Pith. sign in

REVIEW 4 major objections 5 minor 30 references

Predicting Soccer Penalty Kick Direction Using Human Action Recognition

T0 review · 4 major / 5 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 deep learning model trained on kicker motion predicts penalty kick direction at 63.9% accuracy, beating the 54.2% rate of real goalkeepers on the same clips.

desk verdict Useful new penalty-kick dataset and a thorough backbone sweep, but the post-contact input frames make the 'anticipation' claim unverified and the goalkeeper comparison unfair. read the letter →

arxiv 2507.12617 v1 pith:PZ2M5OAI submitted 2025-07-16 cs.CV

David Freire-Obregón , Oliverio J. Santana , Javier Lorenzo-Navarro , Daniel Hernández-Sosa , Modesto Castrillón-Santana This is my paper
classification cs.CV
keywords humanactionrecognitionanticipationpenaltykickpredictionsocceranalyticsvisiontransformersMViTv2videoclassificationsportsdataset
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 a computer vision model can anticipate the direction of a soccer penalty kick from the kicker's body movement alone, and that it does so more reliably than human goalkeepers do in real time. The authors introduce a newly collected, manually annotated dataset of penalty clips and a pipeline that isolates the kicker, extracts video embeddings with pre-trained HAR backbones, and fuses them with metadata such as pitch side and kicking foot. Across twenty-two models in seven architecture families, the best binary left/right classifier reaches 63.9% accuracy, compared with 54.2% for the goalkeepers' actual decisions on the same footage. If the claim holds, it means anticipatory motion cues carry usable directional information that current goalkeeper behavior does not fully exploit.

What carries the argument

The mechanism is a two-stream action-anticipation pipeline. First, ByteTrack isolates the kicker's bounding box in every frame, and that box is superimposed on the clip's average frame, producing a video in which the kicker is the only moving element. The run-up (32 frames) and kicking motion (16 frames) are then passed separately through a pre-trained video backbone; the resulting per-chunk embeddings are pooled (average or max) into one vector per phase. A classifier with fully connected layers concatenates the running embedding, the kicking embedding, and two metadata bits (pitch side and kicking foot), and outputs the shot direction. The name-carrying component is MViTv2, a multiscale vision transformer, whose average-pooled 16-frame configuration gives the best binary accuracy.

What would settle it

Run the same classifier on a version of the dataset trimmed to end at the moment of ball contact; if accuracy falls well below 63.9%, or if a model trained only on the 8 post-contact frames also predicts direction above chance, then the paper's anticipatory framing is not supported. Inspecting saliency maps for attention on the ball region or the follow-through leg would provide additional evidence of leakage.

Watch

Extended reading notes

Core claim

The paper's central claim is that the direction of a penalty kick is readable from the kicker's run-up and kicking motion before the ball's trajectory becomes visible, and that a deep learning system can extract that information better than goalkeepers do. The evidence is a 640-clip dataset drawn from 1,010 collected web videos, each clip trimmed to 48 frames (32 running frames plus 8 frames before and 8 after ball contact), with the kicker isolated by tracking and superimposed on a static background. Feature embeddings from pre-trained HAR backbones, pooled and combined with metadata, are passed through a two-stream classifier. MViTv2 S with average pooling achieves 63.9% accuracy for binary left/right prediction and an F1-score of 62.5%, while the real-goalkeeper baseline on the same clips is 54.2%; the best three-class model, MViT CONV, reaches 51.9% accuracy. The paper also reports that splitting the footage into running and kicking branches and adding metadata each improve accuracy by several points.

Load-bearing premise

The kicking phase extends 8 frames past the moment the ball is kicked, and the paper assumes that isolating the kicker's bounding box and pasting it on a fixed background removes all ball-trajectory and outcome information, so the model is reading anticipatory body motion rather than the result of the kick.

