Pith. sign in

REVIEW 5 major objections 6 minor 46 references

Online Human Action Detection during Escorting

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A vision-only network that simultaneously re-identifies the person being escorted and classifies their action as following, lagging, or stopping, and it runs fast enough for online escorting on a new escorting dataset.

desk verdict A useful escort dataset and joint ReID/action network whose headline numbers rest on a possibly identity-leaking train/test split; worth serious review once that is settled. read the letter →

arxiv 2506.23573 v1 pith:24ZCDOTT submitted 2025-06-30 cs.RO cs.AIcs.LG

Siddhartha Mondal , Avik Mitra , Chayan Sarkar This is my paper · ORCID
classification cs.ROcs.AIcs.LG
keywords onlineactiondetectionpersonre-identificationescortingrobothumanEscortEdatasetvisiontransformerYOLOStripletloss
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a mobile robot can escort a person through a crowded indoor space by watching them continuously rather than assuming they will follow. The authors claim that person re-identification and online action detection can be done together in one transformer-based network, giving the robot the states it needs to adjust speed, pause, or resume escorting. To make this testable they introduce EscortE, a new dataset of 359 video sequences with frame-wise annotations for both tasks, and report that their joint system reaches a mean-average-precision score of 78.09 against 40.88 for the best baseline, while running 97% faster in inference (6.03 s versus 62.03 s). A sympathetic reader would care because this is a path to escortee-aware service robots that need no wearables, depth cameras, or separate pipelines.

What carries the argument

The central object is the joint YOLOS-based network, where YOLOS (You Only Look at One Sequence) is a vision-transformer object detector. Its person detection tokens are shared by both sub-tasks: an embedding layer projects them for re-identification against a reference embedding, and a temporal buffer of the selected person vectors is transformed and attention-weighted to predict the escortee's action. This shared-token design is what lets the system do re-identification and action prediction in one forward pass rather than two separate pipelines.

What would settle it

Make a version of the split where no person who appears in any training sequence appears in any test sequence, and rerun the joint evaluation; if the 78.09 mAP drops toward the 40.88 baseline, the reported advantage comes from subject leakage rather than the architecture.

Watch

Extended reading notes

Core claim

The central claim is that person re-identification and online human action detection can be fused into one network without sacrificing either task, and this fusion is what makes practical real-time escorting possible. Built on YOLOS and a vision-transformer backbone, the network reuses the detection tokens that YOLOS already produces: person vectors go through an embedding layer and are compared by L2 distance to a reference embedding for identity, while the same person vectors, buffered over a window of frames, pass through transformer layers and a self-attention weighted average to classify following, lagging, or stopping. On the new EscortE dataset the joint system reaches 78.09 mAP, and the two tasks together take 6.03 s per window, which the paper argues is a practical path to escortee-aware navigation in crowded indoor spaces.

Load-bearing premise

The dataset is split into training and test sequences without any statement that the same person never appears in both, so the re-identification scores may be inflated if a subject is memorized rather than re-identified.

Editorial extensions

If this is right

  • An escort robot using this module can slow down when the person lags, stop when they stop, and resume after occlusion, because the same network keeps the subject identity and action state updated.
  • The EscortE dataset gives researchers a common benchmark for escort-specific person re-identification and online action detection, where none existed before.
  • Because the system needs only an RGB camera, it can be added to existing navigation-based escort robots without wearables or depth sensors.
  • The shared-token design avoids the latency of chaining a separate detector, re-identifier, and action recognizer, making online escort monitoring more feasible on onboard compute.
  • The three-state semantics (following within 2 m, lagging beyond 2 m, stopping while stationary) translate directly into speed-control commands for the navigation module.

