Pith. sign in

REVIEW 3 major objections 4 minor 42 references

Human Fall Detection using Transfer Learning-based 3D CNN

T0 review · 3 major / 4 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 vision-based fall detector can work without training the video model at all: features from a 3D CNN pre-trained on Sports-1M, fed to a linear SVM, classify falls versus activities of daily living with average accuracy around 98% on one…

desk verdict Standard C3D+SVM transfer-learning paper whose headline metrics contradict each other and whose split protocol risks data leakage; needs a corrected evaluation before the claims are credible. read the letter →

arxiv 2506.03193 v1 pith:UVJYGNXP submitted 2025-05-31 cs.CV cs.LG

Ekram Alam , Abu Sufian , Paramartha Dutta , Marco Leo This is my paper
classification cs.CVcs.LG
keywords falldetection3DCNNtransferlearningC3DSports-1Msupportvectormachinespatio-temporalfeatureshumandatasets
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

This paper tries to establish that transfer learning from a generic large-scale action-recognition model is enough for human fall detection: freeze the C3D network pre-trained on Sports-1M, extract spatio-temporal features from 16-frame video chunks, and train only a linear SVM to separate falls from activities of daily living. The authors report average accuracy 97.8% on GMDCSA and 93.67% on CAUCAFall over five stratified shuffle splits, with zero false negatives on GMDCSA. The point of the design is practicality: 3D CNNs are strong for video but expensive to train, and freezing the backbone removes that cost while keeping the temporal information that 2D frame-by-frame methods miss. If the result holds, it means inexpensive vision-based fall monitoring for elderly or isolated people can be built on existing video models without action-specific deep training.

What carries the argument

The mechanism is the frozen C3D feature extractor plus the linear SVM. C3D is a 3D CNN with eight 3x3x3 convolution layers, five 3D pooling layers, and two fully connected layers; here the first fully connected layer (fc6) is used as the feature representation of a 16-frame, 112x112 chunk. Pre-training on Sports-1M supplies generic motion features, and the SVM replaces the softmax head, so the video model is never fine-tuned on fall data. All spatio-temporal feature extraction is done by the frozen network, and the SVM learns the decision boundary between fall and ADL.

What would settle it

Re-run the same pipeline with leave-one-subject-out or video-level splits, ensuring that no chunk from a video used for training appears in testing; if accuracy drops substantially toward chance on CAUCAFall, the reported averages do not reflect generalization to unseen people.

Watch

Extended reading notes

Core claim

The central claim is that the original learned weights of the C3D 3D CNN, pre-trained on Sports-1M, already encode spatio-temporal features discriminative enough for fall detection; only the final classifier needs to learn the fall/ADL boundary. The authors modify C3D by removing the last fully connected layer and softmax and attaching a linear SVM to the fc6 features, then evaluate on two recent datasets. Average sensitivity, specificity, precision, accuracy, and F1 across five stratified shuffle splits are 96.26, 100, 100, 97.8, and 98.08 for GMDCSA, and 94.55, 92.72, 93.42, 93.67, and 93.95 for CAUCAFall. They interpret the GMDCSA false positives as confusions caused by sleeping ADL activities that resemble falls.

Load-bearing premise

The evaluation splits 16-frame chunks without stating that chunks from the same source video are kept entirely in training or entirely in testing; if they are not, near-identical frames from one video can appear in both sets and inflate the reported accuracy.

Editorial extensions

If this is right

  • If the method is correct, fall detection can be deployed without retraining a 3D CNN, substantially reducing training time and hardware requirements.
  • The system can run on continuous video from existing cameras in homes, hospitals, and nursing homes, since it needs only a frozen feature extractor plus a light classifier.
  • Because temporal structure is captured by the 3D convolutions, the method addresses a limitation of frame-by-frame 2D CNN approaches that miss motion patterns.
  • The GMDCSA false positives show that sleeping ADL sequences are the main confusion source, identifying a concrete target for future fall-detection datasets and models.
  • The reported averages on CAUCAFall, a dataset with multiple subjects, lighting changes, and occlusions, suggest the approach can transfer beyond a single person or camera setup.

