Pith. sign in

REVIEW 3 major objections 4 minor 32 references

HRTR: A Single-stage Transformer for Fine-grained Sub-second Action Segmentation in Stroke Rehabilitation

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 This paper claims that HRTR, a single-stage transformer using sliding windows and focal loss, achieves state-of-the-art Edit Scores on fine-grained sub-second action segmentation—70.1 on StrokeRehab Video, 69.4 on StrokeRehab IMU, and…

desk verdict The abstract's 'no refinements' SOTA claim is contradicted by the paper's own table; the model itself is a sensible single-stage transformer worth a second look after honest reporting. read the letter →

arxiv 2506.02472 v2 pith:2MN5RSVI submitted 2025-06-03 cs.CV

Halil Ismail Helvaci , Justin Philip Huber , Jihye Bae , Sen-ching Samson Cheung This is my paper
classification cs.CV
keywords actionsegmentationstrokerehabilitationsub-secondactionstransformerslidingwindowfocallossEditScorefine-grained
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

HRTR is a single-stage transformer for action segmentation that detects fine-grained, sub-second movements from continuous video or IMU data. The paper argues that a sliding-window transformer with focal loss can match or beat multi-stage segmentation systems that rely on refinement stages and boundary post-processing. On the StrokeRehab dataset—a benchmark of 120,891 annotated functional primitives from stroke-impaired patients—HRTR reaches Edit Scores of 69.8 on video and 68.9 on IMU without any post-processing, and 70.1 and 69.4 with a simple smoothing window (HRTR+). On the general 50Salads dataset, HRTR scores 85.1 and HRTR+ scores 88.4, the latter surpassing the previous best of 87.5. The claim matters because precise, sub-second action tracking is central to monitoring progress in stroke rehabilitation, and a single-stage model would be simpler and faster to deploy in clinical settings.

What carries the argument

The load-bearing mechanism is the sliding-window transformer encoder. The input feature sequence is cut into overlapping windows (width 200 for video, 500 for IMU, 5000 for 50Salads; strides of 10 or 500), each window is projected into 1024-dimensional embeddings with sinusoidal positional encoding, and a 3-layer transformer encoder with 4 attention heads (2 for 50Salads) produces per-frame class probabilities. The overlap between windows preserves global context while the limited window width keeps attention focused on fine-grained, sub-second transitions, and focal loss with α=25, γ=2 counteracts the class imbalance typical of rehabilitation motion primitives. The optional smoothing window (size 25 on StrokeRehab, 200 on 50Salads) averages outputs during inference and gives the HRTR+ variant its extra gain.

What would settle it

Run HRTR on the same three benchmarks with window size, stride, and smoothing width fixed before seeing test labels, selected on a held-out validation fold; if the Edit Scores drop below the published multi-stage baselines (e.g., ASPnet's 87.5 on 50Salads), the claim of single-stage superiority would fail. A second check: apply the exported model to a new stroke-rehab cohort and see whether the 70.1/69.4 edit scores reproduce.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single-stage transformer encoder—featuring sinusoidal positional encoding, overlapping sliding windows, and focal loss—can time-localize and classify sub-second actions as well as or better than multi-stage alternatives, without a separate boundary-refinement network or uncertainty-based post-processing. In its own experiments, unrefined HRTR surpasses the prior best on StrokeRehab video (Edit Score 69.8 vs. 67.6) and IMU (68.9 vs. 68.8), and the smoothing-enhanced HRTR+ sets new state-of-the-art scores of 70.1 on video, 69.4 on IMU, and 88.4 on 50Salads, beating ASPnet's 87.5 on the general dataset. The same architecture, with minor dimension changes, transfers across modalities and datasets, which the authors take as evidence that the sliding-window design captures both local temporal detail and global context.

Load-bearing premise

The window sizes, strides, and smoothing-window widths are tuned against the same datasets on which the model is evaluated, with no described validation split; if those hyperparameters were in effect chosen on the test set, the reported state-of-the-art gains could be selection artifacts rather than a property of the architecture.

