Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

AugmentGest: Can Random Data Cropping Augmentation Boost Gesture Recognition Performance?

T0 review · 4 major / 5 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 The paper shows that a pipeline of random cropping, rotation, zoom, and brightness/contrast adjustments on 2D skeleton images improves gesture-recognition accuracy on every model and dataset tested, including a new state-of-the-art result…

desk verdict Useful augmentation recipe with plausible gains, but the SOTA claim compares against reproduced baselines, not published numbers, and needs statistical grounding. read the letter →

arxiv 2506.07216 v1 pith:LYQOB5VA submitted 2025-06-08 cs.CV

Nada Aboudeshish , Dmitry Ignatov , Radu Timofte This is my paper · ORCID
classification cs.CV
keywords GestureRecognitionDataAugmentationRandomCroppingSkeletonData-LevelFusionSpatio-temporalIntegritySHREC'17DHG14/28
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 standard image-augmentation recipe—random cropping, rotation, zoom, and brightness/contrast adjustment—applied to 2D skeleton-based gesture images reliably improves gesture-recognition accuracy across architectures and datasets. It reports consistent gains in every model–dataset pair tested: e2eET rises from 96.67% to 98.21% on SHREC'17 14-gesture and from 91.67% to 92.98% on DHG 28-gesture, DD-Net rises from 81.82% to 86.36% on JHMDB, and FPPR-PCD rises from 95.90% to 96.40% on SHREC'17. The proposed pipeline quadruples the training set by generating three augmented versions per sample while preserving the spatio-temporal structure of gestures, which the paper argues is why it beats mixing-based augmentations such as MixUp and CutMix. If correct, this gives a cheap, model-agnostic way to push skeleton-based gesture recognition past previously reported state-of-the-art results.

What carries the argument

The load-bearing object is the 2D spatiotemporal gesture image: a gesture sequence is collapsed into one image where $x$/$y$ positions of the hand joints are spatial coordinates and time is mapped to color intensity (following [31]). The AugmentGest pipeline composes four transformations, $\tilde{x}=T_{\mathrm{brightness,contrast}}(T_{\mathrm{zoom}}(T_{\mathrm{rotation}}(T_{\mathrm{crop}}(x))))$, generating three augmented versions per sample so the dataset grows fourfold. The crop is the step named in the title and the step tied to the spatio-temporal-integrity argument; the ablation attributes a 0.83-point gain to cropping alone and a 0.83-point gain to rotation, with the full combination reaching about 98.2%.

What would settle it

Train e2eET on SHREC'17 14G with crops replaced by deterministic region removals that excise the image area where the hand trajectory starts or ends; if accuracy drops substantially below the random-crop result, the claim that cropping preserves spatio-temporal integrity fails. A second check is comparing random crops of 90–95% against fixed center crops of the same retained area: if accuracies are equal, the crop position, and thus the occlusion-robustness story, is not the active ingredient.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a fixed pipeline of transformations—random crop to 90% or 95% of the image, rotation in $[-15^\circ,15^\circ]$, zoom factor in $[0.90,1.10]$, and brightness/contrast factors in $[0.8,1.2]$—creates training samples that improve accuracy on every model and dataset evaluated. This works because the gestures are represented as static 2D spatiotemporal images in which joint positions are spatial and time is encoded as color intensity, following [31]; cropping changes the visible spatial window without destroying the temporal information channel. The paper reports that the augmented e2eET model surpasses the previously published state of the art on SHREC'17 14-gesture (98.21% vs 96.67% reproduced baseline) and DHG 28-gesture (92.98% vs 91.67%), while DD-Net gains 4.54 percentage points on JHMDB. It further argues that mixing-based augmentations break gesture coherence, showing DD-Net accuracy dropping from 81.82% to roughly 71–73% under MixUp and CutMix.

Load-bearing premise

The whole result rests on the assumption that a random crop keeping only 90–95% of the 2D gesture image does not remove the spatially localized pixels that carry the beginning or end of the hand's trajectory, since those pixels hold part of the color-encoded time information.