Reading between the lines

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

  • A natural test not run in the paper is subject-disjoint splitting: train on some people's videos and test on unseen people; because the shuffle splits are not guaranteed to keep the same source video's chunks together, the reported averages may overstate generalization to new individuals.
  • Until the split strategy is clarified, the accuracy numbers are best treated as upper bounds; if chunk-level leakage exists, near-duplicate frames from one video could appear in both training and testing.
  • The same frozen-extractor-plus-linear-classifier recipe could be applied to other small-label surveillance tasks, such as detecting aggressive behavior or medical episodes, whenever a generic action model is already available.
  • Re-measuring the pipeline with video-level or leave-one-subject-out splits would give a more practical estimate of how the method performs when deployed on a person the model has never seen.
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 / 4 minor

Summary. The paper proposes a vision-based fall detection system that uses a C3D 3D CNN pre-trained on Sports-1M as a frozen feature extractor, followed by a linear SVM classifier trained on the extracted fc6 features. The method is evaluated on two public datasets, GMDCSA and CAUCAFall, using a stratified shuffle split into five 70/30 train/test splits. The authors report high average sensitivity, specificity, precision, accuracy, and F1 scores on both datasets, and provide a link to source code. The central claim is that transfer learning from a pre-trained C3D model, with only the SVM trained, yields strong fall-detection performance while saving training time.

Significance. If the reported results are reliable, the paper offers a simple and computationally inexpensive baseline for vision-based fall detection: frozen C3D features plus a linear SVM. The public code link and use of standard, externally pre-trained weights are strengths that support reproducibility. However, the evaluation protocol has a potentially serious data-leakage issue, and the GMDCSA metrics in Table 2 are internally inconsistent. These problems currently prevent the quantitative claims from being accepted as evidence of generalization to unseen falls or subjects.

major comments (3)
  1. [Section 4, Table 2] The GMDCSA results in Table 2 are internally inconsistent. The FNR row reports 0 for all five splits, yet sensitivity is below 100 for splits 2 through 5 (e.g., 96.83 for split 2); since FNR = 1 - sensitivity in a binary setting, these cannot both hold. Additionally, the FPR row reports positive values (4.17, 8.33, 6.25, 6.25) while specificity is listed as 100 for every split, which is arithmetically impossible. The prose in Section 4 also states that splits 2-5 contain 2, 4, 3, and 3 false positives, which is incompatible with specificity = 100 and precision = 100. These contradictions mean Table 2 cannot be interpreted as a reliable record of the experiments; the authors should recompute all metrics from the confusion matrices and correct the table.
  2. [Section 3.3 and Figure 4] The evaluation split protocol is underspecified and potentially leaky. The method divides each input video into 16-frame chunks and then applies a stratified shuffle split at a 70/30 ratio, but the paper never states whether chunks originating from the same original video are kept entirely within the training set or the test set. Consecutive 16-frame chunks from a 30 fps video are near-duplicates, so a chunk-level split can place almost identical clips in both training and testing, allowing the SVM to memorize video-specific appearance and motion rather than learn a general fall/ADL decision boundary. This concern is especially acute for GMDCSA, which contains only a single subject and short videos. The authors must specify the grouping unit used in the split; if the split was performed at the chunk level, the experiments should be rerun with video-level grouping and the results reported.
  3. [Section 3.1 and Section 5] The paper does not discuss the subject-dependence of the evaluation. GMDCSA was generated by a single person, so even a video-level split would not test generalization to unseen individuals; the model would be evaluated on the same subject it was trained on. For CAUCAFall, the same subjects appear in both fall and ADL classes, and a chunk-level split can mix clips of the same subject across folds, letting the classifier exploit subject identity. Section 5 acknowledges that the data come from healthy individuals rather than seniors, but it does not address the more immediate limitation that the reported numbers cannot be interpreted as generalization to unseen subjects. The authors should state this limitation and, if possible, add a subject-independent evaluation.
minor comments (4)
  1. [Section 3.3] There is a typo in the dataset name: 'CACUCAFall' should be 'CAUCAFall'.
  2. [Throughout] The name of the pretraining dataset is written inconsistently as 'Sports-1M' in the abstract and Section 3.2, and 'Sport-1M' in Section 3.3; please standardize.
  3. [Section 4] The sentence 'The GMDCSA dataset contains many sleeping activities as ADL' would be clearer as 'The GMDCSA dataset contains many sleeping activities among its ADL videos'.
  4. [Section 3.3] The authors state that a stratified shuffle split into five splits with 70% training and 30% testing was used, but they do not report the random seed or the exact split indices; providing this information or relying on the version-controlled code would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: frozen C3D features and a trained SVM are evaluated on held-out splits, with the evaluation not reducing to fitted parameters or self-cited theorems.