Editorial extensions

If this is right

  • Multi-stage refinement is not required for state-of-the-art action segmentation: a single transformer encoder with sliding windows suffices on the tested benchmarks.
  • The same model configuration transfers across sensor modalities (video and IMU) and to a general action dataset (50Salads), suggesting the approach is not stroke-specific.
  • Window size is a decisive hyperparameter: video peaks at width 200, IMU at 500, and the choice trades off boundary precision against over-segmentation in a way the ablation quantifies.
  • The smoothing-window post-process improves Edit Score in all settings, with the largest gain on 50Salads (85.1 to 88.4).

Reading between the lines

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

  • Because the model processes fixed-width windows, memory use is bounded and the architecture could run in real time on streaming IMU or video, which the paper motivates but does not demonstrate.
  • The ablation's modality-dependent window optimum hints at an adaptively sized window—or a learned stride—as a natural next step; the paper only announces stride analysis as future work.
  • If the gains hold under a proper validation-based hyperparameter selection, the result would push the field toward lighter single-stage models for fine-grained temporal analysis in other clinical video tasks, such as surgical phase detection or movement disorder assessment.
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 introduces HRTR, a single-stage transformer encoder for fine-grained, sub-second action segmentation, evaluated on the StrokeRehab Video, StrokeRehab IMU, and 50Salads datasets. The method uses a sliding window, sinusoidal positional encoding, a small transformer encoder, and focal loss. The authors claim that HRTR 'without any refinements' outperforms state-of-the-art systems on all three datasets, with Edit Scores of 70.1, 69.4, and 88.4, and that it eliminates the need for multi-stage methods and post-processing. The experimental section reports comparisons against prior action-segmentation models (MS-TCN, ASRF, Seg2Seq, Raw2Seq, ASFormer, DiffAct, ASPnet, BaFormer), an ablation over window sizes, and per-class classification metrics.

Significance. If the claims held, a single-stage transformer that surpasses multi-stage and post-processing-based action segmentation methods on short-duration actions would be a practically valuable contribution, particularly for stroke rehabilitation monitoring. The paper's core architectural idea is simple and plausible, and the use of two benchmark datasets plus a clinical dataset is appropriate. However, the advertised state-of-the-art claim is central to the paper's contribution, and it is internally contradicted by the paper's own table: the numbers quoted in the abstract are the smoothed HRTR+ variants, and the unrefined HRTR does not achieve state-of-the-art on two of the three datasets. The absence of a validation protocol and of error bars further weakens the empirical case. The paper provides no code or machine-checked artifacts, so the contribution rests entirely on the reported experiments.

major comments (3)
  1. [Abstract / Table 1] The abstract states that 'Without any refinements, HRTR outperforms state-of-the-art systems ... achieving Edit Score (ES) of 70.1 on StrokeRehab Video, 69.4 on StrokeRehab IMU, and 88.4 on 50Salads.' Table 1 shows these three values in the HRTR+ row, which is defined in Section 5.3 as applying a smoothing-window post-processing step; the unrefined HRTR row reports 69.8, 68.9, and 85.1. Against the listed baselines, unrefined HRTR outperforms the prior best only on StrokeRehab Video (69.8 vs 67.6 for Seg2Seq); on StrokeRehab IMU it ties MS-TCN at 68.9 and has a worse AER than Raw2Seq (0.311 vs 0.305); on 50Salads it is below ASPnet (85.1 vs 87.5). The central claim of a post-processing-free state of the art is therefore not supported by the paper's own data. The abstract, the introduction (Section 1), and the conclusion (Section 7) must be revised to distinguish unrefined HRTR from HRTR+ and to state the results accurately.
  2. [§4.1, §6, Table 3] The window size w and stride s are described in Section 4.1 as 'hyper-parameters tuned to match the action events of interest', and Table 3 motivates the chosen values by their performance on the test datasets (e.g., video ES ranges from 62.0 at w=1500 to 70.1 at w=200; IMU ES ranges from 63.8 at w=100 to 69.4 at w=500). No validation split or selection protocol is described anywhere in Section 5.2 or Section 6. If these hyperparameters were selected based on test-set performance, the reported margins over prior methods (e.g., 70.1 vs 67.6 on video, 88.4 vs 87.5 on 50Salads) could be selection artifacts. The paper should either declare a held-out validation procedure, or report the performance of a fixed default configuration and treat the best-per-dataset choice as an upper bound, not as the headline result.
  3. [§5.2, Table 1] All results are reported as single numbers without error bars, standard deviations, or the number of runs. This is especially problematic for the small achievable margins that support the state-of-the-art claims: HRTR+ on IMU (69.4) is only 0.6 points above MS-TCN+ (68.8), and HRTR+ on 50Salads (88.4) is 0.9 points above ASPnet (87.5). Without a measure of variability it is impossible to judge whether these differences are meaningful. The authors should report mean and standard deviation over at least three random seeds, or otherwise justify that the differences exceed run-to-run noise.
minor comments (4)
  1. [Section 2 (Related Work)] The sentence 'Baformer [10] proposed a transformer-based model...' cites reference [10], which is ASFormer [10]; the BaFormer citation appears to be missing or incorrectly assigned, since BaFormer is not listed in the references.
  2. [Section 5.3] The 'smoothing windows' are introduced only in this section; they should be defined and motivated earlier, and the relationship between smoothing and the paper's claim of eliminating post-processing should be clarified.
  3. [Section 5.1] The definition of Edit Score is written 'ES(G, P) = 1 - L(G,P)/max(|G|,|P|) × 100', which is ambiguous because the multiplication by 100 applies to the entire fraction; it should be parenthesized as (1 - L(G,P)/max(|G|,|P|)) × 100.
  4. [Table 2] Per-class sensitivity, specificity, and F1 are reported only for HRTR, with no comparison to baseline methods; this makes it difficult to interpret whether the classification differences are meaningful.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reasoning found: HRTR is an empirical architecture evaluated on external benchmarks, and its reported results are not derived from its own inputs by construction.

