadi, M. Nikouei, and H. Sarabi, “Human activity recognition based on multiple inertial sensors through feature-based knowledge distillation paradigm,”Information Sciences, vol. 640, p. 119073, 2023. JOURNAL OF LATEX CLASS FILES, VOL...

    2023

  37. [46]

    Lightweight depth completion network with local similarity-preserving knowledge distillation,

    Y . Jeong, J. Park, D. Cho, Y . Hwang, S. B. Choi, and I. S. Kweon, “Lightweight depth completion network with local similarity-preserving knowledge distillation,”Sensors, vol. 22, no. 19, 2022

    2022

  38. [47]

    Ae-cnn-based multisource data fusion for gait motion step length estimation,

    Z.-Q. Ling, Y .-P. Zhang, G.-Z. Cao, J.-C. Chen, L.-L. Li, and D.-P. Tan, “Ae-cnn-based multisource data fusion for gait motion step length estimation,”IEEE Sensors Journal, vol. 22, no. 21, pp. 20 805–20 815, 2022

    2022

  39. [48]

    Auto-encoding variational bayes,

    D. P. Kingma and M. Welling, “Auto-encoding variational bayes,” International Conference on Learning Representations, 2014

    2014

  40. [49]

    Variational recurrent auto-encoders,

    O. Fabius, J. R. van Amersfoort, and D. P. Kingma, “Variational recurrent auto-encoders,” inInternational Conference on Learning Rep- resentations Workshops, 2015

    2015

  41. [50]

    Wasserstein auto-encoders,

    I. Tolstikhin, O. Bousquet, S. Gelly, and B. Sch ¨olkopf, “Wasserstein auto-encoders,” inInternational Conference on Learning Representa- tions, 2018

    2018

  42. [51]

    Model compression,

    C. Bucilu ˇa, R. Caruana, and A. Niculescu-Mizil, “Model compression,” inProceedings of the ACM International Conference on Knowledge Discovery and Data Mining (KDD), 2006, pp. 535–541

    2006

  43. [52]

    Paying more attention to attention: Improving the performance of convolutional neural networks via at- tention transfer,

    S. Zagoruyko and N. Komodakis, “Paying more attention to attention: Improving the performance of convolutional neural networks via at- tention transfer,” inProceedings of the International Conference on Learning and Representations (ICLR), 2017, pp. 1–13

    2017

  44. [53]

    Similarity-preserving knowledge distillation,

    F. Tung and G. Mori, “Similarity-preserving knowledge distillation,” in Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 1365–1374

    2019

  45. [54]

    Learning spatiotemporal features with 3d convolutional networks,

    D. Tran, L. Bourdev, R. Fergus, L. Torresani, and M. Paluri, “Learning spatiotemporal features with 3d convolutional networks,” inProceedings of the IEEE international conference on computer vision, 2015, pp. 4489–4497

    2015

  46. [55]

    Quo vadis, action recognition? a new model and the kinetics dataset,

    J. Carreira and A. Zisserman, “Quo vadis, action recognition? a new model and the kinetics dataset,” inproceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2017, pp. 6299–6308

    2017

  47. [56]

    A closer look at spatiotemporal convolutions for action recognition,

    D. Tran, H. Wang, L. Torresani, J. Ray, Y . LeCun, and M. Paluri, “A closer look at spatiotemporal convolutions for action recognition,” in Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2018, pp. 6450–6459

    2018

  48. [57]

    Lightweight depth completion network with local similarity-preserving knowledge distillation,

    Y . Jeong, J. Park, D. Cho, Y . Hwang, S. B. Choi, and I. S. Kweon, “Lightweight depth completion network with local similarity-preserving knowledge distillation,”Sensors, vol. 22, no. 19, p. 7388, 2022

    2022

  49. [58]

    3d convolutional neural networks for human action recognition,

    S. Ji, W. Xu, M. Yang, and K. Yu, “3d convolutional neural networks for human action recognition,”IEEE transactions on pattern analysis and machine intelligence, vol. 35, no. 1, pp. 221–231, 2012

    2012

  50. [59]

    A review of video action recognition based on 3d convolution,

    X. Huang and Z. Cai, “A review of video action recognition based on 3d convolution,”Computers and Electrical Engineering, vol. 108, p. 108713, 2023

    2023

  51. [60]