Editorial extensions

If this is right

  • If the 63.9% binary result holds on new footage, the same pipeline can support goalkeeper training by showing which body cues in the run-up and kicking motion are most informative.
  • With a real-time tracker, the classifier could be extended to online match support, offering a probabilistic left/right read while the kick is being taken.
  • The dataset becomes a benchmark for HAR action anticipation, letting future backbones be compared on a real-world sports prediction task rather than only on curated activity datasets.
  • The reported 3-5% gains from two-stream phase separation and 2-3% gains from metadata indicate that both temporal structure and context matter; follow-up work on other sports with staged actions may benefit from the same design.

Reading between the lines

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

  • A direct test of the anticipatory claim would be to retrain the classifier on a version of the kicking phase truncated at the moment of ball contact; if accuracy stays near 63.9%, the model is reading pre-contact cues, and if it falls substantially, the current 8-frame post-contact window is leaking outcome information.
  • The model's internal saliency could be inspected to see whether it attends to the kicking leg, the support leg, or the run-up angle; a human-interpretable cue list could yield a simple rule-based early-warning system for goalkeepers.
  • The same framework likely transfers to other penalty-like actions in which an athlete reveals direction through approach motion, such as free kicks, baseball pitches, or tennis serves, though each would need its own dataset and metadata.
  • Because the dataset is web-sourced and heterogeneous, an adversarial test with held-out competitions or camera angles would indicate whether the learned cues are general or partly dataset-specific.
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 / 5 minor

Summary. The paper introduces a manually annotated dataset of 640 soccer penalty-kick clips and proposes a two-stream classifier that combines HAR embeddings (from running and kicking phases) with contextual metadata (pitch side, kicking foot) to predict shot direction. Twenty-two backbones across seven architecture families are evaluated in three-class (left/center/right) and binary (left/right) settings. The best binary model, MViTv2, achieves 63.9% accuracy, which the authors compare favorably with a 54.2% goalkeeper-decision baseline and interpret as evidence that anticipatory motion cues can outperform human goalkeepers.

Significance. If the central temporal-window concern is resolved, the dataset is a useful contribution: the authors correctly note the absence of a public penalty-kick HAR dataset, and the systematic comparison of 22 backbones across seven families provides a solid benchmark. The inclusion of a human goalkeeper baseline is a valuable reference point. However, the paper's main claim—that the model predicts direction from pre-kick player movements—depends on the assertion that the 48-frame window, which includes 8 frames after the kicking moment, does not leak outcome information. Without addressing this, the headline 63.9% figure cannot be attributed to anticipation, and the comparison with goalkeepers is unfair.

