Pith. sign in

REVIEW 4 major objections 6 minor 38 references

TEMSET-24K: Densely Annotated Dataset for Indexing Multipart Endoscopic Videos using Surgical Timeline Segmentation

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read This paper introduces TEMSET-24K, a densely annotated dataset of 24,306 trans-anal endoscopic microsurgery (TEMS) video micro-clips, and claims that its STALNet model can segment surgical timelines with accuracy and F1 up to 0.99 on…

desk verdict Useful new TEMS dataset, but the validation protocol is under-specified and the benchmark numbers are not yet credible. read the letter →

arxiv 2502.06708 v1 pith:37U7GHHB submitted 2025-02-10 cs.CV

Muhammad Bilal , Mahmood Alam , Deepa Bapu , Stephan Korsgen , Neeraj Lal , Simon Bach , Amir M Hajivanand , Muhammed Ali
show 11 more authors
Kamran Soomro Iqbal Qasim Pawe{l} Capik Aslam Khan Zaheer Khan Hunaid Vohra Massimo Caputo Andrew Beggs Adnan Qayyum Junaid Qadir Shazad Ashraf
This is my paper · ORCID
classification cs.CV
keywords TEMSdatasetsurgicaltimelinesegmentationphasetaskactiontripletsendoscopicvideoindexingvideo-basedanalysisdeeplearningbenchmarkspatiotemporalLSTM
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 releases TEMSET-24K, a dataset of 24,306 trans-anal endoscopic microsurgery (TEMS) video micro-clips, each labeled by clinical experts with a hierarchical phase-task-action triplet. The authors' claim is that this dense annotation makes it possible to automatically segment a surgical timeline, marking when Setup, Dissection, Specimen Removal, Closure, and Scope Removal occur, using deep learning on the video alone. They report that their STALNet model reaches accuracy up to 0.99 and F1 up to 0.99 on well-represented phases, with weaker performance on rare actions. If true, this gives surgical teams a practical way to index and search large video libraries without manual review, and gives the field a public benchmark for surgical timeline segmentation.

What carries the argument

STALNet (Spatiotemporal Adaptive LSTM Network) is the central model: it runs a time-distributed image encoder over each micro-clip, feeds the feature sequence through an adaptive multi-layer LSTM with an attention layer, and outputs three classification heads for phase, task, and action. The complementary machinery is the micro-clip construction, which pairs each labeled keyframe with the 30 to 60 seconds of video leading up to it, giving the model motion context instead of a single still image. The phase-task-action taxonomy itself is the organizing structure that turns a continuous video stream into discrete, searchable timeline labels.

What would settle it

Inspect the validation column in timeline labels.csv: if microclips from the same TEMS surgery appear in both training and validation, recompute accuracy and F1 with a leave-one-surgery-out split; if the scores drop materially, the reported numbers measure memorization of nearby context rather than generalization to new procedures.

Watch

Extended reading notes

Core claim

The central claim is that a TEMS procedure can be represented as a three-level timeline of phase, task, and action, and that a spatiotemporal model can predict all three levels directly from short video micro-clips. The paper reports STALNet, tested with ConvNeXt, ViT, and SWIN V2 encoders, achieving accuracy up to 0.99 and F1 up to 0.99 on well-represented phases such as Setup and Suturing, while rare actions like Debris Wash and Haemostatis sit near F1 0.50. The dataset itself, 24,306 micro-clips with keyframes, JSON annotation exports, and a timeline labels table that also records remaining surgical time, is the primary contribution, along with the proposed taxonomy and the preprocessing library that turns raw multi-part videos into machine-learning-ready clips.

Load-bearing premise

The reported accuracy and F1 figures assume that no overlapping microclips from the same surgery are split across the training and validation sets, but the paper does not state how the validation column was constructed.

Editorial extensions

If this is right

  • A surgeon or training committee could search a large TEMS video library for specific events, such as bleeding or suturing, without manually reviewing hours of footage.
  • Because the dataset includes a time-to-finish column, the same micro-clips can be used to train models that estimate remaining operative time.
  • The phase-task-action triplet and the preprocessing recipe transfer to other video-recorded surgical specialties, so the dataset provides a template for similar benchmarks elsewhere.
  • Any future model trained on TEMSET-24K can be compared against the reported STALNet numbers, giving the field a common reference point.

