Pith. sign in

REVIEW 3 major objections 4 minor 41 references

EgoCHARM: Resource-Efficient Hierarchical Activity Recognition using an Egocentric IMU Sensor

T0 review · 3 major / 4 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read A hierarchical network using one head-mounted IMU recognizes nine long activities at 0.826 F1 and, by freezing its 22k-parameter motion encoder, three brief motions at 0.855 F1.

desk verdict Useful CHARM-on-the-head extension with a compact architecture, but the headline F1 scores are selection maxima because the hyperparameter search optimizes directly on the test set. read the letter →

arxiv 2504.17735 v1 pith:WBCYXQV4 submitted 2025-04-24 cs.CV cs.LG

Akhil Padmanabha , Saravanan Govindarajan , Hwanmun Kim , Sergio Ortiz , Rahul Rajan , Doruk Senkal , Sneha Kadetotad This is my paper · ORCID
classification cs.CVcs.LG
keywords humanactivityrecognitionegocentricinertialmeasurementunitsmartglasseshierarchicalmodelsemi-supervisedlearningmotionembeddingswearablesEgoCHARM
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

EgoCHARM claims that a single head-mounted inertial measurement unit, paired with a hierarchical network of about 22k low-level and 63k high-level parameters, can recognize nine long everyday activities at 0.826 test F1 and three brief motions at 0.855 test F1. The training scheme deliberately needs only high-level labels: the low-level encoder learns motion embeddings through the high-level task, and a frozen copy plus a 99-parameter probe separates stationary, walking, and running. The point is to make always-on activity recognition on smartglasses cheap enough to run on the IMU chip itself, saving power and memory on the main processor. If true, smartglasses could track daily routines and feed context-aware assistants from motion alone, without continuous camera processing.

What carries the argument

The carrying mechanism is the low-level motion encoder: a CNN-GRU that consumes six raw IMU channels (3-axis accelerometer, 3-axis gyroscope) at 50 Hz in 1-second windows and emits a 32-dimensional embedding per second. Its 1-D CNN layers run several parallel kernels with different dilation rates so the same small network can catch both fast and slow periodic patterns in head motion; the GRU then folds the temporal sequence into a single embedding. During training only high-level labels supervise the stack, so the encoder is forced to preserve motion distinctions useful over 30-second horizons. At inference, the same frozen encoder can be paired with a single tunable probing layer, which is what makes the 0.855 F1 result a test of embedding generalizability rather than of a separately trained classifier.

What would settle it

Pick a random sample of the reannotated low-level windows, watch the synchronized Aria video frame-by-frame, and assign a per-second label; if the video-derived labels disagree with the assigned uniform label in a nontrivial fraction of seconds, the reported 0.855 probing F1 is an upper bound on real per-second recognition.

Watch

Extended reading notes

Core claim

On the paper's own terms, EgoCHARM establishes that one head-mounted IMU is enough to recognize both coarse and fine activities with a model of only tens of thousands of parameters. The system splits each 30-second high-level window into 1-second windows, encodes each second into a 32-dimensional motion embedding with a CNN-GRU using variable-dilation convolutions, and feeds the sequence of embeddings into a high-level GRU. Training uses only high-level labels, yet the frozen encoder, probed by one 99-parameter layer, separates stationary, walking, and running with 0.855 F1 (90.64% accuracy), and the full model classifies nine high-level activities with 0.826 F1 (82.86% accuracy). The reported parameter counts are 21,868 for the low-level encoder and 63,369 for the high-level head, with about 1.04M and 1.86M FLOPs respectively. Sensitivity experiments further claim that 15 Hz sampling, 20-second windows, and a few hundred training samples per class each preserve most of the performance.

Load-bearing premise

The per-second low-level labels are inferred from 30-second text summaries by assuming each chosen segment is uniformly stationary, walking, or running; if a segment hides brief stops or other motions, the probing F1 overstates true per-second accuracy.