Reading between the lines

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

  • Beyond the paper: if a subject-disjoint evaluation confirms the reported precision, the same shared-token approach could generalize to other human-robot following tasks, such as warehouse co-workers or guidance for people with visual impairments.
  • Beyond the paper: the action boundary is currently a fixed distance heuristic; feeding the network monocular depth estimates, which the paper names as future work, could make the following/lagging distinction continuous and more robust to camera motion.
  • Beyond the paper: since detection, re-identification, and action prediction share one backbone, the main latency bottleneck is likely the object-detection pass; swapping the detector head for a lighter one or tracking tokens across frames is a testable route to true real-time operation.
  • Beyond the paper: the three-class scheme could be extended to include direction and head orientation to distinguish a deliberate pause from a stop caused by obstruction, which would let the robot decide whether to wait or re-plan.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. This paper introduces EscortE, a new dataset of 359 video sequences from 29 subjects in a shopping mall, annotated for person re-identification and online action detection with three classes (following, lagging, stopping). The authors propose a joint architecture built on YOLOS/ViT that computes person embeddings for re-identification and classifies the escortee's action over a sliding window of 60 frames. The reported results are EscortE-ReID precision 90.44%, EscortE-Action precision 79.40%, and a joint precision of 78.09% versus 40.88% for the best baseline, with an inference time of 6.03 s versus 62.03 s for the closest baseline. The paper claims real-time operation and superiority over strong baselines.

Significance. If the reported numbers survive scrutiny, the dataset is a useful community resource and the joint single-camera architecture is a practical contribution for escorting robots. The paper's strengths are the new dataset definition, the integration of ReID and action prediction in one network, and the explicit inference-time formula. However, the evaluation as written does not establish identity-disjoint generalization, uses precision while claiming mAP, and reports no variance, so the headline performance claims are not yet credible.

major comments (5)
  1. [Section IV.B] The split is described only as 250 training, 49 development, and 60 test sequences, with no statement that subject identities are disjoint across splits. With 29 subjects and roughly 12 sequences per subject, a sequence-level split almost certainly places the same subject on both sides of the training/test boundary. Because EscortE-ReID is trained on person appearance, evaluating on subjects seen during training can inflate ReID precision and propagate into the joint system's precision in Table VI. Please provide per-sequence subject IDs and a subject-disjoint split, or an explicit analysis showing that identity overlap does not affect the results, and recompute Tables III, IV, and VI under that split.
  2. [Section VI.A and Tables III, IV, VI] The text states that mean average precision (mAP) is used to evaluate both sub-tasks and the joint system, but all result tables report a single scalar labeled 'Precision.' For multi-class action detection and for re-identification, mAP and precision are different quantities and require different computation protocols (e.g., ranking by confidence). Please clarify what is actually computed, report per-class average precision and mAP for the action task, and report the corresponding ReID ranking metric; otherwise the numerical claims are not interpretable.
  3. [Section VI.D and Eq. (3)] The real-time claim is not supported by the reported numbers. For a 2 s window (w=60 at 30 fps), the reported inference time of 6.03 s for the proposed system is larger than the 2 s capture interval, and Eq. (3) subtracts capture time in a way that is not operationally defined: it is unclear whether t_i is wall-clock latency, per-window processing cost, or throughput. Please report per-frame ReID latency and action-detection latency separately, define the real-time criterion explicitly, and apply the same timing protocol to all baselines.
  4. [Section VI.D and Table V] The baseline systems in Table V are assembled from named components, but the paper does not describe the training or fine-tuning protocol for the baselines (e.g., learning rates, input resolutions, number of epochs, whether the same person detections and the same reference frames were used). Since the joint comparison is the paper's central claim, the baselines must be described to a reproducible level and evaluated under the same oracle or automatic ReID conditions; otherwise the large precision gap in Table VI may reflect setup differences rather than architectural superiority.
  5. [Section VI.D and Table VI] No error bars, standard deviations, or significance tests are reported for any of the precision numbers, and no per-subject or per-sequence breakdown is given. The differences between systems are large, but without variance information or an identity-disjoint evaluation, one cannot determine whether the improvement is robust or an artifact of the split issue raised above. Please report results disaggregated by subject or sequence, with confidence intervals or significance tests.
minor comments (6)
  1. [Section VI.A] The phrase 'the frames are recoded at 30 fps' should be 'recorded at 30 fps.'
  2. [Section V, Eq. (1)] The triplet loss is stated without specifying the triplet sampling strategy or the margin value; the margin is a free parameter that should be reported along with other training hyperparameters.
  3. [Section VI.B] The VGGFace baseline is mentioned in Table III but the text appears to cite only DeepFace [45]; please provide the correct citation for VGGFace.
  4. [Table I] The VIPeR row cites reference [27], which appears to be Richter et al., 'Playing for benchmarks'; the standard VIPeR citation is Gray and Tao, CVPR 2008. Please correct the reference.
  5. [Section IV.B] The paper states that participants agreed to make the dataset publicly available, but no dataset URL or release mechanism is provided; please add a link or availability statement.
  6. [Figure 3 caption] The caption says 'If the person becomes stationary, it is marked as stopping' but the action definitions in Section III describe stopping by stationarity; please clarify whether the annotation is based purely on distance thresholds or on a combination of distance and motion, since the two could conflict in some frames.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: the claimed predictions are grounded in an external dataset and held-out test sequences; the identity-overlap concern is a validity issue, not circularity.