full rationale

The paper's central claim is that a C3D model pretrained on Sports-1M, used as a frozen feature extractor, plus a linear SVM trained only on the training portion of each stratified shuffle split, yields the reported fall/ADL classification metrics. This derivation chain is self-contained in the relevant sense: the feature extractor weights are external (Sports-1M via Tran et al.), the SVM is trained on held-out training data, and the reported sensitivity, specificity, precision, accuracy, and F1 values come from confusion matrices on the 30% test portions, not from the construction of the model. No equation or parameter in the paper is defined in terms of the target result, and no fitted value is renamed as a prediction. The GMDCSA dataset is introduced in the authors' own in-press work [15], and [4] and [42] are self-citations for background and metric definitions, but these citations are not load-bearing for the empirical result; the CAUCAFall dataset and the Sports-1M pretrained C3D are external sources. The reader-identified concern about chunk-level versus video-level splitting is a potential data-leakage and generalization-validity issue, not a circularity: it does not make the prediction equivalent to the input by construction. Such evaluation-protocol concerns belong under correctness risk rather than circularity analysis, so no circular step is identified.

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

The paper introduces no invented entities and no fitted free parameters beyond the standard trained SVM weights. The central empirical claim rests on the transferability of pre-trained C3D features, on a leakage-free data split, and on dataset representativeness, all of which are domain assumptions rather than demonstrated facts.

assumptions (3)
  • domain assumption Sports-1M pre-trained C3D features transfer to the fall-detection domain without fine-tuning.
    The method freezes C3D weights and relies on features learned from sports videos to separate falls from ADL; the paper provides no theoretical or empirical justification beyond its own results.
  • domain assumption The stratified shuffle split prevents frame-level leakage between training and test sets.
    Section 3.3 describes a 70/30 five-split shuffle but does not state whether the split unit is the original video or the 16-frame chunk; the validity of the reported metrics depends on this.
  • domain assumption GMDCSA and CAUCAFall are representative of real-world fall scenarios for elderly people.
    The authors note in the conclusion that the datasets were collected from healthy individuals, so applicability to actual seniors is an unverified assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Human Fall Detection using Transfer Learning-based 3D CNN." pith.science (2026). https://pith.science/paper/UVJYGNXP

@misc{pith2026250603193,
  author       = {Pith},
  title        = {Pith review of: Human Fall Detection using Transfer Learning-based 3D CNN},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/UVJYGNXP}},
  note         = {Machine review of arXiv:2506.03193}
}
read the original abstract

Unintentional or accidental falls are one of the significant health issues in senior persons. The population of senior persons is increasing steadily. So, there is a need for an automated fall detection monitoring system. This paper introduces a vision-based fall detection system using a pre-trained 3D CNN. Unlike 2D CNN, 3D CNN extracts not only spatial but also temporal features. The proposed model leverages the original learned weights of a 3D CNN model pre-trained on the Sports1M dataset to extract the spatio-temporal features. Only the SVM classifier was trained, which saves the time required to train the 3D CNN. Stratified shuffle five split cross-validation has been used to split the dataset into training and testing data. Extracted features from the proposed 3D CNN model were fed to an SVM classifier to classify the activity as fall or ADL. Two datasets, GMDCSA and CAUCAFall, were utilized to conduct the experiment. The source code for this work can be accessed via the following link: https://github.com/ekramalam/HFD_3DCNN.

Figures

Figures reproduced from arXiv: 2506.03193 by the authors.