Editorial extensions

If this is right

  • The low-level encoder fits within the 25k-parameter budget of current IMU chips with on-chip compute, so the frequent 1-second inference can run on the sensor itself while the main processor handles only the 30-second high-level pass.
  • Dropping sampling to 15 Hz keeps high-level F1 at 0.807 and cuts low-level encoder FLOPs roughly 3.5x, so an always-on system can be made substantially cheaper with minimal accuracy loss.
  • A 20-second high-level window reaches 0.808 F1, close to the 30-second peak of 0.826, so high-level labels can be produced with lower latency than the training window.
  • Only about 500 high-level samples per class are enough for 0.748 F1, suggesting that useful egocentric recognizers can be trained with far less data than current large-scale datasets.
  • The frozen encoder's PCA clusters already separate stationary, walking, and running without seeing their labels, so the same embeddings are plausible inputs for other on-device tasks such as context for an assistant or triggering higher-power sensors.

Reading between the lines

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

  • An extension the paper leaves implicit: push the frozen-encoder probe to low-level classes beyond stationary, walking, and running; the PCA separation suggests the embeddings carry general motion structure, but only three probe classes are tested.
  • The FLOP counts are estimates, not measured power; the decisive resource test is deployment on a real IMU chip with on-chip compute at 15 Hz, measuring current draw and latency.
  • The mechanism is not fundamentally head-specific, so the same high-level-label-only recipe could be tested on wrist- or body-worn IMUs; success would extend the sample-efficiency claim beyond egocentric sensing.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces EgoCHARM, a hierarchical activity-recognition architecture that uses a single head-mounted IMU. A small CNN-GRU encoder processes 1-second windows of 6-channel IMU data and outputs motion embeddings; these embeddings are aggregated over a 30-second window and passed to a GRU high-level classifier that recognizes 9 long activities. The encoder is trained with only high-level labels, then frozen and probed with a linear layer to recognize 3 low-level activities (stationary, walking, running). Data are taken from Ego-Exo4D and Nymeria, with participant-disjoint train/test stratification. The paper reports a test F1 of 0.826 for high-level recognition and 0.855 for low-level probing, with 21,868 low-level and 63,369 high-level parameters, plus a sensitivity analysis over sample counts, sampling frequency, and window size.

Significance. The architecture idea is attractive: a sub-25k-parameter low-level encoder is genuinely deployable on current IMU chips, and the semi-supervised training from high-level labels is a practical way to obtain low-level embeddings without dense annotation. The paper is also honest about the limitations of head-mounted IMU signals, and the confusion-matrix discussion is useful. The comparison against several architecture combinations and the sensitivity analysis are valuable empirical contributions, provided the evaluation protocol is sound. However, the central quantitative claims are not currently supported because the hyperparameter search directly optimized the reported test F1, making the headline numbers selection-biased maxima rather than unbiased estimates.

major comments (3)
  1. [Section 3.7] The hyperparameter search is explicitly described as 'optimizing over the test F1 score.' With 200 runs per architecture combination and 15 evaluated combinations, the final high-level model is selected as the best of roughly 3,000 evaluations of the same test set. The reported test F1 of 0.826 and accuracy of 82.86% in Table 2 are therefore maxima of a selection procedure over the test data, not unbiased estimates of generalization. The low-level probing result (0.855) is also affected, because the frozen encoder was chosen using high-level test performance, so the embedding quality is not validated independently. The paper needs a validation split (or nested cross-validation) used for all architecture and hyperparameter choices, with the test set evaluated only once. This is a load-bearing issue for the central claims in the abstract and Section 4.1.
  2. [Section 3.1] The low-level labels are derived by manually choosing 30-second windows from the Nymeria text summaries and assigning that single label to every 1-second window inside the chosen segment, with video refinement only for the running class. If a chosen segment contains brief stops, pauses, or other actions, the per-second labels are wrong. This label noise is not measured or modeled, and without per-second verification the 0.855 low-level probing F1 is not a reliable estimate of true per-second recognition accuracy. Please provide per-second verification (or a second annotator agreement study, or an evaluation that removes boundary seconds from each segment) before treating the low-level result as established.
  3. [Section 4.2] The sensitivity analysis in Fig. 6 and Tables S1–S4 uses the same 200-run protocol with test-F1 optimization described in Section 3.7. The comparisons across sample counts, sampling frequencies, and window sizes are therefore also subject to the same selection bias. The claim in Section 5 that a 15 Hz sampling frequency yields performance comparable to 50 Hz is based on these selection-biased numbers. The resource-efficiency conclusions require an unbiased validation protocol before they can be accepted.
minor comments (4)
  1. [Section 4.1] In the sentence 'In line with our PCA visualization, Fig 3.5', the reference should be to Figure 5 (the PCA plot), not 'Fig 3.5'.
  2. [Table 2] The table header 'LL + HL' followed by repeated column labels for F1, Accuracy, Parameters, and FLOPs is difficult to parse; please restructure with clear subheaders indicating LL, HL, and PL columns.
  3. [Section 3.3] The manuscript inconsistently uses 'hand picked' and 'hand-picked'; please unify the spelling.
  4. [Abstract] After the evaluation protocol is corrected, the reported test F1 values in the abstract, introduction, and sensitivity tables will need to be updated to reflect the unbiased results.