full rationale

The paper proposes a single-stage transformer for action segmentation and evaluates it on StrokeRehab and 50Salads using standard train/test splits. No step in the derivation reduces to the paper's own inputs by definition. The sliding window size, stride, and smoothing window are hyperparameters tuned on the same datasets, which raises a selection-bias concern and is appropriately noted by the reader, but this is not logical circularity. The abstract's attribution of HRTR+ (smoothed) scores to unrefined HRTR is an internal reporting inconsistency, not a circular derivation. Self-citations to prior work by the authors are not load-bearing: reference [7] is the external StrokeRehab dataset paper, and reference [16] is only cited in related work as an example of transformer applications. Accordingly, no specific circular step can be exhibited, and the circularity score is 0.

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

The central results are empirical and depend on a modest set of manually chosen hyperparameters, pre-extracted features, and dataset assumptions. The most consequential choices are per-dataset window sizes and smoothing sizes, which the ablation study shows can swing Edit Scores by several points. No new theoretical entities or first-principles derivations are involved.

free parameters (5)
  • Window size w = 200 (video), 500 (IMU), 5000 (50Salads)
    Per-dataset hyperparameters described in Section 4.1 as tuned to match action events; Table 3 shows Edit Score varies from 62.0 to 70.1 with w, so the choice is load-bearing.
  • Stride s = 10 (IMU/video), 500 (50Salads)
    Chosen with the window size; the paper says detailed analysis of stride effects is left to future work.
  • Smoothing window size = 25 (StrokeRehab), 200 (50Salads)
    Used to produce HRTR+ results in Table 1; the headline Edit Scores come from the smoothed variant.
  • Focal loss alpha and gamma = alpha=25, gamma=2
    Set without sensitivity analysis in Section 5.2; focal loss is used because of class imbalance.
  • Transformer depth and width = 3 layers, 4 heads, hidden 512 (StrokeRehab); 3 layers, 2 heads, hidden 256 (50Salads)
    Architecture choices in Section 4.1; no search or justification beyond dataset-specific embedding differences.
assumptions (5)
  • domain assumption Pre-extracted X3D and I3D video features adequately represent sub-second actions
    The paper uses features from dataset authors [7] and standard I3D features [22] without extracting or validating them; performance is bounded by feature quality.
  • domain assumption StrokeRehab ground truth is reliable
    The paper quotes Cohen's kappa >= 0.96 from [7] but does not re-audit labels.
  • standard math Levenshtein-based Edit Score and AER are suitable evaluation metrics
    Section 5.1 defines metrics; ranking can differ across metrics, as shown by IMU AER.
  • domain assumption Train/test splits are identical to prior work
    Section 5.2 states 'we follow the same train/test splits as previous methods [7, 10]' without providing split files.
  • domain assumption Single training runs are representative
    No seeds or repeated runs are reported, so optimization variability is not characterized.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HRTR: A Single-stage Transformer for Fine-grained Sub-second Action Segmentation in Stroke Rehabilitation." pith.science (2026). https://pith.science/paper/2MN5RSVI