major comments (4)
  1. [Section 5, Dataset paragraph; Section 4, Eq. (4)] The 48-frame window is defined as 32 run-up frames plus 8 frames before and 8 frames after the kicking moment, and the paper states that 'This interval excludes visual information about the ball's trajectory or outcome, focusing solely on kick motion cues.' This assertion is not credible as stated. The 8 post-contact frames contain the kicker's follow-through, hip rotation, and weight transfer, which are strong, direct signals of shot direction; the preprocessing in Section 4 removes background and other players but not the kicker's own post-contact motion. At the contact frame, the ball can also be inside or adjacent to the kicker's bounding box. Therefore the 63.9% binary accuracy in Table 2 may reflect reading the outcome of the kick rather than anticipating it, and the comparison with the 54.2% goalkeeper baseline (which must commit before or during the kick) is not apples-to-apples. I request an ablation that removes post-contact frames (e.g., 32 run-up + 8 pre-contact frames only) and a curve showing accuracy as a function of the number of post-contact frames (0, 2, 4, 6, 8) to quantify leakage.
  2. [Section 5, Experimental Setup; Tables 1 and 2] No confidence intervals, standard deviations, or significance tests are reported for any accuracy value. On 640 samples with 10-fold cross-validation, the gap between the best model (MViTv2, 63.9%) and the second-best (MViTv1, 61.8%) is small enough to be sampling noise, and the same holds for the difference from the 54.2% goalkeeper baseline. The paper should report per-fold results, confidence intervals, and paired statistical tests (e.g., McNemar's test) for the central comparisons, otherwise the claim of 'outperforming' goalkeepers is not statistically substantiated.
  3. [Section 5, Experimental Setup; Section 6] For each of the 22 backbones, the reported result is the best-performing variant among different frame counts and pooling strategies (e.g., Table 2 selects MViTv2 S with 16 frames and Average pooling). Selecting the best variant per architecture on validation folds and then reporting the best over 22 architectures induces a multiple-comparisons optimism that is not quantified. The paper should use nested cross-validation or report all variants plus the explicit selection rule, and should discuss the expected inflation of the reported 63.9%.
  4. [Section 5, results narrative (after Table 2)] The claims that 'Metadata incorporation improved accuracy by 2-3% and F1-Score by 5%' and 'Using two-stream embedding branches instead of one increased accuracy by 3-5%' are central to justifying the proposed two-stream classifier design, but they are presented without a supporting table, figure, or error bars. These ablations should be documented under the same 10-fold cross-validation protocol as the main results, with variance estimates.
minor comments (5)
  1. [Introduction] The text contains typographical errors: 'FIF A' and 'UEF A' should be 'FIFA' and 'UEFA'.
  2. [Acknowledgments] The Acknowledgments contain a duplicated word: 'partially funded funded' should be 'partially funded'.
  3. [Section 4, Eq. (4)] The notation for the bounding box BB(i)(t) and the union operation in Eq. (4) is informal; please clarify whether BB(i)(t) is a binary mask, a crop, or a transparency-weighted region, and how the superposition onto the average frame is implemented.
  4. [Section 5, Dataset paragraph] The paper states that original videos last between 3 and 6 seconds at 1280x720 resolution but then excludes clips with fewer than 64 frames; please state the assumed frame rate and explain how temporal padding to 48 frames interacts with clips whose running phase is shorter than 32 frames.
  5. [Figures 5 and 6] The boxplot captions do not specify whether the distributions are over cross-validation folds or over model variants; please clarify the aggregation to make the plots interpretable.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation: held-out cross-validation makes the accuracy a genuine prediction; self-citations are background only.

full rationale

The claimed result (63.9% binary direction accuracy, Tables 1 and 2) is obtained by training the mapping C:(T_run, T_kick, Gamma) -> y and evaluating on held-out test folds in a 10-fold cross-validation (Section 5, Experimental Setup). No fitted parameter or derived equation is reused as a prediction: the label y is manually annotated shot direction, and the HAR embeddings and metadata are inputs, not outputs, of the fitted model. The comparison to the goalkeeper baseline is an external benchmark, and the goalkeeper's 54.2% accuracy is not used to train or calibrate the model. The only self-references ([2], [10], [22]) are background/preprocessing citations and are not load-bearing for the central claim. The potential leakage from the 8 post-contact frames is a data-validity concern about whether the task truly measures anticipation, not a circular derivation: those frames are not constructed from the labels, and the evaluation remains held-out. Therefore no circular step can be exhibited under the required standard.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper contributes a dataset and an empirical benchmark; it introduces no mathematical constants or invented entities. The hand-chosen frame split and pooling selection are the main tuning choices that affect the headline accuracy, and the post-kick frames assumption is the least-verified load-bearing premise.

free parameters (3)
  • Running/kicking frame split = 32/16 frames
    Hand-chosen clip lengths; the 16-frame kicking stage includes 8 frames after the kick moment, which is load-bearing for the anticipation claim.
  • Embedding pooling selection = Max or Average per architecture
    Pooling method chosen per architecture based on highest validation performance during cross-validation; affects reported scores.
  • MLP classifier widths = 256/128/64 units
    Chosen without a stated search; part of the classification module that fuses embeddings and metadata.
assumptions (4)
  • domain assumption The 48-frame window (32 run, 8 pre, 8 post kick) captures anticipatory motion cues and excludes ball trajectory
    Section 5 Dataset; the paper asserts this without demonstrating that the post-kick frames do not reveal outcome information.
  • domain assumption Pre-trained Kinetics backbones transfer to penalty kick motion
    Step 2 relies on transfer from generic human action recognition to soccer-specific kicker motion.
  • domain assumption Human-annotated shot direction labels are reliable ground truth
    Section 5; the center class is admitted to be ambiguous, which motivated the two-class setup, yet labels are used for training and evaluation.
  • domain assumption Goalkeeper dive direction is a valid human baseline
    GK baseline in Tables 1 and 2; assumes the annotated dive direction represents real-time GK decisions with no annotation error.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting Soccer Penalty Kick Direction Using Human Action Recognition." pith.science (2026). https://pith.science/paper/PZ2M5OAI