Circularity Check

1 steps flagged · score 6.0 of 10

Hyperparameters are optimized on the test F1, so the headline 0.826/0.855 scores are selection-maximized values rather than unbiased test estimates.

  1. fitted input called prediction [Section 3.7 (Experiments and Hyperparameter Searches); Table 2 note]
    "For each architecture combination, we conduct 200 runs and use Sobol sampling [2] in addition to a Bayesian optimization method, SAASBO [10], to select parameters for each run, optimizing over the test F1 score. After selecting the best performing model for every architecture combination, we evaluate the frozen low level encoder’s effectiveness at low level activity recognition using our low level activity dataset and probing layer detailed in Section 3.6."

    The paper's headline high-level number is the direct objective of its own hyperparameter optimization: the same test set is evaluated roughly 3,000 times across 15 architecture combinations (200 runs each), and the reported 0.826 test F1 is the best result of that selection, not an unbiased estimate of generalization. Equivalently, the test F1 is used as a selection signal and then reported as if it were an untouched test measurement. The low-level 0.855 F1 is reported for an encoder and architecture chosen by the same test-F1 optimization, so although the probing layer is cross-validated, the reported low-level number inherits the same architecture-selection bias. A held-out validation split or nested cross-validation would be required to support the claimed test F1 scores.

full rationale

Aside from the model-selection step, the paper's derivation chain is empirical and largely self-contained. The low-level encoder is trained only on high-level labels, then frozen and probed; that claim is supported by the reported PCA visualization and probing experiments, neither of which is definitionally tied to the low-level labels. The self-citation to CHARM [36], whose second author is also an author of the present paper, is not load-bearing: CHARM is identified as prior inspiration, and the paper evaluates its own architecture against baselines rather than invoking CHARM to justify a uniqueness claim. The remaining concern is not self-citation but the explicit use of the test F1 score as the hyperparameter-search objective in Section 3.7. Because the final model is selected from hundreds of test-set evaluations, the reported test F1 is a fitted quantity; this is the one concrete circular step. The coarse 30-second textual summaries used to derive per-second low-level labels are a label-quality and annotation-leakage concern, but they are not a construction-level circularity in the claimed derivation.

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

The central claim rests on empirical machine-learning design choices: window sizes, sampling rate, embedding dimension, and undisclosed final hyperparameters. The key domain assumptions are the sufficiency of high level labels for low level embedding learning and the accuracy of manual low level reannotation. No new entities are introduced.