@misc{pith2026250602472,
  author       = {Pith},
  title        = {Pith review of: HRTR: A Single-stage Transformer for Fine-grained Sub-second Action Segmentation in Stroke Rehabilitation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2MN5RSVI}},
  note         = {Machine review of arXiv:2506.02472}
}
read the original abstract

Stroke rehabilitation often demands precise tracking of patient movements to monitor progress, with complexities of rehabilitation exercises presenting two critical challenges: fine-grained and sub-second (under one-second) action detection. In this work, we propose the High Resolution Temporal Transformer (HRTR), to time-localize and classify high-resolution (fine-grained), sub-second actions in a single-stage transformer, eliminating the need for multi-stage methods and post-processing. Without any refinements, HRTR outperforms state-of-the-art systems on both stroke related and general datasets, achieving Edit Score (ES) of 70.1 on StrokeRehab Video, 69.4 on StrokeRehab IMU, and 88.4 on 50Salads.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

32 extracted references · 30 canonical work pages

  1. [1]

    INTRODUCTION Stroke is a leading cause of disability, affecting over 795,000 individuals annually in the United States. Among stroke survivors, 77.4% experience arm impairments, which signifi- cantly hinder their ability to perform daily activities indepen- dently and diminish their quality of life [1, 2]. Rehabilitation focused on arm movements plays a c...

    work page Pith review arXiv 2025

  2. [2]

    introducing a single-stage transformer model that ef- fectively captures fine-grained, sub-second actions, address- ing the specific challenges of stroke rehabilitation, and

    work page

  3. [3]

    demonstrating the model’s generalization capability across diverse datasets, such as 50Salads, and establishing a new state of the art on the StrokeRehab Video and IMU datasets

    work page

  4. [4]

    RELATED WORK Action segmentation has traditionally relied on fixed-duration temporal models, which often struggled to capture long- range dependencies and complex temporal dynamics. Recur- rent Neural Networks (RNNs), particularly Long Short-Term Memory (LSTM) networks [8], improved temporal modeling but encountered challenges with vanishing gradients in ...

    work page

  5. [5]

    StrokeRehab The StrokeRehab dataset, developed by Kaku et al

    DATASET 3.1. StrokeRehab The StrokeRehab dataset, developed by Kaku et al. [7], in- cludes 3,372 trials from 51 stroke-impaired patients and 20 healthy subjects, designed for stroke rehabilitation research. It contains 120,891 annotated functional primitives across nine activities, such as feeding and brushing teeth. The an- notations, labeled by trained ...

    work page

  6. [6]

    , xT }, whereTdenotes the total number of discrete time steps, the goal of action segmentation is to generate the corresponding sequence of action labels,Y={y 1, y2, ..., yT }

    ACTION SEGMENTATION Given a sequence of input featuresX={x 1, x2, . . . , xT }, whereTdenotes the total number of discrete time steps, the goal of action segmentation is to generate the corresponding sequence of action labels,Y={y 1, y2, ..., yT }. This section describes the details of our proposed HRTR system in solving the action segmentation problem. 4...

    work page

  7. [7]

    EXPERIMENTS 5.1. Evaluation Metrics Segmentation performance is evaluated using the Levenshtein distance [7], which computes the minimum number of inser- tions, deletions, and substitutions required to transform the predicted sequencePinto the ground-truth sequenceG, de- noted asL(G, P). For instance,G= [reach, idle, retract] andP= [reach, stabilize]yield...

    work page 2021

  8. [8]

    The study evaluated various win- dow sizew, ranging from 100 to 1500, with stridesselected based on the best-performing ranges observed during prelim- inary experiments

    ABLATION STUDY An ablation study was conducted to investigate the impact of window size on HRTR+performance using the StrokeRehab dataset, detailed in Table 3. The study evaluated various win- dow sizew, ranging from 100 to 1500, with stridesselected based on the best-performing ranges observed during prelim- inary experiments. Smaller windows, 200 and 50...

    work page

Show all 32 references
  1. [9]

    Evaluated on the StrokeRehab and 50Salads datasets, HRTR achieved superior performance

    DISCUSSION AND CONCLUSION In this study we presented HRTR, a single-stage transformer model for high temporal resolution action segmentation, ad- dressing fine-grained and sub-second actions without the need for multi-stage frameworks. Evaluated on the StrokeRehab and 50Salads...

