Pith. sign in

REVIEW 3 major objections 4 minor 32 references

A dual-path network that splits IMU signals into low- and high-frequency components claims a 33 percent cut in pedestrian trajectory error.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 21:19 UTC pith:G7GENUXR

load-bearing objection A plausible frequency-decomposed dual-path IO architecture with an internally consistent ablation, but the headline 33% ATE gain is unverifiable as reported because the baseline comparison protocol is never stated. the 3 major comments →

arxiv 2511.15645 v3 pith:G7GENUXR submitted 2025-11-19 cs.CV cs.RO

FDIO: Frequency Decomposed Inertial Odometry

Shanshan Zhang , Liqin Wu , Wenying Cao , Lingxiang Zheng , Yu Yang This is my paper
classification cs.CV cs.RO
keywords inertial odometrypedestrian localizationLaplacian pyramidMambafrequency decompositionIMUglobal framedeep learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper argues that pedestrian inertial odometry improves when IMU readings are first separated by frequency: a low-frequency path captures whole-body motion trends, and a high-frequency path captures fine local dynamics. It proposes MambaIO, which uses a Laplacian pyramid for the split, a Mamba module for long-range low-frequency context, and multi-scale convolutions for high-frequency detail. On five public datasets, it reports an average ATE of 3.221 m, a 33.3 percent reduction over the RoNIN ResNet baseline. The paper also argues, with kinematic analysis and visualizations, that a gravity-aligned global frame is preferable to the body frame for pedestrian IO.

Core claim

On the paper's own terms, the central claim is that frequency-decoupled modeling is a better inductive bias for learning pedestrian motion from IMU data than single-path architectures. The Laplacian pyramid isolates a smooth low-frequency component, modeled by Mamba for long-range contextual motion, and a high-frequency residual, modeled by multi-scale depthwise convolutions for local details. Combining the two pathways yields the lowest average absolute trajectory error (3.221 m) and relative trajectory error (2.550 m) across RIDI, RoNIN, RNIN, IMUNet, and OxIOD, beating the RoNIN ResNet baseline by 33.3 percent and 16.7 percent respectively. The ablation study shows either pathway alone un

What carries the argument

The central mechanism is the Laplacian pyramid decomposition, implemented with differentiable depthwise convolutions and nearest-neighbor upsampling: it splits each six-channel IMU window (acceleration and angular velocity, 200 samples) into a low-pass trend and a high-pass residual. The low-frequency path is handled by a Mamba state-space module with linear-complexity long-range modeling; the high-frequency path is handled by three parallel depthwise convolutions with kernel sizes 1, 3, and 7, followed by a squeeze-and-excitation block. The two outputs are concatenated and fused for velocity regression.

Load-bearing premise

The 33.3 percent ATE improvement assumes every baseline in Table I was trained and evaluated under the same preprocessing, windowing, and train/test splits as MambaIO, which the paper never specifies for the baselines.

What would settle it

Re-running RoNIN ResNet and the other baselines under the paper's exact protocol (window length 200, same coordinate-frame preprocessing, same splits) and seeing whether the gap persists; if the baseline ATE already drops to about 3.221 m, the claimed improvement collapses.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • If the frequency-decoupling claim holds, architectures for inertial odometry should explicitly separate global trends from local perturbations rather than feeding raw signals to a single network.
  • The global-frame analysis provides a reason for pedestrian IO to keep the gravity-aligned frame, contradicting recent drone-based results that favor the body frame; this distinction is tied to non-rigid mounting.
  • The differentiable Laplacian pyramid is a reusable component for other time-series tasks that need frequency separation without losing end-to-end training.
  • The reported gains (33.3 percent ATE average) would make MambaIO a practical choice for smartphone localization where only IMU data is available.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper leaves baseline protocol details unstated; until the baselines are re-run under identical preprocessing and splits, the 33.3 percent figure should be read as a point estimate rather than a controlled measurement.
  • The frequency-splitting principle could extend to other wearable sensors (e.g., foot-mounted IMUs or earbuds) where the same low/high-frequency distinction between body motion and limb jitter exists; a testable extension would be applying the same architecture to foot-mounted zero-velocity-update scenarios.
  • The authors note in the conclusion that high-frequency noise is not precisely isolated; a learnable or adaptive filter bank could refine the split and potentially improve results further, though this is not tested in the paper.
  • The ablation table shows Mamba-only is worse than Conv-only on IMUNet and OxIOD, suggesting the optimal frequency split may vary by device placement or motion type, which the paper does not address.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. This manuscript proposes MambaIO (also referred to as FDIO in the submitted abstract), a supervised pedestrian inertial odometry architecture. The paper first argues that a global coordinate frame is preferable to the body frame for pedestrian IMU signals, then decomposes the 6-axis windowed input with a differentiable Laplacian pyramid into low- and high-frequency components. A Mamba block models the low-frequency path and a multi-scale depthwise convolution module with an SE block models the high-frequency path; the two paths are fused for velocity/pose regression. Experiments on RIDI, RoNIN, RNIN, IMUNet, and OxIOD report an average ATE of 3.221 m and RTE of 2.550 m, claimed to be state-of-the-art and 33.3%/16.7% better than RoNIN ResNet. An ablation comparing Conv-only and Mamba-only variants supports the complementary benefit of the dual-path design.