@misc{pith2026250712617,
  author       = {Pith},
  title        = {Pith review of: Predicting Soccer Penalty Kick Direction Using Human Action Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PZ2M5OAI}},
  note         = {Machine review of arXiv:2507.12617}
}
read the original abstract

Action anticipation has become a prominent topic in Human Action Recognition (HAR). However, its application to real-world sports scenarios remains limited by the availability of suitable annotated datasets. This work presents a novel dataset of manually annotated soccer penalty kicks to predict shot direction based on pre-kick player movements. We propose a deep learning classifier to benchmark this dataset that integrates HAR-based feature embeddings with contextual metadata. We evaluate twenty-two backbone models across seven architecture families (MViTv2, MViTv1, SlowFast, Slow, X3D, I3D, C2D), achieving up to 63.9% accuracy in predicting shot direction (left or right), outperforming the real goalkeepers' decisions. These results demonstrate the dataset's value for anticipatory action recognition and validate our model's potential as a generalizable approach for sports-based predictive tasks.

Figures

Figures reproduced from arXiv: 2507.12617 by the authors.

Figure 1
Figure 1. Penalty-kick direction estimation. Our method predicts ball direction by combining metadata and HAR embeddings, using only kicker actions without ball trajectory data. formations [18], and key moment detection [5]. These advances support applica￾tions such as strategy analysis and predictive modeling. In this context, penalty kick outcome prediction can benefit from incorporating factors like tempo, tech￾nique, and … view at source ↗
Figure 2
Figure 2. Embeddings extraction module. The video is downsampled into n clips. A pre-trained HAR model extracts features, combined via pooling (average or max) to produce a final tensor for the classifier. kicker, is isolated using ByteTrack [29], a precise multi-object tracking network. Context-constrained pre-processing creates an ideal experimental setting. For a kicker (i) at time (t) within interval ([0, T]), the boundin… view at source ↗
Figure 3
Figure 3. The proposed classifier. It combines HAR backbone features and penalty￾kick metadata for enhanced decision-making during running and kicking stages [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Penalty-kick dataset samples. The dataset, sourced from the Internet, ex￾hibits pose, scale, and lighting variability. Clips are edited to focus on the running and kicking stages. HAR backbones, ranging from MViT and Slowfast to C2D and X3D, are tested for kick directi…
Figure 5
Figure 5. Figure 5: Accuracy Analysis. X3D SlowFast Slow MViT C2D I3D SlowFastNLN SlowNLN I3DNLN Architecture 0.48 0.50 0.52 0.54 0.56 0.58 0.60 0.62 F1-Score [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

30 extracted references · 29 canonical work pages

  1. [1]

    Akan, S., Varli, S.: Reidentifying soccer players in broadcast videos using body fea- ture alignment based on pose. In: Int. Conf. on Computing, Networks and Internet of Things. p. 440–444 (2023)

    work page 2023

  2. [2]

    Artiles, J., Hern´ andez-Sosa, D., Santana, O., Lorenzo-Navarro, J., Freire-Obreg´ on, D.: Classifying soccer ball-on-goal position through kicker shooting action. In: Int. Conf. on Pattern Recognition Applications and Methods. pp. 79–90. SciTePress (2024)

    work page 2024

  3. [3]

    In: IEEE Conf

    Carreira, J., Zisserman, A.: Quo Vadis, Action Recognition? A New Model and the Kinetics Dataset. In: IEEE Conf. on Computer Vision and Pattern Recognition. pp. 4724–4733 (2017)

    work page 2017

  4. [4]

    In: IEEE Conf

    Cioppa, A., Deli` ege, A., Van Droogenbroeck, M.: A bottom-up approach based on semantics for the interpretation of the main camera stream in soccer games. In: IEEE Conf. on Computer Vision and Pattern Recognition Workshops. pp. 1846– 184609 (2018)

    work page 2018

  5. [5]