  2. [10]

    Heart disease and stroke statis- tics—2023 update: a report from the american heart associ- ation,

    Connie W Tsao, Aaron W Aday, Zaid I Almarzooq, Cheryl AM Anderson, Pankaj Arora, Christy L Avery, Carissa M Baker-Smith, Andrea Z Beaton, Amelia K Boehme, Alfred E Buxton, et al., “Heart disease and stroke statis- tics—2023 update: a report from the american heart associ- atio...

    2023

  3. [11]

    Estimates of the prevalence of acute stroke impair- ments and disability in a multiethnic population,

    Enas S Lawrence, Catherine Coshall, Ruth Dundas, Judy Stewart, Anthony G Rudd, Robin Howard, and Charles DA Wolfe, “Estimates of the prevalence of acute stroke impair- ments and disability in a multiethnic population,”Stroke, vol. 32, no. 6, pp. 1279–1284, 2001

    2001

  4. [12]

    World report on disability,

    World Health Organization et al., “World report on disability,” World Health Organization, 2011

    2011

  5. [13]

    Kinematic variables quantifying upper-extremity perfor- mance after stroke during reaching and drinking from a glass,

    Margit Alt Murphy, Carin Will ´en, and Katharina S Sunnerha- gen, “Kinematic variables quantifying upper-extremity perfor- mance after stroke during reaching and drinking from a glass,” Neurorehabilitation and neural repair, vol. 25, no. 1, pp. 71– 80, 2011

    2011

  6. [14]

    Does task-specific train- ing improve upper limb performance in daily life poststroke?,

    Kimberly J Waddell, Michael J Strube, Ryan R Bailey, Joseph W Klaesner, Rebecca L Birkenmeier, Alexander W Dromerick, and Catherine E Lang, “Does task-specific train- ing improve upper limb performance in daily life poststroke?,” Neurorehabilitation and neural repair, vol. 31,...

    2017

  7. [15]

    A novel combination of accelerometry and ecolog- ical momentary assessment for post-stroke paretic arm/hand use: feasibility and validity,

    Yi-An Chen, Marika Demers, Rebecca Lewthwaite, Nicolas Schweighofer, John R Monterosso, Beth E Fisher, and Carolee Winstein, “A novel combination of accelerometry and ecolog- ical momentary assessment for post-stroke paretic arm/hand use: feasibility and validity,”Journal of C...

    2021

  8. [16]

    Strokerehab: A benchmark dataset for sub-second action identification,

    Aakash Kaku, Kangning Liu, Avinash Parnandi, Haresh Ren- garaj Rajamohan, Kannan Venkataramanan, Anita Venkate- san, Audre Wirtanen, Natasha Pandit, Heidi Schambra, and Carlos Fernandez-Granda, “Strokerehab: A benchmark dataset for sub-second action identification,”Advances in...

    2022

  9. [17]

    Lattice long short-term memory for hu- man action recognition,

    Lin Sun, Kui Jia, Kevin Chen, Dit-Yan Yeung, Bertram E Shi, and Silvio Savarese, “Lattice long short-term memory for hu- man action recognition,” inProceedings of the IEEE interna- tional conference on computer vision, 2017, pp. 2147–2156

    2017

  10. [18]

    Ms-tcn: Multi-stage tem- poral convolutional network for action segmentation,

    Yazan Abu Farha and Jurgen Gall, “Ms-tcn: Multi-stage tem- poral convolutional network for action segmentation,” inPro- ceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 3575–3584

    2019

  11. [19]

    Asformer: Transformer for action segmentation,

    Fangqiu Yi, Hongyu Wen, and Tingting Jiang, “Asformer: Transformer for action segmentation,” inThe British Machine Vision Conference (BMVC), 2021

    2021

  12. [20]