Significance. If the performance claims survive a controlled comparison, the frequency-decomposed dual-path design is a genuinely interesting inductive bias for pedestrian inertial odometry and is among the first applications of Mamba to this task. The decomposition is simple and fully differentiable, and the ablation is internally consistent: the full model beats both single-path variants on nearly every dataset/metric pair. These are real strengths. However, the paper's central empirical claim stands or falls on Table I, and the current reporting does not establish that the comparison with prior methods was controlled. The significance of the contribution is therefore contingent on clarifying or re-running the baselines under a common protocol.

major comments (3)
  1. [IV-A-2, IV-C, Table I] Table I is the sole support for the state-of-the-art claim, but the paper never states how the baseline numbers were obtained. It gives training details only for MambaIO (Sec. IV-A-4) and does not say whether RoNIN ResNet, TLIO, RNIN, Tartan IMU, etc. were retrained under the same coordinate-frame preprocessing, window length L=200, train/test splits, loss function, and ATE/RTE alignment protocol. PIO results are highly sensitive to these choices (e.g., OxIOD subject splits), so the averaged 33.3% ATE / 16.7% RTE improvements over RoNIN ResNet are not a valid controlled measurement unless the baselines were evaluated in this codebase. Please state the source of every baseline number and, ideally, retrain all baselines under the same protocol; report per-seed variability.
  2. [III-C, Fig. 6] Equations (4)-(6) define a single one-level low/high decomposition, yet the method is described as a Laplacian pyramid and the stage diagram shows repeated MSFD blocks. The paper does not specify the number of pyramid levels, where downsampling occurs relative to the four feature-extraction stages, or how the channel schedule [64,128,256,512] interacts with the decomposition. Without this information the architecture cannot be reproduced exactly, and the 'multi-scale' claim is underspecified. Please provide a precise dataflow for one stage and for the full network, including the number of pyramid levels and the downsampling schedule.
  3. [IV-D, Table II] No error bars, confidence intervals, or multiple seeds are reported for either the main comparison or the ablation. Several differences in Table I are small enough to be within run-to-run noise (e.g., OxIOD RTE: MambaIO 0.980 vs Tartan IMU 0.940; average RTE: MambaIO 2.550 vs SBIPTVTL 2.733). The absence of variance reporting weakens both the SOTA claim and the conclusion that the full model is consistently better than the ablations. At minimum, report mean and standard deviation over at least three seeds for the proposed method and the key baselines.
minor comments (4)
  1. [Title/Abstract] The submission header and initial abstract use 'FDIO: Frequency Decomposed Inertial Odometry', but the full text title and abstract use 'MambaIO: Global-Coordinate Inertial Odometry for Pedestrians via Multi-Scale Frequency-Decoupled Modeling'. Please harmonize this inconsistency before publication.
  2. [Section IV-C] The trajectory discussion refers to 'Fig. 4 (a), (c), (d) and (c)', but the sample trajectories are in Fig. 8; Fig. 4 is the t-SNE plot. There is also a typo 'strATEht-line'.
  3. [Appendix] The text repeatedly refers to the Appendix for details (e.g., t-SNE on other datasets, additional coordinate-frame comparisons), but the submitted manuscript contains no appendix. Either include the material or remove the references.
  4. [Table I and References] The best/second-best color coding in Table I is not visible in plain text or grayscale; use boldface or a textual note. In addition, iMoT is listed in Table I as AAAI 2025 but reference [24] is an arXiv preprint; please verify the publication venue.