free parameters (5)
  • High level window size = 30 s
    Chosen by hand; sensitivity analysis shows performance degrades at smaller windows. Central to the architecture.
  • Low level window size = 1 s
    Chosen to match typical atomic action annotation schemes; determines input dimension R6x50.
  • IMU sampling frequency = 50 Hz
    Chosen using Nyquist rationale in Section 3.2; sensitivity analysis shows 15 Hz performs similarly, so 50 Hz is not critical.
  • Output embedding dimension = 32
    The low level embedding dimension is 32, seen in Section 3.5 (PCA); final value from hyperparameter search not disclosed.
  • Final architecture hyperparameters (CNN dilation, channels, layers, GRU hidden size) = Not disclosed
    Selected by a 200-run Sobol/SAASBO search optimizing test F1; exact values absent from paper and supplement.
assumptions (5)
  • domain assumption Head-mounted IMU signals sampled at 50 Hz capture human movement frequencies up to 25 Hz (Nyquist)
    Section 3.2 justifies the sampling rate; if human motion has higher-frequency components, aliasing could affect features, though sensitivity analysis shows 15 Hz works.
  • domain assumption High level activity labels provide sufficient weak supervision to learn discriminative low level motion embeddings
    Core semi-supervised assumption of the paper; if high level classes do not implicitly cover stationary/walking/running motions, the probing result would not hold.
  • domain assumption Manually reannotated low level labels are accurate per 1s window
    Section 3.1: labels are derived from 30s textual summaries and visual refinement; heterogeneous windows would corrupt labels.
  • domain assumption The selected high level scenarios are homogeneous enough to serve as classes
    Section 3.1 excludes scenarios like 'health' because they contain multiple sub-activities; remaining classes may still contain varied motions.
  • domain assumption Train/test split has no participant overlap
    Section 3.1 states stratification ensures no participant appears in both sets; this is a stated design choice and reasonable.

how reviews work

0 comments
Cite this review

Pith. "Pith review of EgoCHARM: Resource-Efficient Hierarchical Activity Recognition using an Egocentric IMU Sensor." pith.science (2026). https://pith.science/paper/WBCYXQV4

@misc{pith2026250417735,
  author       = {Pith},
  title        = {Pith review of: EgoCHARM: Resource-Efficient Hierarchical Activity Recognition using an Egocentric IMU Sensor},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBCYXQV4}},
  note         = {Machine review of arXiv:2504.17735}
}
read the original abstract

Human activity recognition (HAR) on smartglasses has various use cases, including health/fitness tracking and input for context-aware AI assistants. However, current approaches for egocentric activity recognition suffer from low performance or are resource-intensive. In this work, we introduce a resource (memory, compute, power, sample) efficient machine learning algorithm, EgoCHARM, for recognizing both high level and low level activities using a single egocentric (head-mounted) Inertial Measurement Unit (IMU). Our hierarchical algorithm employs a semi-supervised learning strategy, requiring primarily high level activity labels for training, to learn generalizable low level motion embeddings that can be effectively utilized for low level activity recognition. We evaluate our method on 9 high level and 3 low level activities achieving 0.826 and 0.855 F1 scores on high level and low level activity recognition respectively, with just 63k high level and 22k low level model parameters, allowing the low level encoder to be deployed directly on current IMU chips with compute. Lastly, we present results and insights from a sensitivity analysis and highlight the opportunities and limitations of activity recognition using egocentric IMUs.

Figures

Figures reproduced from arXiv: 2504.17735 by the authors.

Figure 1
Figure 1. We propose a resource-efficient hierarchical architecture, EgoCHARM, to classify both high and low level activities [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. EgoCHARM Low Level Encoder Architecture. Our en [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Low Level Encoder Probing. To enable low level activity [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (6 more)
Figure 5
Figure 5. Figure 5: Principal component analysis (PCA) on unseen low level [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 4
Figure 4. Figure 4: High Level Activity Recognition Confusion Matrix using [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 6
Figure 6. Figure 6: EgoCHARM Sensitivity Analysis. For all plots, the dotted grey line represents the best F1 score for our EgoCHARM model, [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Low Level Activity Recognition Confusion Matrix using [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: Egocentric camera frames and IMU Signals are shown from a 2 second clip of the cooking high level activity. In this clip, the [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]
Figure 9
Figure 9. Figure 9: Egocentric camera frames and IMU Signals are shown from a 2 second clip of the bike repair high level activity. In this clip, the [PITH_FULL_IMAGE:figures/full_fig_p008_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