    Efficient temporal action segmentation via boundary- aware query voting,

    Peiyao Wang, Yuewei Lin, Erik Blasch, Jie Wei, and Haibin Ling, “Efficient temporal action segmentation via boundary- aware query voting,”arXiv preprint arXiv:2405.15995, 2024

    2024 arXiv

  13. [21]

    Diffusion action segmenta- tion,

    Daochang Liu, Qiyue Li, Anh-Dung Dinh, Tingting Jiang, Mubarak Shah, and Chang Xu, “Diffusion action segmenta- tion,” inProceedings of the IEEE/CVF International Confer- ence on Computer Vision, 2023, pp. 10139–10149

    2023

  14. [22]

    Aspnet: Action seg- mentation with shared-private representation of multiple data sources,

    Beatrice van Amsterdam, Abdolrahim Kadkhodamohammadi, Imanol Luengo, and Danail Stoyanov, “Aspnet: Action seg- mentation with shared-private representation of multiple data sources,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp....

    2023

  15. [23]

    Attention is all you need,

    A Vaswani, “Attention is all you need,”Advances in Neural Information Processing Systems, 2017

    2017

  16. [24]

    Alleviating over-segmentation errors by detecting action boundaries,

    Yuchi Ishikawa, Seito Kasai, Yoshimitsu Aoki, and Hirokatsu Kataoka, “Alleviating over-segmentation errors by detecting action boundaries,” inProceedings of the IEEE/CVF win- ter conference on applications of computer vision, 2021, pp. 2322–2331

    2021

  17. [25]

    Localizing moments of actions in untrimmed videos of infants with autism spectrum disor- der,

    Halil Ismail Helvaci, Chen-Nee Chuah, Sally Ozonoff, and Sen-Ching Samson Cheung, “Localizing moments of actions in untrimmed videos of infants with autism spectrum disor- der,” in2024 IEEE International Conference on Image Pro- cessing (ICIP). IEEE, 2024, pp. 3841–3847

    2024

  18. [26]

    Rele- vance detection in cataract surgery videos by spatio-temporal action localization,

    Negin Ghamsarian, Mario Taschwer, Doris Putzgruber- Adamitsch, Stephanie Sarny, and Klaus Schoeffmann, “Rele- vance detection in cataract surgery videos by spatio-temporal action localization,” in2020 25th International conference on pattern recognition (ICPR). IEEE, 2021, pp....

    2021

  19. [27]

    Uncertainty-aware representation learning for action segmen- tation.,

    Lei Chen, Muheng Li, Yueqi Duan, Jie Zhou, and Jiwen Lu, “Uncertainty-aware representation learning for action segmen- tation.,” inIJCAI, 2022, vol. 2, p. 6

    2022

  20. [28]

    Combining embed- ded accelerometers with computer vision for recognizing food preparation activities,

    Sebastian Stein and Stephen J McKenna, “Combining embed- ded accelerometers with computer vision for recognizing food preparation activities,” inProceedings of the 2013 ACM in- ternational joint conference on Pervasive and ubiquitous com- puting, 2013, pp. 729–738

    2013

  21. [29]

    X3d: Expanding architectures for efficient video recognition,

    Christoph Feichtenhofer, “X3d: Expanding architectures for efficient video recognition,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 203–213

    2020

  22. [30]

    The kinetics hu- man action video dataset,

    Will Kay, Joao Carreira, Karen Simonyan, Brian Zhang, Chloe Hillier, Sudheendra Vijayanarasimhan, Fabio Viola, Tim Green, Trevor Back, Paul Natsev, et al., “The kinetics hu- man action video dataset,”arXiv preprint arXiv:1705.06950, 2017

    2017 arXiv

  23. [31]

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

    Joao Carreira and Andrew Zisserman, “Quo vadis, action recognition? a new model and the kinetics dataset,” inpro- ceedings of the IEEE Conference on Computer Vision and Pat- tern Recognition, 2017, pp. 6299–6308

    2017

  24. [32]

    Fully convolutional one-stage 3d object de- tection on lidar range images,

    Zhi Tian, Xiangxiang Chu, Xiaoming Wang, Xiaolin Wei, and Chunhua Shen, “Fully convolutional one-stage 3d object de- tection on lidar range images,”Advances in Neural Informa- tion Processing Systems, vol. 35, pp. 34899–34911, 2022

    2022

Pith tools

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