Circularity Check

0 steps flagged

No significant circularity; the central claims are empirical measurements and architectural ablations, not quantities defined by fitted parameters or self-citation.

full rationale

The paper's derivation chain is empirical rather than definitional. The coordinate-frame analysis (Sections III-A and III-B) compares body versus global representations through the kinematic equations (1)-(2), t-SNE visualizations, and trajectory plots; the conclusion that the global frame is beneficial is an experimentally tested hypothesis, not a conclusion built into its definitions. The Laplacian pyramid construction (Equations 4-6) is a standard, fully differentiable decomposition, and the routing of the low-frequency component to Mamba and the high-frequency component to multi-scale convolution is an architectural hypothesis evaluated by the ablations in Table II, not a fitted parameter renamed as a prediction. Table I reports measured ATE and RTE values on five public datasets; the headline 33.3% ATE and 16.7% RTE improvements are direct ratios of the reported errors and are not forced by construction. There are no self-citations, no imported uniqueness theorems, and no ansatz smuggled in via the authors' prior work. The legitimate concern — that the paper does not specify whether baseline numbers in Table I were produced under identical preprocessing, windowing, splits, and metrics — is an experimental-reporting and external-validity issue, not circularity. The title/abstract inconsistency (FDIO versus MambaIO) is a manuscript consistency issue, also not circularity. Honest non-finding: no significant circularity.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The central claim rests on a network whose components are assembled from prior literature (Laplacian pyramids, Mamba, depthwise multi-scale convolutions, SE blocks); the genuinely new content is the specific combination and its empirical evaluation. Everything load-bearing beyond the known components is a hand-chosen hyperparameter (window length, kernel sizes, frequency split point, channel schedule) or an inherited domain assumption (global-frame preprocessing, ground-truth labels, orientation estimation). No theorem is derived and no data is shipped; the missing loss function and missing appendix are additional unstated dependencies.

free parameters (5)
  • Input window length L=200 = 200 samples
    Hand-chosen temporal window, no ablation. It interacts with the frequency split: the stride-2 low-pass halves the effective context length (Sec. III-C-1).
  • Laplacian pyramid kernel size and stride (k=5, s=2) = k=5, s=2
    The fixed averaging kernel defines the low/high-frequency split point (Eqs. 4–6). No analysis shows this cutoff is appropriate for pedestrian gait; it determines how much of the signal each path sees.
  • MPC kernel sizes k=1,3,7 and SE block design = 1, 3, 7
    Hand-chosen receptive fields for the high-frequency path (Eq. 7); no ablation of alternative kernel sets; part of the architecture's capacity budget.
  • Channel schedule [64,128,256,512] in 4 stages
    Hand-chosen model width/depth (Sec. IV-A-4); no ablation; standard scaling chosen by the authors.
  • Number of decomposition levels = 1 (in equations)
    Eqs. 4–6 describe a single-level decomposition though Fig. 6 and the prose suggest multi-resolution; the number of levels used in the actual experiments is never stated.
axioms (4)
  • domain assumption Gravity-aligned global-frame (HACF) preprocessing is the appropriate input representation for pedestrian IO
    Section III-A/B argues this from kinematics and t-SNE; the conclusion matches RoNIN [5], so it is an inherited domain assumption, not proven by this paper.
  • ad hoc to paper Low-frequency components carry the long-range motion structure and high-frequency components carry local dynamics, so a split architecture with Mamba on low and convolutions on high is the right inductive bias
    Section III-C-1 asserts this by analogy with convolution/attention literature; it is the core design premise, validated only post hoc by ablation.
  • domain assumption The body-to-global rotation R (Eqs. 2–3) is estimated accurately enough by orientation estimation/preintegration
    The paper relies on standard IMU orientation estimation inherited from the RoNIN pipeline; no error analysis of R is given, yet the global-frame processing depends on it.
  • standard math The training pipeline (Adam, learning-rate decay, 40 epochs, multi-GPU) suffices to train the model, and an unstated loss function is appropriate
    Standard practice stated in Sec. IV-A-4, but no loss function is ever defined; the training objective is an unstated but load-bearing part of the pipeline.

pith-pipeline@v1.3.0-alltime-deepseek · 11618 in / 18982 out tokens · 186551 ms · 2026-08-03T21:19:47.072537+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of FDIO: Frequency Decomposed Inertial Odometry." pith.science (2026). https://pith.science/paper/G7GENUXR