Figure 1
Figure 1. 3D Convolution Mechanism 3D convolution, 3D pooling works. 3D Pooling layer downsamples input 3D data. 3D CNNs use 3D [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the C3D model [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Proposed 3D CNN model architecture 3D CNN model comprises eight “convolutional layers”, five “pooling layers”, one “fully connected” layer [4], and an SVM classifier. The proposed methodology to detect human falls is illustrated in the flow diagram depicted in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Proposed work methodology CAUCAFall datasets, respectively. The GMDCSA dataset’s confusion matrices show no false negative value for any split but some false positives for each split except the first split. The false positive values for split 1, split 2, split 3, split…
Figure 5
Figure 5. Figure 5: Confusion Matrix of the five splits of the GMDCSA and the CAUCAFall Dataset [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

42 extracted references · 42 canonical work pages

  1. [1]

    Ageing and health — who.int,https://www.who.int/news-room/fact-sheets/detail/ageing-and-health, [Accessed 24-08-2023]

    work page 2023

  2. [2]

    Martínez-Villaseñor, H

    L. Martínez-Villaseñor, H. Ponce, J. Brieva, E. Moya-Albor, J. Núñez-Martínez, C. Peñafort-Asturiano, Up- fall detection dataset: A multimodal approach, Sensors 19 (9) (2019) 1988

    work page 2019

  3. [3]

    Paulauskaite-Taraseviciene, J

    A. Paulauskaite-Taraseviciene, J. Siaulys, K. Sutiene, T. Petravicius, S. Navickas, M. Oliandra, A. Rapalis, J. Balciunas, Geriatric care management system powered by the iot and computer vision techniques, in: Healthcare, Vol. 11, MDPI, 2023, p. 1152

    work page 2023

  4. [4]

    E. Alam, A. Sufian, A. K. Das, A. Bhattacharya, M. F. Ali, M. H. Rahman, Leveraging deep learning for computer vision: A review, in: 2021 22nd International Arab Conference on Information Technology (ACIT), IEEE, 2021, pp. 1–8

    work page 2021

  5. [5]

    Ghosh, A

    A. Ghosh, A. Sufian, F. Sultana, A. Chakrabarti, D. De, Fundamental concepts of convolutional neural network, Recent trends and advances in artificial intelligence and Internet of Things (2020) 519–567

    work page 2020

  6. [6]

    Maturana, S

    D. Maturana, S. Scherer, Voxnet: A 3d convolutional neural network for real-time object recognition, in: 2015 IEEE/RSJ international conference on intelligent robots and systems (IROS), IEEE, 2015, pp. 922–928

    work page 2015

  7. [7]

    Y. Liu, D. Jiang, H. Duan, Y. Sun, G. Li, B. Tao, J. Yun, Y. Liu, B. Chen, Dynamic gesture recognition algo- rithm based on 3d convolutional neural network, Computational Intelligence and Neuroscience 2021 (2021)

    work page 2021

  8. [8]

    D. Tran, L. Bourdev, R. Fergus, L. Torresani, M. Paluri, Learning spatiotemporal features with 3d con- volutional networks, in: Proceedings of the IEEE international conference on computer vision, 2015, pp. 4489–4497

    work page 2015

Show all 42 references
  1. [9]

    Vrskova, R

    R. Vrskova, R. Hudec, P. Kamencay, P. Sykora, Human activity classification using the 3dcnn architecture, Applied Sciences 12 (2) (2022) 931

    2022

  2. [10]

    Alanazi, G

    T. Alanazi, G. Muhammad, Human fall detection using 3d multi-stream convolutional neural networks with fusion, Diagnostics 12 (12) (2022) 3060

    2022

  3. [11]

    Sufian, E

    A. Sufian, E. Alam, A. Ghosh, F. Sultana, D. De, M. Dong, Deep learning in computer vision through mobile edge computing for iot, Mobile Edge Computing (2021) 443–471

    2021

  4. [12]