    Dueholm, S.G., Nasrollahi, K., Moeslund, T.B., Droogenbroeck, M.V.: Soccernet-v2 : A dataset and benchmarks for holistic understanding of broadcast soccer videos

    Deli` ege, A., Cioppa, A., aand Jacob V. Dueholm, S.G., Nasrollahi, K., Moeslund, T.B., Droogenbroeck, M.V.: Soccernet-v2 : A dataset and benchmarks for holistic understanding of broadcast soccer videos. In: The IEEE Conf. on Computer Vision and Pattern Recognition Workshops (June 2021)

    work page 2021

  6. [6]

    IEEE Int

    Fan, H., Xiong, B., Mangalam, K., Li, Y., Yan, Z., Malik, J., Feichtenhofer, C.: Multiscale vision transformers. IEEE Int. Conf. on Computer Vision pp. 6804–6815 (2021)

    work page 2021

  7. [7]

    IEEE Conf

    Feichtenhofer, C.: X3D: Expanding Architectures for Efficient Video Recognition. IEEE Conf. on Computer Vision and Pattern Recognition pp. 200–210 (2020)

    work page 2020

  8. [8]

    IEEE Int

    Feichtenhofer, C., Fan, H., Malik, J., He, K.: Slowfast networks for video recogni- tion. IEEE Int. Conf. on Computer Vision (ICCV) pp. 6201–6210 (2018)

    work page 2018

Show all 30 references
  1. [9]

    IEEE Conf

    Feichtenhofer, C., Fan, H., Xiong, B., Girshick, R.B., He, K.: A large-scale study on unsupervised spatiotemporal representation learning. IEEE Conf. on Computer Vision and Pattern Recognition pp. 3298–3308 (2021)

    2021

  2. [10]

    Freire-Obreg´ on, D., Lorenzo-Navarro, J., Santana, O.J., Hern´ andez-Sosa, D., Cas- trill´ on-Santana, M.: Towards cumulative race time regression in sports: I3D Con- vNet transfer learning in ultra-distance running events. In: Int. Conf. on Pattern Recognition (ICPR). pp. 80...

    2022

  3. [11]

    IEEE Access 8, 195062–195073 (2020) 12 D

    Guo, T., Tao, K., Hu, Q., Shen, Y.: Detection of ice hockey players and teams via a two-phase cascaded cnn model. IEEE Access 8, 195062–195073 (2020) 12 D. Freire-Obreg´ on et al

    2020

  4. [12]

    Soft Computing 26(20), 10971–10979 (2022)

    He, X.: Application of deep learning in video target tracking of soccer players. Soft Computing 26(20), 10971–10979 (2022)

    2022

  5. [13]

    Neurocomputing 491, 395–413 (2022)

    Hu, X., Dai, J., Li, M., Peng, C., Li, Y., Du, S.: Online human action detection and anticipation in videos: A survey. Neurocomputing 491, 395–413 (2022)

    2022

  6. [14]

    Sports 6(3) (2018)

    Hunter, A.H., Murphy, S.C., Angilletta, M.J., Wilson, R.S.: Anticipating the di- rection of soccer penalty shots depends on the speed and technique of the kick. Sports 6(3) (2018)

    2018

  7. [15]

    Journal of Biomechanics 141, 111208 (2022)

    Hunter, A.H., Pavlic, T.P., Angilletta, M.J., Wilson, R.S.: Identifying the best strategy for soccer penalty success: A predictive model for optimising behavioural and biomechanical trade-offs. Journal of Biomechanics 141, 111208 (2022)

    2022

  8. [16]

    In: Proceedings of the British Machine Vision Conf

    Jiyang Gao, Z.Y., Nevatia, R.: Red: Reinforced encoder-decoder networks for action anticipation. In: Proceedings of the British Machine Vision Conf. (BMVC). pp. 92.1–92.11 (September 2017)

    2017

  9. [17]

    Kong, Y., Fu, Y.: Human action recognition and prediction: A survey. Int. Journal of Computer Vision 130(5), 1366–1401 (2022)

    2022

  10. [18]