41 extracted references · 36 canonical work pages

  1. [1]

    InCrowd-VI: A Realistic Visual-Inertial Dataset for Evaluating SLAM in Indoor Pedestrian-Rich Spaces for Human Navigation

    M. Bamdad, H.-P. Hutter, and A. Darvishy. Incrowd-vi: A realis- tic visual-inertial dataset for evaluating slam in indoor pedestrian-rich spaces for human navigation. arXiv preprint arXiv:2411.14358, 2024. 2

    work page Pith review arXiv 2024

  2. [2]

    Bergstra and Y

    J. Bergstra and Y . Bengio. Random search for hyper-parameter opti- mization. The journal of machine learning research, 13(1):281–305,

    work page

  3. [3]

    Bmi263: Inertial measurement unit (imu)

    Bosch Sensortec. Bmi263: Inertial measurement unit (imu). https://www.bosch-sensortec.com/products/ motion-sensors/imus/bmi263/, 2023. 2

    work page 2023

  4. [4]

    Bhi360: Inertial measurement unit (imu)

    Bosch Sensortec. Bhi360: Inertial measurement unit (imu). https://www.bosch-sensortec.com/products/ smart-sensor-systems/bhi360/, 2024. 1, 4

    work page 2024

  5. [5]

    B. Chen, W. Wongso, Z. Li, Y . Khaokaew, H. Xue, and F. Salim. Co- modo: Cross-modal video-to-imu distillation for efficient egocentric human activity recognition. arXiv preprint arXiv:2503.07259, 2025. 2

    arXiv 2025

  6. [6]

    K. Chen, D. Zhang, L. Yao, B. Guo, Z. Yu, and Y . Liu. Deep learning for sensor-based human activity recognition: Overview, challenges, and opportunities. ACM Computing Surveys (CSUR) , 54(4):1–40,

    work page

  7. [7]

    Cristiano, A

    A. Cristiano, A. Sanna, and D. Trojaniello. Daily physical activ- ity classification using a head-mounted device. In 2019 IEEE In- ternational Conference on Engineering, Technology and Innovation (ICE/ITMC), pp. 1–7. IEEE, 2019. 2

    work page 2019

  8. [8]

    A. M. Das, C. I. Tang, F. Kawsar, and M. Malekzadeh. Primus: Pretraining imu encoders with multimodal self-supervision. arXiv preprint arXiv:2411.15127, 2024. 2, 4, 7

    work page arXiv 2024

Show all 41 references
  1. [9]

    Engel, K

    J. Engel, K. Somasundaram, M. Goesele, A. Sun, A. Gamino, A. Turner, A. Talattof, A. Yuan, B. Souti, B. Meredith, et al. Project aria: A new tool for egocentric multi-modal ai research.arXiv preprint arXiv:2308.13561, 2023. 1, 3

    2023 arXiv

  2. [10]

    Eriksson and M

    D. Eriksson and M. Jankowiak. High-dimensional bayesian optimiza- tion with sparse axis-aligned subspaces. In Uncertainty in Artificial Intelligence, pp. 493–503. PMLR, 2021. 4

    2021

  3. [11]

    L. Fan, Z. Wang, and H. Wang. Human activity recognition model based on decision tree. In2013 International Conference on Advanced Cloud and Big Data, pp. 64–68. IEEE, 2013. 2

    2013

  4. [12]

    Gjoreski, I

    H. Gjoreski, I. Kiprijanovska, S. Stankoski, S. Kalabakov, J. Brouli- dakis, C. Nduka, and M. Gjoreski. Head-ar: Human activity recogni- tion with head-mounted imu using weighted ensemble learning. Ac- tivity and Behavior Computing, pp. 153–167, 2021. 2

    2021

  5. [13]

    Grauman, A

    K. Grauman, A. Westbury, E. Byrne, Z. Chavis, A. Furnari, R. Gird- har, J. Hamburger, H. Jiang, M. Liu, X. Liu, et al. Ego4d: Around the world in 3,000 hours of egocentric video. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pp. 18995–1...