    Karpathy, G

    A. Karpathy, G. Toderici, S. Shetty, T. Leung, R. Sukthankar, L. Fei-Fei, Large-scale video classification with convolutional neural networks, in: Proceedings of the IEEE conference on Computer Vision and Pattern Recognition, 2014, pp. 1725–1732

    2014

  5. [13]

    W. S. Noble, What is a support vector machine?, Nature biotechnology 24 (12) (2006) 1565–1567

    2006

  6. [14]

    Cervantes, F

    J. Cervantes, F. Garcia-Lamont, L. Rodríguez-Mazahua, A. Lopez, A comprehensive survey on support vector machine classification: Applications, challenges and trends, Neurocomputing 408 (2020) 189–215

    2020

  7. [15]

    E. Alam, A. Sufian, P. Dutta, M. Leo, Real-time human fall detection using a lightweight pose estimation technique, in: Proceedings of the Computational Intelligence in Communications and Business Analytics (CICBA-2023) Conference, in press

    2023

  8. [16]

    J. C. E. Guerrero, E. M. España, M. M. Añasco, J. E. P. Lopera, Dataset for human fall recognition in an uncontrolled environment, Data in brief 45 (2022) 108610

    2022

  9. [17]

    Mekruksavanich, P

    S. Mekruksavanich, P. Jantawong, A. Jitpattanakul, Deep learning approaches for har of daily living activities using imu sensors in smart glasses, in: 2023 Joint International Conference on Digital Arts, Media and Tech- nology with ECTI Northern Section Conference on Electrica...

    2023

  10. [18]

    Seenath, M

    S. Seenath, M. Dharmaraj, Conformer-based human activity recognition using inertial measurement units, Sensors 23 (17) (2023) 7357. Human Fall Detection using Transfer Learning-based 3D CNN 9

    2023

  11. [19]

    Y. A. Andrade-Ambriz, S. Ledesma, M.-A. Ibarra-Manzano, M. I. Oros-Flores, D.-L. Almanza-Ojeda, Hu- man activity recognition using temporal convolutional neural network architecture, Expert Systems with Applications 191 (2022) 116287

    2022

  12. [20]

    Tahir, W

    A. Tahir, W. Taylor, A. Taha, M. Usman, S. A. Shah, M. A. Imran, Q. H. Abbasi, Iot based fall detection system for elderly healthcare, in: Internet of Things for Human-Centered Design: Application to Elderly Healthcare, Springer, 2022, pp. 209–232

    2022

  13. [21]

    Rezaee, M

    K. Rezaee, M. R. Khosravi, M. K. Moghimi, Intelligent elderly people fall detection based on modified deep learning deep transfer learning and iot using thermal imaging-assisted pervasive surveillance, in: Intelligent Healthcare: Infrastructure, Algorithms and Management, Spri...

    2022

  14. [22]

    D. Arun, H. Sumukh Subramanya, T. Goel, N. Tanush, J. S. Nayak, Video-based elderly fall detection using convolutional neural networks, in: Proceedings of Third International Conference on Intelligent Computing, Information and Control Systems: ICICCS 2021, Springer, 2022, pp. 803–814

    2021

  15. [23]

    Zhang, X

    X. Zhang, X. Zhou, M. Lin, J. Sun, Shufflenet: An extremely efficient convolutional neural network for mobile devices, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 6848– 6856

    2018

  16. [24]

    A. R. Inturi, V. Manikandan, V. Garrapally, A novel vision-based fall detection scheme using keypoints of human skeleton with long short-term memory network, Arabian Journal for Science and Engineering 48 (2) (2023) 1143–1155

    2023

  17. [25]

    H.-S. Fang, J. Li, H. Tang, C. Xu, H. Zhu, Y. Xiu, Y.-L. Li, C. Lu, Alphapose: Whole-body regional multi- person pose estimation and tracking in real-time, IEEE Transactions on Pattern Analysis and Machine Intel- ligence (2022)

    2022

  18. [26]

    Saurav, R

    S. Saurav, R. Saini, S. Singh, Vision-based techniques for fall detection in 360°videos using deep learning: Dataset and baseline results, Multimedia Tools and Applications 81 (10) (2022) 14173–14216

    2022

  19. [27]