full rationale

After walking the derivation chain, I find no step in which a predicted quantity is equivalent to an input by construction. The action labels are defined by physical distance criteria (Section IV.A: <2 m following, >2 m lagging, stationary stopping), not by the model's outputs, and the network is trained with triplet margin loss (Eq. 1) and cross-entropy loss (Eq. 2) against these external annotations. ReID is evaluated on held-out test sequences (Section IV.B), and the only hand-set scalar, the 1.5 L2 threshold in Section V, is a hyperparameter rather than a fitted prediction. The joint system numbers in Table VI combine independently trained ReID and action components, and the baselines use the same test protocol. The inference-time expression (Eq. 3) is a definitional accounting formula, not a learned result. The paper's self-citations ([20]–[23]) appear only in the illustrative system overview (ASR, NLU, route description) and are not load-bearing for the dataset or architecture claims. The reader-flagged train/test split by sequence rather than by subject is a legitimate validity concern—if identities overlap, ReID precision and hence the joint precision could be inflated—but that is a data-leakage/correctness issue, not a circularity of derivation. No quoted equation or citation chain reduces the claimed predictions to their inputs.

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

The paper introduces no new physical entities. It relies on standard deep learning components and a dataset of real humans. The main free parameters are a threshold and window size, and the key unstated assumption is the subject independence of the data split.

free parameters (3)
  • ReID distance threshold = 1.5
    The threshold for deciding whether a detected person is the subject is set empirically (Section V) and is central to re-identification performance.
  • Window length w = 60 frames (2 s at 30 fps)
    The action detection window size is chosen as 2 seconds and is used throughout the evaluation (Section VI-A).
  • Triplet loss margin = not specified
    The triplet loss for re-identification uses a margin hyperparameter that is not reported, but it influences the learned embeddings.
assumptions (3)
  • domain assumption The train/test split is by sequence and may allow the same subject to appear in both training and test sets.
    Section IV-B states sequences are split into training/development/test without confirming subject independence. This is a load-bearing assumption for person re-identification evaluation.
  • domain assumption The action labels (following, lagging, stopping) are accurately defined by distance thresholds (2 m) and the annotations are reliable.
    Section IV-A defines the three actions using distances and manual labeling. No inter-annotator agreement or label noise analysis is provided.
  • domain assumption YOLOS person detections are sufficient for both re-identification and action detection.
    The architecture relies on YOLOS detection tokens as the sole source of person features (Section V). Detection errors propagate to the downstream tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Online Human Action Detection during Escorting." pith.science (2026). https://pith.science/paper/24ZCDOTT

@misc{pith2026250623573,
  author       = {Pith},
  title        = {Pith review of: Online Human Action Detection during Escorting},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/24ZCDOTT}},
  note         = {Machine review of arXiv:2506.23573}
}
read the original abstract