Reading between the lines

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

  • Editorial extension: if the validation column splits microclips from the same surgery across training and validation, the headline 0.99 accuracy may overstate generalization; a leave-one-surgery-out evaluation would be the safer estimate of real-world performance.
  • Editorial extension: because each label is attached to a 30-to-60-second window ending at a keyframe, the model is likely good at recognizing which stage a moment belongs to, but the dataset does not test precise localization of the boundary frames where one phase becomes another.
  • Editorial extension: since access to the dataset is controlled by a restricted application process, reproducibility of the benchmark will depend on approval turnaround; the downloadable code alone does not make the underlying videos public.
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

4 major / 6 minor

Summary. The paper introduces TEMSET-24K, a dataset of microclips extracted from transanal endoscopic microsurgery (TEMS) videos, each annotated by clinical experts with a hierarchical phase/task/action label triplet. The authors also propose STALNet, a spatiotemporal model combining a time-distributed vision encoder (ConvNeXt, ViT, or SWIN V2) with an LSTM and attention, and report high accuracy and F1 scores for phase, task, and action prediction. The paper additionally describes an annotation pipeline, post-processing steps, and a released Python library for video processing.

Significance. If the dataset and its annotations are validated properly, TEMSET-24K would fill a clear gap in surgical data science: a densely annotated, expert-validated dataset for multipart endoscopic video indexing, with a clinically motivated three-level taxonomy. The release of preprocessing code and a benchmark model is a useful contribution to the community. However, the central evidence supporting the dataset's usability—the STALNet benchmark results—currently rests on an undocumented validation split and inconsistent dataset statistics, which must be corrected before the work can be fully credited.

major comments (4)
  1. [Post-Processing of Annotations to Generate ML-ready Dataset; Model Validation] The construction of the validation set is not specified. The paper only states that 'we added a valid column to indicate the suggested validation set' without describing whether the split is by surgery, by video, or by individual clip. Because Algorithm 1 creates each microclip as a temporal window ending at a keyframe, and keyframes are extracted at approximately 550 per video, microclips from the same surgery share substantial temporal context and, with a 60-second offset, can overlap directly. If the split is at clip level, the reported accuracy and F1 scores (Tables 1-4) may reflect memorization of nearby context rather than generalization to new procedures. This issue is load-bearing for the headline performance claim; the authors must document the split construction and, ideally, report results with a surgery-level split.
  2. [Abstract; Contributions; Data Records; Technical Validation] There are unreconciled counts across the manuscript. The abstract and contributions state 24,306 microclips and 84 actions, while Data Records states 24,704 microclips and Technical Validation states 21 actions. Additionally, the Discussion refers to '11 tasks' whereas the taxonomy and Table 3 list 12 tasks. These contradictions must be resolved, as they directly affect the dataset integrity and the interpretation of the benchmark results.
  3. [Technical Validation / Annotation Assessment] No inter-annotator agreement is reported. The paper describes that annotations were initially performed by one surgeon and validated by at least two other surgeons, with conflicts resolved through discussion, but it provides no quantitative agreement metrics such as Cohen's kappa, Fleiss' kappa, or percentage agreement for the phase, task, and action labels. For a dataset whose central contribution is expert annotation, inter-annotator agreement is a standard validation expectation; its absence weakens the claim of consistent, 'densely' annotated labels and makes it difficult to assess label quality.
  4. [Post-Processing of Annotations; Data Records] The microclip duration is inconsistent. The post-processing section states that microclips capture 'the 30 seconds of video leading up to that frame,' while the Data Records section states that the microclips folder contains '60-second video clips leading up to each keyframe.' This inconsistency directly affects the temporal context of each sample and the leakage analysis. The actual offset used in the released dataset must be clarified.
minor comments (6)
  1. [Tables 2-4 and Statistical Analysis] Please clarify whether the '±' values are standard deviations or standard errors; the text uses 'variance' but the reported values appear to be standard deviations, and some intervals (e.g., 0.99 ± 0.10) exceed the [0,1] range for accuracy and F1 scores.
  2. [References] The text says the annotation tool is Label Studio version 1.12.1, but the cited reference [33] is for WebAnno; please provide the correct citation for Label Studio.
  3. [Data Records] The 'valid column' is mentioned as indicating the suggested validation set, but its contents and usage are not documented; please include an example or a schema of the CSV columns.
  4. [Figure 6] The ROC curves are presented without AUC values or a clear explanation of how multi-class ROC curves are aggregated; please report AUC scores or specify the averaging method (e.g., one-vs-rest).
  5. [Equation (2)] The metrics in (2) are defined for binary classification; please clarify how accuracy and F1 are computed for the multiclass phase/task/action predictions, especially the handling of the 'No Action' class and class imbalance.
  6. [Fig. 5 and Problem Formulation] The term 'Adaptive LSTM' is not clearly explained; the statement that the number of LSTM layers depends on the input sequence length T is unusual and would benefit from a precise definition or rephrasing.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reported benchmark is an empirical evaluation against clinician ground truth, not a derivation that reduces to its inputs.