    V.Patel,S.Kaple,V.R.Satpute,Indoorhumanfalldetectionusingdeeplearning,in:InternationalConference on Advancements in Interdisciplinary Research, Springer, 2022, pp. 235–242

    2022

  20. [28]

    K. Fei, C. Wang, J. Zhang, Y. Liu, X. Xie, Z. Tu, Flow-pose net: an effective two-stream network for fall detection, The Visual Computer 39 (6) (2023) 2305–2320

    2023

  21. [29]

    Zhang, H

    S. Zhang, H. Tong, J. Xu, R. Maciejewski, Graph convolutional networks: a comprehensive review, Compu- tational Social Networks 6 (1) (2019) 1–23

    2019

  22. [30]

    Egawa, A

    R. Egawa, A. S. M. Miah, K. Hirooka, Y. Tomioka, J. Shin, Dynamic fall detection using graph-based spatial temporal convolution and attention network, Electronics 12 (15) (2023) 3234

    2023

  23. [31]

    Amsaprabhaa, et al., Multimodal spatiotemporal skeletal kinematic gait feature fusion for vision-based fall detection, Expert Systems with Applications 212 (2023) 118681

    M. Amsaprabhaa, et al., Multimodal spatiotemporal skeletal kinematic gait feature fusion for vision-based fall detection, Expert Systems with Applications 212 (2023) 118681

    2023

  24. [32]

    Z. Chen, Y. Wang, W. Yang, Video based fall detection using human poses, in: CCF Conference on Big Data, Springer, 2022, pp. 283–296

    2022

  25. [33]

    Osigbesan, S

    A. Osigbesan, S. Barrat, H. Singh, D. Xia, S. Singh, Y. Xing, W. Guo, A. Tsourdos, Vision-based fall detec- tion in aircraft maintenance environment with pose estimation, in: 2022 IEEE International Conference on Multisensor Fusion and Integration for Intelligent Systems (MFI)...

    2022

  26. [34]

    S. Li, X. Song, Future frame prediction network for human fall detection in surveillance videos, IEEE Sensors Journal (2023)

    2023

  27. [35]

    L. Wu, C. Huang, L. Fei, S. Zhao, J. Zhao, Z. Cui, Y. Xu, Video-based fall detection using human pose and constrained generative adversarial network, IEEE Transactions on Circuits and Systems for Video Technology (2023)

    2023

  28. [36]

    Creswell, T

    A. Creswell, T. White, V. Dumoulin, K. Arulkumaran, B. Sengupta, A. A. Bharath, Generative adversarial networks: An overview, IEEE signal processing magazine 35 (1) (2018) 53–65

    2018

  29. [37]

    X. Zi, K. Chaturvedi, A. Braytee, J. Li, M. Prasad, Detecting human falls in poor lighting: Object detection and tracking approach for indoor safety, Electronics 12 (5) (2023) 1259

    2023

  30. [38]

    J. Leal, H. Moayyed, Z. Vale, Detection of human falls via computer vision for elderly care–an i3d/rnn approach, in: International Symposium on Distributed Computing and Artificial Intelligence, Springer, 2023, pp. 113–122

    2023

  31. [39]

    T. V. Ha, H. M. Nguyen, S. H. Thanh, B. T. Nguyen, Fall detection using mixtures of convolutional neural networks, Multimedia Tools and Applications (2023) 1–28

    2023

  32. [40]

    Alanazi, K

    T. Alanazi, K. Babutain, G. Muhammad, A robust and automated vision-based human fall detection system using 3d multi-stream cnns with an image fusion technique, Applied Sciences 13 (12) (2023) 6916

    2023

  33. [41]

    Szeghalmy, A

    S. Szeghalmy, A. Fazekas, A comparative study of the use of stratified cross-validation and distribution- balanced stratified cross-validation in imbalanced learning, Sensors 23 (4) (2023) 2333. 10 Ekram Alam et al

    2023

  34. [42]

    E. Alam, A. Sufian, P. Dutta, M. Leo, Vision-based human fall detection systems using deep learning: A review, Computers in biology and medicine 146 (2022) 105626

    2022

Pith tools

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