The deployment of robot assistants in large indoor spaces has seen significant growth, with escorting tasks becoming a key application. However, most current escorting robots primarily rely on navigation-focused strategies, assuming that the person being escorted will follow without issue. In crowded environments, this assumption often falls short, as individuals may struggle to keep pace, become obstructed, get distracted, or need to stop unexpectedly. As a result, conventional robotic systems are often unable to provide effective escorting services due to their limited understanding of human movement dynamics. To address these challenges, an effective escorting robot must continuously detect and interpret human actions during the escorting process and adjust its movement accordingly. However, there is currently no existing dataset designed specifically for human action detection in the context of escorting. Given that escorting often occurs in crowded environments, where other individuals may enter the robot's camera view, the robot also needs to identify the specific human it is escorting (the subject) before predicting their actions. Since no existing model performs both person re-identification and action prediction in real-time, we propose a novel neural network architecture that can accomplish both tasks. This enables the robot to adjust its speed dynamically based on the escortee's movements and seamlessly resume escorting after any disruption. In comparative evaluations against strong baselines, our system demonstrates superior efficiency and effectiveness, showcasing its potential to significantly improve robotic escorting services in complex, real-world scenarios.

Figures

Figures reproduced from arXiv: 2506.23573 by the authors.

Figure 1
Figure 1. The three human action during escorting are – [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. An overview of a robotic escorting system with online [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Sample frames of a sequence from EscortE dataset – each frames of each sequence is annotated with action class, [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Network architecture for online human action detec [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Comparison of confusion matrices for action predic [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]
Figure 2
Figure 2. Figure 2: REFERENCES [1] M.-h. Jiao, H.-x. Wei, B.-w. Zhang, J.-q. Jin, Z.-q. Jia, and J.-l. Yan, “Path planning of escort robot based on improved quantum particle swarm optimization,” in 2019 Chinese Control And Decision Conference (CCDC). IEEE, 2019, pp. 3730–3735. [2] A. Ohya…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

46 extracted references · 41 canonical work pages

  1. [1]

    Path planning of escort robot based on improved quantum particle swarm optimization,

    M.-h. Jiao, H.-x. Wei, B.-w. Zhang, J.-q. Jin, Z.-q. Jia, and J.-l. Yan, “Path planning of escort robot based on improved quantum particle swarm optimization,” in 2019 Chinese Control And Decision Conference (CCDC). IEEE, 2019, pp. 3730–3735

    work page 2019

  2. [2]

    Intelligent escort robot moving together with human-interaction in accompanying behavior,

    A. Ohya and T. Munekata, “Intelligent escort robot moving together with human-interaction in accompanying behavior,” in Proceedings 2002 FIRA Robot World Congress , 2002, pp. 31–35

    work page 2002

  3. [3]

    An experiment on squad navigation of human and robots,

    L. Nomdedeu, J. Sales, E. Cervera, J. Alemany, R. Sebastia, J. Penders, and V . Gazi, “An experiment on squad navigation of human and robots,” in 2008 10th International Conference on Control, Automa- tion, Robotics and Vision . IEEE, 2008, pp. 1212–1218

    work page 2008

  4. [4]

    Follow me: interleaving human tracking and interacting with a new assistant robot,

    G. P ´erez, N. J. Zapata, S. Barroso, A. Torrej ´on, P. Bustos, and P. N ´u˜nez, “Follow me: interleaving human tracking and interacting with a new assistant robot,” in 2023 IEEE International Conference on Autonomous Robot Systems and Competitions (ICARSC) . IEEE, 2023, pp. 169–174

    work page 2023

  5. [5]

    Autonomous robotic escort incorporating motion prediction and human intention,

    D. Conte and T. Furukawa, “Autonomous robotic escort incorporating motion prediction and human intention,” in 2021 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2021, pp. 3480–3486

    work page 2021

  6. [6]

    A survey of open-world person re- identification,

    Q. Leng, M. Ye, and Q. Tian, “A survey of open-world person re- identification,” IEEE Transactions on Circuits and Systems for Video Technology, vol. 30, no. 4, pp. 1092–1108, 2019

    work page 2019

  7. [7]