full rationale

TEMSET-24K is a dataset curation and benchmarking study. Its central outputs are a clinician-defined phase/task/action taxonomy, a set of 24k labeled micro-clips, and measured accuracy/F1 values for STALNet with ConvNeXt, ViT, and SWIN V2 encoders. These are empirical claims. There is no first-principles derivation whose conclusion could be equivalent to an input: the model is trained and evaluated on the same human-annotated labels, which is standard supervised benchmarking rather than a definitional circularity. The encoders and the fastai training library are external, and the clinical context citation (ref. 32, the TREc trial) is an externally published study, not a load-bearing self-citation. The main unresolved issue is that the 'valid column' introduced in Post-Processing is not documented as a surgery-level or patient-level split; because microclips are temporal windows ending at keyframes (Algorithm 1), overlapping clips could leak between training and validation. That is a data-leakage/soundness threat, not a circularity: even if leakage occurred, the reported numbers would be inflated rather than true by construction. No circular step can be exhibited from the paper's equations or citations, so the score is 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central benchmark rests on human annotations as ground truth, on keyframe-sampling coverage, and on the assumption that overlapping clips can be split without leakage. The loss weights, sequence length, microclip offset, and keyframe threshold are free choices whose values are not fully reported. No physical entities are introduced.

free parameters (4)
  • Microclip offset (temporal context length) = 30 s in Methods; 60 s in Data Records
    Chosen by authors; inconsistent values affect what each clip contains and therefore the temporal context available to the model.
  • Keyframe cosine-similarity threshold = not reported
    Controls how many keyframes are kept (about 550 per video clip); sampling density directly affects which short actions are represented.
  • STALNet sequence length T = not reported
    Number of frames per sequence is a training choice; no value is given.
  • Loss weights alpha, beta, gamma = not reported
    Weights for phase, task, and action losses are introduced but no values or tuning procedure is specified.
assumptions (5)
  • domain assumption Expert annotations exported from Label Studio are treated as ground truth.
    The benchmark compares models to these labels without inter-annotator agreement metrics; disagreements are resolved by discussion, so label noise is unmeasured. Technical Validation section.
  • domain assumption Cosine-distance keyframe selection preserves timeline coverage.
    The paper assumes scene-change-based keyframes represent all phases, tasks, and actions; no experiment verifies that rare or brief actions are not systematically dropped. Post-Processing section.
  • domain assumption A microclip's label equals the label of its terminal keyframe.
    Thirty to sixty seconds of video are collapsed to one triplet; short actions earlier in the clip may be mislabeled or missed. Algorithm 1 and Data Records.
  • ad hoc to paper Boundary overlaps can be resolved by keeping the trailing label.
    The transition column records overlaps, but the training labels are formed by the trailing label; the impact on boundary frames is not analyzed. Post-Processing section.
  • domain assumption Pretrained ImageNet encoders transfer to endoscopic imagery.
    ConvNeXt, ViT, and SWIN V2 are used with pretrained weights; no domain adaptation or fine-tuning-on-medical-data evidence is provided. DL Model Implementation section.