    2022

  6. [14]

    Grauman, A

    K. Grauman, A. Westbury, L. Torresani, K. Kitani, J. Malik, T. Afouras, K. Ashutosh, V . Baiyya, S. Bansal, B. Boote, et al. Ego- exo4d: Understanding skilled human activity from first-and third- person perspectives. In Proceedings of the IEEE/CVF Conference on Computer Vision...

    2024

  7. [15]

    Ionut-Cristian and D

    S. Ionut-Cristian and D. Dan-Marius. Using inertial sensors to deter- mine head motion—a review. Journal of Imaging, 7(12):265, 2021. 2

    2021

  8. [16]

    L. G. Jaimes, Y . De La Hoz, C. Eggert, and I. J. Vergara-Laurens. Pat: A power-aware decision tree algorithm for mobile activity recogni- tion. In 2016 13th IEEE Annual Consumer Communications & Net- working Conference (CCNC), pp. 54–59. IEEE, 2016. 2

    2016

  9. [17]

    S.-R. Ke, H. L. U. Thuc, Y .-J. Lee, J.-N. Hwang, J.-H. Yoo, and K.-H. Choi. A review on video-based human activity recognition. Comput- ers, 2(2):88–131, 2013. 2

    2013

  10. [18]

    Kumar, S

    A. Kumar, S. Pundlik, E. Peli, and G. Luo. Comparison of visual slam and imu in tracking head movement outdoors.Behavior research methods, 55(6):2787–2799, 2023. 2

    2023

  11. [19]

    O. D. Lara and M. A. Labrador. A survey on human activity recogni- tion using wearable sensors. IEEE communications surveys & tutori- als, 15(3):1192–1209, 2012. 2

    2012

  12. [20]

    Lemieux and R

    N. Lemieux and R. Noumeir. A hierarchical learning approach for human action recognition. Sensors, 20(17):4946, 2020. 2

    2020

  13. [21]

    Li and H

    H. Li and H. Hu. Head gesture recognition combining activity detec- tion and dynamic time warping. Journal of Imaging, 10(5):123, 2024. 2

    2024

  14. [22]

    Z. Li, S. Deldari, L. Chen, H. Xue, and F. D. Salim. Sensorllm: Align- ing large language models with motion sensors for human activity recognition. arXiv preprint arXiv:2410.10624, 2024. 2

    2024 arXiv

  15. [23]

    M. Liu, L. Ma, K. Somasundaram, Y . Li, K. Grauman, J. M. Rehg, and C. Li. Egocentric activity recognition and localization on a 3d map. In European Conference on Computer Vision, pp. 621–638. Springer,

  16. [24]

    Z. Lv, N. Charron, P. Moulon, A. Gamino, C. Peng, C. Sweeney, E. Miller, H. Tang, J. Meissner, J. Dong, et al. Aria everyday activities dataset. arXiv preprint arXiv:2402.13349, 2024. 2

    2024 arXiv

  17. [25]

    L. Ma, Y . Ye, F. Hong, V . Guzov, Y . Jiang, R. Postyeni, L. Pesqueira, A. Gamino, V . Baiyya, H. J. Kim, et al. Nymeria: A massive collec- tion of multimodal egocentric daily motion in the wild. In European Conference on Computer Vision, pp. 445–465. Springer, 2024. 3

    2024

  18. [26]

    M. Ma, H. Fan, and K. M. Kitani. Going deeper into first-person activ- ity recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pp. 1894–1903, 2016. 1

    1903

  19. [27]

    Project aria, 2025

    Meta. Project aria, 2025. https://www.projectaria.com/. 1, 3

    2025

  20. [28]

    S. Moon, A. Madotto, Z. Lin, A. Dirafzoon, A. Saraf, A. Bearman, and B. Damavandi. Imu2clip: Multimodal contrastive learning for imu motion sensors from egocentric videos and text. arXiv preprint arXiv:2210.14395, 2022. 2, 4, 7

    2022 arXiv

  21. [29]

    S. Moon, A. Madotto, Z. Lin, T. Nagarajan, M. Smith, S. Jain, C.-F. Yeh, P. Murugesan, P. Heidari, Y . Liu, et al. Anymal: An efficient and scalable any-modality augmented language model. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing...