    ArXiv abs/2302.00123 (2023)

    Li, L., Zhang, T., Kang, Z., Zhang, W.H.: Design and implementation of a soccer ball detection system with multiple cameras. ArXiv abs/2302.00123 (2023)

    2023 arXiv

  11. [19]

    IEEE Conf

    Li, Y., Wu, C., Fan, H., Mangalam, K., Xiong, B., Malik, J., Feichtenhofer, C.: Mvitv2: Improved multiscale vision transformers for classification and detection. IEEE Conf. on Computer Vision and Pattern Recognition pp. 4794–4804 (2021)

    2021

  12. [20]

    Journal of Sports Sciences 33(1), 1–10 (2015)

    No¨ el, B., Furley, P., van der Kamp, J., Dicks, M., Memmert, D.: The development of a method for identifying penalty kick strategies in association football. Journal of Sports Sciences 33(1), 1–10 (2015)

    2015

  13. [21]

    In: IEEE Conf

    Parmar, P., Morris, B.T.: What and how well you performed? A multitask learning approach to action quality assessment. In: IEEE Conf. on Computer Vision and Pattern Recognition. pp. 304–313. IEEE (2019)

    2019

  14. [22]

    Multimedia Tools and Applications 82(8), 11395–11415 (2023)

    Santana, O.J., Freire-Obreg´ on, D., Hern´ andez-Sosa, D., Lorenzo-Navarro, J., S´ anchez-Nielsen, E., Castrill´ on-Santana, M.: Facial expression analysis in a wild sporting environment. Multimedia Tools and Applications 82(8), 11395–11415 (2023)

    2023

  15. [23]

    IEEE Transactions on Circuits and Systems for Video Technology 28(5), 1212–1231 (2018)

    Shih, H.C.: A survey of content-aware video analysis for sports. IEEE Transactions on Circuits and Systems for Video Technology 28(5), 1212–1231 (2018)

    2018

  16. [24]

    ArXiv abs/1406.2199 (2014)

    Simonyan, K., Zisserman, A.: Two-stream convolutional networks for action recog- nition in videos. ArXiv abs/1406.2199 (2014)

    2014 arXiv

  17. [25]

    In: IEEE Conf

    Teranishi, M., Fujii, K., Takeda, K.: Trajectory prediction with imitation learn- ing reflecting defensive evaluation in team sports. In: IEEE Conf. on Consumer Electronics (GCCE). pp. 124–125 (2020)

    2020

  18. [26]

    IEEE Transactions on Instrumentation and Measurement 68(8), 2742–2754 (2019)

    Wang, S., Xu, Y., Zheng, Y., Zhu, M., Yao, H., Xiao, Z.: Tracking a golf ball with high-speed stereo vision system. IEEE Transactions on Instrumentation and Measurement 68(8), 2742–2754 (2019)

    2019

  19. [27]

    IEEE Conf

    Wang, X., Girshick, R.B., Gupta, A.K., He, K.: Non-local neural networks. IEEE Conf. on Computer Vision and Pattern Recognition pp. 7794–7803 (2017)

    2017

  20. [28]

    IEEE Transactions on Circuits and Systems for Video Technology 30(12), 4578–4590 (2020)

    Xu, C., Fu, Y., Zhang, B., Chen, Z., Jiang, Y.G., Xue, X.: Learning to score figure skating sport videos. IEEE Transactions on Circuits and Systems for Video Technology 30(12), 4578–4590 (2020)

    2020

  21. [29]

    In: European Conf

    Zhang, Y., Sun, P., Jiang, Y., Yu, D., Yuan, Z., Luo, P., Liu, W., Wang, X.: Byte- Track: Multi-Object Tracking by Associating Every Detection Box. In: European Conf. on Computer Vision (2021)

    2021

  22. [30]

    Journal of Sports Sciences 42(13), 1224–1231 (2024)

    Zheng, R., van der Zijden, B., Janssen, T., van der Kamp and, J.: Goalkeepers benefit from using deceptive actions in the soccer penalty kick. Journal of Sports Sciences 42(13), 1224–1231 (2024)

    2024

Pith tools

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