how reviews work

0 comments
Cite this review

Pith. "Pith review of TEMSET-24K: Densely Annotated Dataset for Indexing Multipart Endoscopic Videos using Surgical Timeline Segmentation." pith.science (2026). https://pith.science/paper/37U7GHHB

@misc{pith2026250206708,
  author       = {Pith},
  title        = {Pith review of: TEMSET-24K: Densely Annotated Dataset for Indexing Multipart Endoscopic Videos using Surgical Timeline Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/37U7GHHB}},
  note         = {Machine review of arXiv:2502.06708}
}
read the original abstract

Indexing endoscopic surgical videos is vital in surgical data science, forming the basis for systematic retrospective analysis and clinical performance evaluation. Despite its significance, current video analytics rely on manual indexing, a time-consuming process. Advances in computer vision, particularly deep learning, offer automation potential, yet progress is limited by the lack of publicly available, densely annotated surgical datasets. To address this, we present TEMSET-24K, an open-source dataset comprising 24,306 trans-anal endoscopic microsurgery (TEMS) video micro-clips. Each clip is meticulously annotated by clinical experts using a novel hierarchical labeling taxonomy encompassing phase, task, and action triplets, capturing intricate surgical workflows. To validate this dataset, we benchmarked deep learning models, including transformer-based architectures. Our in silico evaluation demonstrates high accuracy (up to 0.99) and F1 scores (up to 0.99) for key phases like Setup and Suturing. The STALNet model, tested with ConvNeXt, ViT, and SWIN V2 encoders, consistently segmented well-represented phases. TEMSET-24K provides a critical benchmark, propelling state-of-the-art solutions in surgical data science.

Figures

Figures reproduced from arXiv: 2502.06708 by the authors.

Figure 1
Figure 1. TEMS surgical workflow: A typical surgical flow from landmarking of the rectal polyp to dissection, lesion removal and closure of the rectal wall defect. The key milestones of a TEMS procedure are detailed in images A-J: [A] Baseline lesion in view after setup; [B] Application of landmark dots to outline the lesion; [C] Dissection of the wall through the mucosa and muscle; [D-E] Circumferential removal of the lesion… view at source ↗
Figure 2
Figure 2. Proposed Taxonomy of TEMS Surgical Workflow. The TEMS operation can be split into three levels: [A] High level activity phase (such as Set-up, Dissection, Specimen Removal, Closure and Scope Removal), [B] Task based activities (such as scope insertion, instrument movement, site wash and pressure increase), [C] Small unit tasks (such as tissue marking, tissue retraction, smoke identification, bleeding identification … view at source ↗
Figure 3
Figure 3. Side-by-side comparison of videos before and after pre-processing. Panels (a) and (b) show image quality before and after pre-processing, respectively. This shows that despite reduction in the size of the ESV file by a factor of 10 (from 1GB to 0.1GB), there was no loss in quality. Co-Creation of Dense Taxonomy Labels for Timeline Segmentation In our effort to develop a comprehensive taxonomy for annotation, the pro… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Methodology adopted for annotating TEMS surgical videos for surgical timeline segmentation includes six major steps: (1) TEMS surgical data acquisition (2) Data Preprocessing, (3) Data Annotation, (4) Annotation Verification, (5) Data Post Processing, and (6) Surgical …
Figure 5
Figure 5. Figure 5: Proposed SpatioTemporal Adaptive LSTM Network (STALNet) for Surgical Timeline Segmentation: This network diagram shows the process by which ESV clips are analysed by encoders in order to apply reliable timeline segments. Problem Formulation A key objective of this stud…
Figure 6
Figure 6. Figure 6: STALNet Performance Review using ROC Curves for Taxonomy Triplets: The top row of ROC curves shows the performance of ConvNeXt, ViT and SWIN V2 encoders on labelling high level TEMS surgical ”Phases”. The next two rows show the performance of STALNet encoders on labell…
Figure 7
Figure 7. Figure 7: STALNet: Batch of results for visual inspection: This figure illustrates the output of the STALNet model compared to human annotations—the ground truth (GT). Each tile displays the first, middle, and last frames of a video clip, along with predictions and GT for each t…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 35 canonical work pages