    2024

  22. [30]

    Nakabayashi and H

    T. Nakabayashi and H. Saito. Multimodal human activity recogni- tion on edge devices. In 2024 IEEE International Symposium on Mixed and Augmented Reality Adjunct (ISMAR-Adjunct) , pp. 136–

    2024

  23. [31]

    Nguyen, J.-C

    T.-H.-C. Nguyen, J.-C. Nebel, and F. Florez-Revuelta. Recognition of activities of daily living with egocentric vision: A review. Sensors, 16(1):72, 2016. 1

    2016

  24. [32]

    Padmanabha, J

    A. Padmanabha, J. Gupta, C. Chen, J. Yang, V . Nguyen, D. J. Weber, C. Majidi, and Z. Erickson. Independence in the home: A wearable interface for a person with quadriplegia to teleoperate a mobile ma- nipulator. In Proceedings of the 2024 ACM/IEEE International Con- ference o...

    2024

  25. [33]

    Padmanabha, C

    A. Padmanabha, C. Majidi, and Z. Erickson. Towards wearable inter- faces for robotic caregiving. arXiv preprint arXiv:2502.05343, 2025. 2

    2025 arXiv

  26. [34]

    Padmanabha, Q

    A. Padmanabha, Q. Wang, D. Han, J. Diyora, K. Kacker, H. Khalid, L.-J. Chen, C. Majidi, and Z. Erickson. Hat: Head-worn assistive tele- operation of mobile manipulators. In 2023 IEEE International Con- ference on Robotics and Automation (ICRA), pp. 12542–12548. IEEE,

    2023

  27. [35]

    D. Ravi, C. Wong, B. Lo, and G.-Z. Yang. Deep learning for hu- man activity recognition: A resource efficient implementation on low- power devices. In 2016 IEEE 13th international conference on wear- able and implantable body sensor networks (BSN) , pp. 71–76. IEEE,

    2016

  28. [36]

    Rosen and D

    E. Rosen and D. Senkal. Charm: A hierarchical deep learning model for classification of complex human activities using motion sensors. arXiv preprint arXiv:2207.07806, 2022. 2, 3, 4

    2022 arXiv

  29. [37]

    I. C. Severin, D. M. Dobrea, and M. C. Dobrea. Head gesture recog- nition using a 6dof inertial imu. International Journal of Computers Communications & Control, 15(3), 2020. 2

    2020

  30. [38]

    Y . Song, E. Byrne, T. Nagarajan, H. Wang, M. Martin, and L. Torre- sani. Ego4d goal-step: Toward hierarchical understanding of proce- dural activities. Advances in Neural Information Processing Systems, 36:38863–38886, 2023. 1

    2023

  31. [39]

    Lsm6dso16is

    STMicroelectronics. Lsm6dso16is. https://www.st.com/en/ mems-and-sensors/lsm6dso16is.html , 2025. 1, 4

    2025

  32. [40]

    Vrigkas, C

    M. Vrigkas, C. Nikou, and I. A. Kakadiaris. A review of human ac- tivity recognition methods. Frontiers in Robotics and AI, 2:28, 2015. 2

    2015

  33. [41]

    H. Yu, G. Pan, M. Pan, C. Li, W. Jia, L. Zhang, and M. Sun. A hierarchical deep fusion framework for egocentric activity recognition using a wearable hybrid sensor system. Sensors, 19(3):546, 2019. 2 SUPPLEMENTAL MATERIALS S1 S AMPLES PER ACTIVITY CLASS In Fig. S1 and Fig. S2,...

    2019

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.