@misc{pith2026251115645,
  author       = {Pith},
  title        = {Pith review of: FDIO: Frequency Decomposed Inertial Odometry},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G7GENUXR}},
  note         = {Machine review of arXiv:2511.15645}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Pedestrian inertial odometry (PIO) estimates autonomous pedestrian motion using only acceleration and angular velocity measurements collected by an inertial measurement unit (IMU), making it highly valuable for consumer level localization applications. However, under a dual device acquisition setting, IMU signals collected by a freely carried mobile device are inherently composite signals in which the global motion of the human torso is coupled with perturbations induced by local limb motion. This coupling makes accurate human motion modeling more challenging. To address this issue, this paper proposes frequency decomposed inertial odometry (FDIO). The proposed method first decomposes input IMU signals into low frequency and high frequency components using a Laplacian pyramid. It then adopts a Mamba module to model long range motion information from the low frequency component and uses a multi scale convolution module to extract fine grained local dynamic features from the high frequency component. Experiments on five public PIO datasets show that FDIO achieves an average absolute trajectory error of 3.221~m and an average relative trajectory error of 2.550~m, reducing the errors by 33.3\% and 16.7\% compared with the RoNIN ResNet baseline, respectively. These results validate the effectiveness of the proposed frequency decomposition strategy. To the best of our knowledge, this work is among the first efforts to introduce Mamba and a frequency decomposition architecture into inertial odometry.

Figures

Figures reproduced from arXiv: 2511.15645 by Lingxiang Zheng, Liqin Wu, Shanshan Zhang, Wenying Cao, Yu Yang.

Figure 1
Figure 1. Figure 1: Performance comparison of algorithms on the RoNIN dataset [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Schematic diagram of the coordinate system transformation [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of IO Learning Processes in body frame and [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: t-SNE visualization of features from the RIDI dataset, demon [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Schematic diagram of the proposed MambaIO architecture, including the Laplacian pyramid and the multi-path convolutional kernel [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Comparison of trajectory error for RoNIN ResNet under global [PITH_FULL_IMAGE:figures/full_fig_p006_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Visualization of sample trajectories across five datasets, comparing MambaIO with Tartan IMU. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

32 extracted references · 3 linked inside Pith

  1. [1]

    Tartan imu: A light foundation model for inertial positioning in robotics,

    S. Zhao, S. Zhou, R. Blanchard, Y . Qiu, W. Wang, and S. Scherer, “Tartan imu: A light foundation model for inertial positioning in robotics,”2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pp. 22 520–22 529, 2025. [Online]. Available: https://api.semanticscholar.org/CorpusID:280089012

    2025

  2. [2]

    Airio: Learning inertial odometry with enhanced imu feature observability,

    Y . Qiu, C. Xu, Y . Chen, S. Zhao, J. Geng, and S. Scherer, “Airio: Learning inertial odometry with enhanced imu feature observability,” IEEE Robotics and Automation Letters, pp. 1–8, 2025

    2025

  3. [3]

    Survey on inertial sensor- based ils for smartphone users,

    A. K. Panja, C. Chowdhury, and S. Neogy, “Survey on inertial sensor- based ils for smartphone users,”CCF Transactions on Pervasive Com- puting and Interaction, vol. 4, no. 3, pp. 319–337, 2022

    2022

  4. [4]

    Strapdown inertial navigation integration algorithm design part 2: Velocity and position algorithms,

    P. G. Savage, “Strapdown inertial navigation integration algorithm design part 2: Velocity and position algorithms,”Journal of Guidance, Control, and Dynamics, vol. 21, no. 2, pp. 208–221, 1998

    1998

  5. [5]

    Ronin: Robust neural inertial navigation in the wild: Benchmark, evaluations, & new methods,

    S. Herath, H. Yan, and Y . Furukawa, “Ronin: Robust neural inertial navigation in the wild: Benchmark, evaluations, & new methods,” in2020 IEEE International Conference on Robotics and Automation (ICRA), 2020, pp. 3146–3152

    2020

  6. [6]

    End-to- end learning framework for imu-based 6-dof odometry,

    J. P. Silva do Monte Lima, H. Uchiyama, and R.-i. Taniguchi, “End-to- end learning framework for imu-based 6-dof odometry,”Sensors, vol. 19, no. 17, p. 3777, 2019

    2019

  7. [7]

    A survey of indoor inertial positioning systems for pedes- trians,

    R. Harle, “A survey of indoor inertial positioning systems for pedes- trians,”IEEE Communications Surveys & Tutorials, vol. 15, no. 3, pp. 1281–1293, 2013