  1. [1]

    Weiser, T. G. et al. Estimate of the global volume of surgery in 2012: an assessment supporting improved health outcomes. The Lancet 385, S11 (2015)

    work page 2015

  2. [2]

    Nepogodiev, D. et al. Global burden of postoperative death. The Lancet 393, 401 (2019)

    work page 2019

  3. [3]

    Secure data environment

    England, N. Secure data environment. www.digital.nhs.uk/services/secure-data-environment-service(2021)

    work page 2021

  4. [4]

    NHS, N. Z. Delivering a net zero nhs. https://www.england.nhs.uk/greenernhs/a-net-zero-nhs/ (2020)

    work page 2020

  5. [5]

    Maier-Hein, L. et al. Surgical data science for next-generation interventions. Nat. Biomed. Eng. 1, 691–696 (2017)

    work page 2017

  6. [6]

    E., Lin, H

    Reiley, C. E., Lin, H. C., Yuh, D. D. & Hager, G. D. Review of methods for objective surgical skill evaluation. Surg. endoscopy 25, 356–366 (2011)

    work page 2011

  7. [7]

    Goodman, E. D. et al. A real-time spatiotemporal ai model analyzes skill in open surgical videos. arXiv preprint arXiv:2112.07219 (2021)

    arXiv 2021

  8. [8]

    Machine and deep learning for workflow recognition during surgery

    Padoy, N. Machine and deep learning for workflow recognition during surgery. Minim. Invasive Ther. & Allied Technol. 28, 82–90 (2019)

    work page 2019

Show all 38 references
  1. [9]

    Huaulm´e, A. et al. Offline identification of surgical deviations in laparoscopic rectopexy.Artif. Intell. Medicine 104, 101837 (2020)

    2020

  2. [10]

    Kadkhodamohammadi, A. et al. Towards video-based surgical workflow understanding in open orthopaedic surgery. Comput. Methods Biomech. Biomed. Eng. Imaging & Vis. 9, 286–293 (2021)

    2021

  3. [11]

    Holden, M. S. et al. Feasibility of real-time workflow segmentation for tracked needle interventions. IEEE Transactions on Biomed. Eng. 61, 1720–1728 (2014)

    2014

  4. [12]

    Padoy, N. et al. Statistical modeling and recognition of surgical workflow. Med. image analysis 16, 632–641 (2012)

    2012

  5. [13]

    Lin, H. C. et al. Automatic detection and segmentation of robot-assisted surgical motions. In International conference on medical image computing and computer-assisted intervention, 802–810 (Springer, 2005)

    2005

  6. [14]

    & Jannin, P

    Dergachyova, O., Bouget, D., Huaulm ´e, A., Morandi, X. & Jannin, P. Automatic data-driven real-time segmentation and recognition of surgical workflow. Int. journal computer assisted radiology surgery 11, 1081–1089 (2016)

    2016

  7. [15]

    Sv-rcnet: workflow recognition from surgical videos using recurrent convolutional network

    Jin, Y .et al. Sv-rcnet: workflow recognition from surgical videos using recurrent convolutional network. IEEE transactions on medical imaging 37, 1114–1126 (2017). 16/17

    2017

  8. [16]

    & Navab, N

    Blum, T., Feußner, H. & Navab, N. Modeling and segmentation of surgical workflow from laparoscopic video. In Medical Image Computing and Computer-Assisted Intervention–MICCAI 2010: 13th International Conference, Beijing, China, September 20-24, 2010, Proceedings, Part III 13, ...

    2010

  9. [17]

    Twinanda, A. P. et al. EndoNet: a deep architecture for recognition tasks on laparoscopic videos. IEEE transactions on medical imaging 36, 86–97 (2016)

    2016

  10. [18]

    Ramesh, S. et al. Multi-task temporal convolutional networks for joint recognition of surgical phases and steps in gastric bypass procedures. Int. journal computer assisted radiology surgery 16, 1111–1119 (2021)

    2021

  11. [19]

    & Heng, P.-A

    Gao, X., Jin, Y ., Long, Y ., Dou, Q. & Heng, P.-A. Trans-SVNet: Accurate phase recognition from surgical videos via hybrid embedding aggregation transformer. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2021: 24th International Conference, Strasbourg, ...