    Plantar pressure assessment,

    M. N. Orlin and T. G. McPoil, “Plantar pressure assessment,”Physical therapy, vol. 80, no. 4, pp. 399–409, 2000

    2000

  52. [61]

    Knowledge distillation from a stronger teacher,

    T. Huang, S. You, F. Wang, C. Qian, and C. Xu, “Knowledge distillation from a stronger teacher,”Advances in Neural Information Processing Systems, vol. 35, pp. 33 716–33 727, 2022

    2022

  53. [62]

    Semckd: Semantic calibration for cross-layer knowledge distillation,

    C. Wang, D. Chen, J.-P. Mei, Y . Zhang, Y . Feng, and C. Chen, “Semckd: Semantic calibration for cross-layer knowledge distillation,” IEEE Transactions on Knowledge and Data Engineering, vol. 35, no. 6, pp. 6305–6319, 2023

    2023

  54. [63]

    Be your own teacher: Improve the performance of convolutional neural networks via self distillation,

    L. Zhang, J. Song, A. Gao, J. Chen, C. Bao, and K. Ma, “Be your own teacher: Improve the performance of convolutional neural networks via self distillation,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 3713–3722

    2019

  55. [64]

    Distilling knowledge by mimicking fea- tures,

    G.-H. Wang, Y . Ge, and J. Wu, “Distilling knowledge by mimicking fea- tures,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 44, no. 11, pp. 8183–8195, 2021

    2021

  56. [65]

    Benchmarking multimodal variational autoencoders: Cdsprites+ dataset and toolkit,

    G. Sejnova, M. Vavrecka, and K. Stepanova, “Benchmarking multimodal variational autoencoders: Cdsprites+ dataset and toolkit,” 2023

    2023

  57. [66]

    On the limitations of multimodal vaes,

    I. Daunhawer, T. M. Sutter, K. Chin-Cheong, E. Palumbo, and J. E. V ogt, “On the limitations of multimodal vaes,” inInternational Conference on Learning Representations, 2022

    2022

  58. [67]

    On the efficacy of knowledge distillation,

    J. H. Cho and B. Hariharan, “On the efficacy of knowledge distillation,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), 2019, pp. 4794–4802

    2019

  59. [68]

    Cross-modal alignment and translation for missing modality action recognition,

    Y . Park, S. Woo, S. Lee, M. A. Nugroho, and C. Kim, “Cross-modal alignment and translation for missing modality action recognition,” Computer Vision and Image Understanding, vol. 236, p. 103805, 2023

    2023

  60. [69]

    Data augmentation with cross-modal variational autoencoders (dacmva) for cancer survival prediction,

    S. Rajaram and C. S. Mitchell, “Data augmentation with cross-modal variational autoencoders (dacmva) for cancer survival prediction,”Infor- mation, vol. 15, no. 1, p. 7, 2023

    2023

  61. [70]

    Regression modeling strategies with applications to linear models, logistic and ordinal regression, and survival analysis,

    E. H. Frank, “Regression modeling strategies with applications to linear models, logistic and ordinal regression, and survival analysis,” 2015

    2015

  62. [71]

    Obtaining well calibrated probabilities using bayesian binning,

    M. P. Naeini, G. Cooper, and M. Hauskrecht, “Obtaining well calibrated probabilities using bayesian binning,” inProceedings of the AAAI conference on artificial intelligence, vol. 29, no. 1, 2015

    2015

  63. [72]

    On calibration of modern neural networks,

    C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On calibration of modern neural networks,” inProceedings of the International Confer- ence on Machine Learning (ICML), 2017, pp. 1321–1330

    2017

  64. [73]

    Temporal transformer networks: Joint learning of invariant and discriminative time warping,

    S. Lohit, Q. Wang, and P. Turaga, “Temporal transformer networks: Joint learning of invariant and discriminative time warping,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019, pp. 12 426–12 435

    2019

  65. [74]

    Seamless indoor–outdoor foot-mounted inertial pedestrian positioning system enhanced by smartphone ppp/3-d map/barometer,

    J. Wang, C. Shi, M. Xia, F. Zheng, T. Li, Y . Shan, G. Jing, W. Chen, and T. C. Hsia, “Seamless indoor–outdoor foot-mounted inertial pedestrian positioning system enhanced by smartphone ppp/3-d map/barometer,” IEEE Internet of Things Journal, vol. 11, no. 7, pp. 13 051–13 069, 2024

    2024

Pith tools

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