    Deep learning for person re-identification: A survey and outlook,

    M. Ye, J. Shen, G. Lin, T. Xiang, L. Shao, and S. C. Hoi, “Deep learning for person re-identification: A survey and outlook,” IEEE transactions on pattern analysis and machine intelligence , vol. 44, no. 6, pp. 2872–2893, 2021

    work page 2021

  8. [8]

    Recurrent con- volutional network for video-based person re-identification,

    N. McLaughlin, J. M. Del Rincon, and P. Miller, “Recurrent con- volutional network for video-based person re-identification,” in Pro- ceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 1325–1334

    work page 2016

Show all 46 references
  1. [9]

    Top-push video-based person re-identification,

    J. You, A. Wu, X. Li, and W.-S. Zheng, “Top-push video-based person re-identification,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 1345–1353

    2016

  2. [10]

    Diversity regularized spa- tiotemporal attention for video-based person re-identification,

    S. Li, S. Bak, P. Carr, and X. Wang, “Diversity regularized spa- tiotemporal attention for video-based person re-identification,” in Proceedings of the IEEE conference on computer vision and pattern recognition, 2018, pp. 369–378

    2018

  3. [11]

    Video-based person re- identification with spatial and temporal memory networks,

    C. Eom, G. Lee, J. Lee, and B. Ham, “Video-based person re- identification with spatial and temporal memory networks,” in Pro- ceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 12 036–12 045

    2021

  4. [12]

    Spatial- temporal graph convolutional network for video-based person re- identification,

    J. Yang, W.-S. Zheng, Q. Yang, Y .-C. Chen, and Q. Tian, “Spatial- temporal graph convolutional network for video-based person re- identification,” in Proceedings of the IEEE/CVF conference on com- puter vision and pattern recognition , 2020, pp. 3289–3299

    2020

  5. [13]

    Appearance- preserving 3d convolution for video-based person re-identification,

    X. Gu, H. Chang, B. Ma, H. Zhang, and X. Chen, “Appearance- preserving 3d convolution for video-based person re-identification,” in Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part II 16 . Springer, 2020, pp. 228–243

    2020

  6. [14]

    Deep learning for sensor-based human activity recognition: Overview, challenges, and opportunities,

    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) , vol. 54, no. 4, pp. 1–40, 2021

    2021

  7. [15]

    Human activity recognition using tools of convolutional neural networks: A state of the art review, data sets, challenges, and future prospects,

    M. M. Islam, S. Nooruddin, F. Karray, and G. Muhammad, “Human activity recognition using tools of convolutional neural networks: A state of the art review, data sets, challenges, and future prospects,” Computers in Biology and Medicine , p. 106060, 2022

    2022

  8. [16]

    Human activity recognition: A spatio-temporal image encoding of 3d skeleton data for online action detection,

    N. Mokhtari, A. N ´ed´elec, and P. De Loor, “Human activity recognition: A spatio-temporal image encoding of 3d skeleton data for online action detection,” in 17th International Conference on Computer Vision Theory and Applications . SCITEPRESS-Science and Technology Publicati...

    2022

  9. [17]

    En- hanced spatio- temporal image encoding for online human activity recognition,

    N. Mokhtari, V . Fer, A. N ´ed´elec, M. Gilles, and P. de Loor, “En- hanced spatio- temporal image encoding for online human activity recognition,” in 2023 International Conference on Machine Learning and Applications (ICMLA) , 2023, pp. 884–889

    2023

  10. [18]

    Gatehub: Gated history unit with background suppression for online action detection,

    J. Chen, G. Mittal, Y . Yu, Y . Kong, and M. Chen, “Gatehub: Gated history unit with background suppression for online action detection,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 19 925–19 934

    2022

  11. [19]

    Miniroad: Minimal rnn framework for online action detection,

    J. An, H. Kang, S. H. Han, M.-H. Yang, and S. J. Kim, “Miniroad: Minimal rnn framework for online action detection,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 10 341–10 350

    2023

  12. [20]

    Can visual context improve automatic speech recognition for an embodied agent?