    2021

  12. [20]

    & Speidel, S

    Funke, I., Rivoir, D., Krell, S. & Speidel, S. TUNeS: A Temporal U-Net with Self-Attention for Video-based Surgical Phase Recognition. arXiv preprint arXiv:2307.09997 (2023)

    2023 arXiv

  13. [21]

    Video-based content analysis

    Huber, M. Video-based content analysis. In Campbell, A. G., Hong, L., Meinel, F. & Zallio, M. (eds.)Handbook of Research on Multimodal Human Computer Interaction and Pervasive Services, 10.4324/9780429316647-5 (CRC Press, 2020)

    doi

  14. [22]

    & ¨Ozsu, M

    Liu, L. & ¨Ozsu, M. T. Encyclopedia of database systems, vol. 6 (Springer New York, 2009)

    2009

  15. [23]

    Feldman, L. S. et al. SAGES Video-Based Assessment (VBA) program: a vision for life-long learning for surgeons. Surg. endoscopy 34, 3285–3288 (2020)

    2020

  16. [24]

    & Navab, N

    Vercauteren, T., Unberath, M., Padoy, N. & Navab, N. CAI4CAI: the rise of contextual artificial intelligence in computer-assisted interventions. Proc. IEEE 108, 198–214 (2019)

    2019

  17. [25]

    Nwoye, C. I. et al. Recognition of instrument-tissue interactions in endoscopic videos via action triplets. In Medical Image Computing and Computer Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8, 2020, Proceedings, Part III 23, 364–37...

    2020

  18. [26]

    Mascagni, P. et al. Computer vision in surgery: from potential to clinical value. npj Digit. Medicine 5, 163 (2022)

    2022

  19. [27]

    Lewandrowski, K.-U. et al. Regional variations in acceptance, and utilization of minimally invasive spinal surgery techniques among spine surgeons: results of a global survey. J. Spine Surg. 6, S260 (2020)

    2020

  20. [28]

    Richards, M. K. et al. A national review of the frequency of minimally invasive surgery among general surgery residents: assessment of ACGME case logs during 2 decades of general surgery resident training. JAMA surgery 150, 169–172 (2015)

    2015

  21. [29]

    & Buhmann, J

    Paysan, D., Haug, L., Bajka, M., Oelhafen, M. & Buhmann, J. M. Self-supervised representation learning for surgical activity recognition. Int. J. Comput. Assist. Radiol. Surg. 16, 2037–2044 (2021)

    2021

  22. [30]

    Valderrama, N. et al. Towards holistic surgical scene understanding. In International conference on medical image computing and computer-assisted intervention, 442–452 (Springer, 2022)

    2022

  23. [31]

    Ayobi, N. et al. Pixel-wise recognition for holistic surgical scene understanding. arXiv preprint arXiv:2401.11174 (2024)

    2024 arXiv

  24. [32]

    Bach, S. P. et al. Radical surgery versus organ preservation via short-course radiotherapy followed by transanal endoscopic microsurgery for early-stage rectal cancer (trec): a randomised, open-label feasibility study. The Lancet Gastroenterol. & Hepatol. 6, 92–105 (2021)

    2021

  25. [33]

    Gurevych, I., De Castilho, R. E. & Biemann, C. Webanno: A flexible, web-based and visually supported system for distributed annotations. 51st Annu. Meet. (2013)

    2013

  26. [34]

    Liu, Z. et al. A convnet for the 2020s. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 11976–11986 (2022)

    2022

  27. [35]

    Liu, Z. et al. Swin transformer v2: Scaling up capacity and resolution. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 12009–12019 (2022)

    2022

  28. [36]

    Steiner, A. et al. How to train your vit? data, augmentation, and regularization in vision transformers. arXiv preprint arXiv:2106.10270 (2021)

    2021 arXiv

  29. [37]

    Caron, M. et al. Emerging properties in self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, 9650–9660 (2021)

    2021

  30. [38]

    & Gugger, S

    Howard, J. & Gugger, S. Fastai: a layered api for deep learning. Information 11, 108 (2020). 17/17

    2020

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.