    2013

  8. [8]

    EqNIO: Subequivariant neural inertial odometry,

    R. K. Jayanth, Y . Xu, Z. Wang, E. Chatzipantazis, K. Daniilidis, and D. Gehrig, “EqNIO: Subequivariant neural inertial odometry,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview.net/forum?id=C8jXEugWkq

    2025

  9. [9]

    Pedestrian trajectory forecasting using deep ensembles under sensing uncertainty,

    A. Nayak, A. Eskandarian, Z. Doerzaph, and P. Ghorai, “Pedestrian trajectory forecasting using deep ensembles under sensing uncertainty,” IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 9, pp. 11 317–11 329, 2024

    2024

  10. [10]

    Adaptive threshold-based zupt for single imu-enabled wearable pedestrian local- ization,

    H. Li, H. Liu, Z. Li, C. Li, Z. Meng, N. Gao, and Z. Zhang, “Adaptive threshold-based zupt for single imu-enabled wearable pedestrian local- ization,”IEEE Internet of Things Journal, vol. 10, no. 13, pp. 11 749– 11 760, 2023

    2023

  11. [11]

    A visual-inertial approach to human gait estimation,

    A. Ahmed and S. Roumeliotis, “A visual-inertial approach to human gait estimation,” in2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2018, pp. 4614–4621

    2018

  12. [12]

    Fast-livo2: Fast, direct lidar–inertial–visual odometry,

    C. Zheng, W. Xu, Z. Zou, T. Hua, C. Yuan, D. He, B. Zhou, Z. Liu, J. Lin, F. Zhu, Y . Ren, R. Wang, F. Meng, and F. Zhang, “Fast-livo2: Fast, direct lidar–inertial–visual odometry,”IEEE Transactions on Robotics, vol. 41, pp. 326–346, 2025

    2025

  13. [13]

    Rnin- vio: Robust neural inertial navigation aided visual-inertial odometry in challenging scenes,

    D. Chen, N. Wang, R. Xu, W. Xie, H. Bao, and G. Zhang, “Rnin- vio: Robust neural inertial navigation aided visual-inertial odometry in challenging scenes,” in2021 IEEE International Symposium on Mixed and Augmented Reality (ISMAR), 2021, pp. 275–283

    2021

  14. [14]

    A multi-state constraint kalman filter for vision-aided inertial navigation,

    A. I. Mourikis and S. I. Roumeliotis, “A multi-state constraint kalman filter for vision-aided inertial navigation,” inProceedings of the 2007 IEEE International Conference on Robotics and Automation, 2007, pp. 3565–3572

    2007

  15. [15]

    Ridi: Robust imu double integra- tion,

    H. Yan, Q. Shan, and Y . Furukawa, “Ridi: Robust imu double integra- tion,” inECCV, V . Ferrari, M. Hebert, C. Sminchisescu, and Y . Weiss, Eds. Cham: Springer International Publishing, 2018, pp. 641–656

    2018

  16. [16]

    Pdrnet: A deep-learning pedestrian dead reckoning framework,

    O. Asraf, F. Shama, and I. Klein, “Pdrnet: A deep-learning pedestrian dead reckoning framework,”IEEE Sensors Journal, vol. 22, no. 6, pp. 4932–4939, 2022

    2022

  17. [17]

    Tlio: Tight learned inertial odometry,

    W. Liu, D. Caruso, E. Ilg, J. Dong, A. I. Mourikis, K. Daniilidis, V . Kumar, and J. Engel, “Tlio: Tight learned inertial odometry,”IEEE Robotics and Automation Letters, vol. 5, no. 4, pp. 5653–5660, 2020

    2020

  18. [18]

    Enhancing vio robustness under sudden lighting variation: A learning- based imu dead-reckoning for uav localization,

    D. Yang, H. Liu, X. Jin, J. Chen, C. Wang, X. Ding, and K. Xu, “Enhancing vio robustness under sudden lighting variation: A learning- based imu dead-reckoning for uav localization,”IEEE Robotics and Automation Letters, vol. 9, no. 5, pp. 4535–4542, 2024

    2024

  19. [19]

    Imunet: Efficient regression architecture for inertial imu navigation and positioning,

    B. Zeinali, H. Zanddizari, and M. J. Chang, “Imunet: Efficient regression architecture for inertial imu navigation and positioning,”IEEE Transac- tions on Instrumentation and Measurement, vol. 73, no. 2516213, 2024