    P. Pramanick and C. Sarkar, “Can visual context improve automatic speech recognition for an embodied agent?” inProceedings of the 2022 Conference on Empirical Methods in Natural Language Processing . Association for Computational Linguistics, 2022, pp. 1946–1957

    2022

  13. [21]

    Enabling human-like task identification from natural conversation,

    P. Pramanick, C. Sarkar, P. Balamuralidhar, A. Kattepur, I. Bhat- tacharya, and A. Pal, “Enabling human-like task identification from natural conversation,” in 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2019, pp. 6196–6203

    2019

  14. [22]

    tage: Enabling an embodied agent to understand human instructions,

    C. Sarkar, A. Mitra, P. Pramanick, and T. Nayak, “tage: Enabling an embodied agent to understand human instructions,” in Findings of the Association for Computational Linguistics: EMNLP 2023 . Association for Computational Linguistics, 2023, pp. 8846–8857

    2023

  15. [23]

    How much is too much: Exploring the effect of verbal route description length on indoor navigation,

    F. Nourin N, P. Pramanick, and C. Sarkar, “How much is too much: Exploring the effect of verbal route description length on indoor navigation,” in 2024 33rd IEEE International Conference on Robot and Human Interactive Communication (ROMAN) , 2024, pp. 1378– 1385

    2024

  16. [24]

    Scalable person re-identification: A benchmark,

    L. Zheng, L. Shen, L. Tian, S. Wang, J. Wang, and Q. Tian, “Scalable person re-identification: A benchmark,” in Proceedings of the IEEE international conference on computer vision , 2015, pp. 1116–1124

    2015

  17. [25]

    Consistent re-identification in a camera network,

    A. Das, A. Chakraborty, and A. K. Roy-Chowdhury, “Consistent re-identification in a camera network,” in European Conference on Computer Vision, ser. Lecture Notes in Computer Science, vol. 8690. Springer, 2014, pp. 330–345

    2014

  18. [26]

    Deepreid: Deep filter pairing neural network for person re-identification,

    W. Li, R. Zhao, T. Xiao, and X. Wang, “Deepreid: Deep filter pairing neural network for person re-identification,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2014, pp. 152–159

    2014

  19. [27]

    Playing for benchmarks,

    S. R. Richter, Z. Hayder, and V . Koltun, “Playing for benchmarks,” in Proceedings of the IEEE international conference on computer vision , 2017, pp. 2213–2222

    2017

  20. [28]

    Unsupervised person re-identification by deep learning tracklet association,

    M. Li, X. Zhu, and S. Gong, “Unsupervised person re-identification by deep learning tracklet association,” in Proceedings of the European conference on computer vision (ECCV) , 2018, pp. 737–753

    2018

  21. [29]

    Human reidentification with transferred metric learning,

    W. Li, R. Zhao, and X. Wang, “Human reidentification with transferred metric learning,” in ACCV, 2012

    2012

  22. [30]

    Locally aligned feature transforms across views,

    W. Li and X. Wang, “Locally aligned feature transforms across views,” in CVPR, 2013

    2013

  23. [31]

    Custom pictorial structures for re-identification

    D. S. Cheng, M. Cristani, M. Stoppa, L. Bazzani, V . Murino, et al. , “Custom pictorial structures for re-identification.” in Bmvc, vol. 1, no. 2. Citeseer, 2011, p. 6

    2011

  24. [32]

    Online human action detection using joint classification-regression recurrent neural networks,

    Y . Li, C. Lan, J. Xing, W. Zeng, C. Yuan, and J. Liu, “Online human action detection using joint classification-regression recurrent neural networks,” European Conference on Computer Vision , 2016

    2016

  25. [33]

    PAMAP2 Physical Activity Monitoring,

    A. Reiss, “PAMAP2 Physical Activity Monitoring,” UCI Machine Learning Repository, 2012, DOI: https://doi.org/10.24432/C5NW2H

    2012 doi

  26. [34]

    A public domain dataset for human activity recognition using smart- phones

    D. Anguita, A. Ghio, L. Oneto, X. Parra, J. L. Reyes-Ortiz, et al. , “A public domain dataset for human activity recognition using smart- phones.” in Esann, vol. 3, 2013, p. 3

    2013

  27. [35]

    Mm-fit: Multimodal deep learning for automatic exercise logging across sensing devices,

    D. Str ¨omb¨ack, S. Huang, and V . Radu, “Mm-fit: Multimodal deep learning for automatic exercise logging across sensing devices,” Pro- ceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies, vol. 4, no. 4, pp. 1–22, 2020

    2020

  28. [36]

    The daily home life activity dataset: a high semantic activity dataset for online recog- nition,

    G. Vaquette, A. Orcesi, L. Lucat, and C. Achard, “The daily home life activity dataset: a high semantic activity dataset for online recog- nition,” in 2017 12th IEEE International Conference on Automatic Face & Gesture Recognition (FG 2017) . IEEE, 2017, pp. 497–504

    2017

  29. [37]

    Movi: A large multipurpose motion and video dataset. arxiv 2020,

    S. Ghorbani, K. Mahdaviani, A. Thaler, K. Kording, D. Cook, G. Blohm, and N. Troje, “Movi: A large multipurpose motion and video dataset. arxiv 2020,” arXiv preprint arXiv:2003.01888 , 2020

    2020 arXiv

  30. [38]

    Home action genome: Cooperative compositional action understanding,

    N. Rai, H. Chen, J. Ji, R. Desai, K. Kozuka, S. Ishizaka, E. Adeli, and J. C. Niebles, “Home action genome: Cooperative compositional action understanding,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2021, pp. 11 184–11 193

    2021

  31. [39]

    An image is worth 16x16 words: Transformers for image recognition at scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly,et al., “An image is worth 16x16 words: Transformers for image recognition at scale,” arXiv preprint arXiv:2010.11929 , 2020

    2010 arXiv

  32. [40]

    Depth anything: Unleashing the power of large-scale unlabeled data,

    L. Yang, B. Kang, Z. Huang, X. Xu, J. Feng, and H. Zhao, “Depth anything: Unleashing the power of large-scale unlabeled data,” arXiv preprint arXiv:2401.10891, 2024

    2024 arXiv

  33. [41]

    You only look at one sequence: Rethinking transformer in vision through object detection,

    Y . Fang, B. Liao, X. Wang, J. Fang, J. Qi, R. Wu, J. Niu, and W. Liu, “You only look at one sequence: Rethinking transformer in vision through object detection,” Advances in Neural Information Processing Systems, vol. 34, pp. 26 183–26 197, 2021

    2021

  34. [42]

    In defense of the triplet loss for person re-identification,

    A. Hermans, L. Beyer, and B. Leibe, “In defense of the triplet loss for person re-identification,” arXiv preprint arXiv:1703.07737 , 2017

    2017 arXiv

  35. [43]

    Hierarchical clustering with hard-batch triplet loss for person re-identification,

    K. Zeng, M. Ning, Y . Wang, and Y . Guo, “Hierarchical clustering with hard-batch triplet loss for person re-identification,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 13 657–13 665

    2020

  36. [44]

    Generalized cross entropy loss for training deep neural networks with noisy labels,

    Z. Zhang and M. Sabuncu, “Generalized cross entropy loss for training deep neural networks with noisy labels,” Advances in neural information processing systems , vol. 31, 2018

    2018

  37. [45]

    Deepface: Closing the gap to human-level performance in face verification,

    Y . Taigman, M. Yang, M. Ranzato, and L. Wolf, “Deepface: Closing the gap to human-level performance in face verification,” in Pro- ceedings of the IEEE conference on computer vision and pattern recognition, 2014, pp. 1701–1708

    2014

  38. [46]

    Clip-reid: exploiting vision-language model for image re-identification without concrete text labels,

    S. Li, L. Sun, and Q. Li, “Clip-reid: exploiting vision-language model for image re-identification without concrete text labels,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 37, no. 1, 2023, pp. 1405–1413

    2023

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.