    2024

  20. [20]

    Deepils: Towards accu- rate domain invariant aiot-enabled inertial localization system,

    O. Tariq, B. Dastagir, M. Bilal, and D. Han, “Deepils: Towards accu- rate domain invariant aiot-enabled inertial localization system,”IEEE Internet of Things Journal, pp. 1–1, 2025

    2025

  21. [21]

    Deep learning for inertial positioning: A survey,

    C. Chen and X. Pan, “Deep learning for inertial positioning: A survey,” IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 9, pp. 10 506–10 523, 2024

    2024

  22. [22]

    Ctin: Robust contextual transformer network for inertial navigation,

    B. Rao, E. Kazemi, Y . Ding, D. M. Shila, F. M. Tucker, and L. Wang, “Ctin: Robust contextual transformer network for inertial navigation,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 36, no. 5, pp. 5413–5421, Jun. 2022. [Online]. Available: https://ojs.aaai.org/index.php/AAAI/article/view/20479

    2022

  23. [23]

    Smartphone-based indoor pedestrian tracking via transformer,

    X. Li, K. Li, J. Liu, and R. Gao, “Smartphone-based indoor pedestrian tracking via transformer,” in2024 27th International Conference on Computer Supported Cooperative Work in Design (CSCWD), 2024, pp. 1280–1285

    2024

  24. [24]

    iMoT: Inertial Motion Transformer for Inertial Navigation,

    S. M. Nguyen, L. D. Tran, D. Viet Le, and P. J. M. Havinga, “iMoT: Inertial Motion Transformer for Inertial Navigation,”arXiv e-prints, p. arXiv:2412.12190, Dec. 2024

    Pith/arXiv arXiv 2024

  25. [25]

    Spatiotemporal co-attention hybrid neural network for pedestrian localization based on 6d imu,

    Y . Wang, H. Cheng, and M. Q.-H. Meng, “Spatiotemporal co-attention hybrid neural network for pedestrian localization based on 6d imu,” IEEE Transactions on Automation Science and Engineering, vol. 20, no. 1, pp. 636–648, 2023

    2023

  26. [26]

    From imu measurement sequence to velocity estimate sequence: An effective and efficient data-driven inertial odometry approach,

    Y . Wang, H. Cheng, A. Zhang, and M. Q.-H. Meng, “From imu measurement sequence to velocity estimate sequence: An effective and efficient data-driven inertial odometry approach,”IEEE Sensors Journal, vol. 23, no. 15, pp. 17 117–17 126, 2023

    2023

  27. [27]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inFirst Conference on Language Modeling, 2024. [Online]. Available: https://openreview.net/forum?id=tEYskw1VY2

    2024

  28. [28]

    Nanomst: A hardware-aware multiscale trans- former network for tinyml-based real-time inertial motion tracking,

    O. Tariq and D. Han, “Nanomst: A hardware-aware multiscale trans- former network for tinyml-based real-time inertial motion tracking,” IEEE Internet of Things Journal, vol. 12, no. 18, pp. 37 763–37 776, 2025

    2025

  29. [29]

    Repvit: Revisiting mobile cnn from vit perspective,

    A. Wang, H. Chen, Z. Lin, J. Han, and G. Ding, “Repvit: Revisiting mobile cnn from vit perspective,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 909–15 920

    2024

  30. [30]

    MobileNetV4 – Universal Models for the Mobile Ecosys- tem,

    D. Qin, C. Leichner, M. Delakis, M. Fornoni, S. Luo, F. Yang, W. Wang, C. Banbury, C. Ye, B. Akin, V . Aggarwal, T. Zhu, D. Moro, and A. Howard, “MobileNetV4 – Universal Models for the Mobile Ecosys- tem,”arXiv e-prints, p. arXiv:2404.10518, Apr. 2024

    Pith/arXiv arXiv 2024

  31. [31]

    Oxiod: The dataset for deep inertial odometry,

    C. Chen, P. Zhao, C. X. Lu, W. Wang, A. Markham, and N. Trigoni, “Oxiod: The dataset for deep inertial odometry,”arXiv preprint arXiv:1809.07491, 2018

    Pith/arXiv arXiv 2018

  32. [32]

    Ionet: Learning to cure the curse of drift in inertial odometry,

    C. Chen, X. Lu, A. Markham, and N. Trigoni, “Ionet: Learning to cure the curse of drift in inertial odometry,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 32, no. 1, 2018

    2018