Editorial extensions

If this is right

  • Any skeleton-based gesture pipeline that represents gestures as 2D images can adopt AugmentGest without changing its model or loss function.
  • The fourfold-larger training set adds only minutes to e2eET training time on the tested datasets, so the accuracy gains do not come at a large compute cost.
  • Mixing-based augmentations should be avoided for gesture tasks: they drop DD-Net accuracy by roughly ten percentage points where AugmentGest raises it.
  • Fewer training epochs are needed with the augmented data: DD-Net reaches its improved accuracy after 200 epochs instead of 600.
  • The state-of-the-art numbers for SHREC'17 14-gesture and DHG 28-gesture would be moved by data augmentation alone, without architectural changes.

Reading between the lines

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

  • A direct test of the mechanism would be to force the crop to remove the image region containing the hand's first or last color-coded positions; if accuracy falls sharply compared with random crops, the gains depend on preserving trajectory endpoints rather than on augmentation diversity alone.
  • The same pipeline should transfer to other skeleton-to-image encodings that place time in the color or intensity channel; swapping the representation while holding AugmentGest fixed would show whether the crop benefit is tied to this particular encoding.
  • Because training converges in one-third as many epochs on augmented data, AugmentGest likely acts as a regularizer; a control that quadruples the dataset by repeating original samples, with no geometric variation, would isolate the diversity contribution from the size contribution.
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 proposes AugmentGest, a data augmentation pipeline for skeleton-based gesture recognition that applies random cropping, rotation, zooming, and brightness/contrast adjustments to 2D spatiotemporal images derived from skeleton sequences. It evaluates the pipeline on three models (e2eET, FPPR-PCD, DD-Net) and three datasets (SHREC'17, DHG14/28, JHMDB), reporting consistent accuracy gains in Table 2 and claiming state-of-the-art results. The paper also includes an ablation study, a runtime comparison, and a public code release.

Significance. If confirmed, the paper would show that a simple, compute-efficient augmentation pipeline can yield consistent accuracy gains across diverse architectures and datasets, which is practically useful. The strengths are the breadth of evaluation (six model-dataset rows), the inclusion of an ablation study, and the public release of the code. The consistency of the improvements across all rows is a positive signal. However, the paper's SOTA claim is not supported by the reported comparisons, and the lack of error bars leaves the magnitude of the smaller gains uncertain.

major comments (4)
  1. [Section 4.2 / 4.4, Table 2] The central claim that AugmentGest 'surpasses the previously reported SOTA results' (Section 4.4 and the abstract) is not supported by the evidence. Table 2 compares against reproduced baselines only, and Section 4.2 explicitly states that 'the published SOTA results are slightly higher' than these reproduced baselines. For example, the +1.54% gain over the reproduced e2eET baseline (96.67% to 98.21%) does not establish that 98.21% exceeds the published SOTA from [31], since that published number is admitted to be higher. Please add the published SOTA accuracies to Table 2 and compare against them directly, or revise all SOTA claims to 'improvements over reproduced baselines.'
  2. [Section 4.2, Table 2] All accuracy numbers in Table 2 appear to come from single runs with no error bars, multiple seeds, or statistical tests. Several improvements are small (0.47%, 0.5%, 0.72%) and could plausibly be within run-to-run variance, especially since the experiments were run on cloud notebooks with variable hardware. Please report the mean and standard deviation over at least three independent training runs for each configuration, or provide a paired significance test, so the reader can judge whether the reported gains are reliable.
  3. [Section 3.1 and Figure 2] The random cropping step removes 5-10% of each 2D spatiotemporal image (crop size 0.9W x 0.9H or 0.95W x 0.95H). Because temporal information is encoded as color intensity in this representation (as described in Section 3.1 after [31]), a spatial crop can cut off the beginning or end of a gesture trajectory, removing exactly the temporal information the representation encodes. The paper asserts that cropping 'preserves spatio-temporal integrity' but provides no analysis or visualization of which parts of the trajectory survive cropping, or of how frequently a crop severs the trajectory endpoints. Please add such an analysis to substantiate the central motivation for cropping.
  4. [Section 2.1 and References] The e2eET baseline is attributed to reference [31], but the cited paper, 'Development of a lightweight real-time application for dynamic hand gesture recognition' (ICMA 2023), does not obviously correspond to the multi-stream e2eET model with data-level fusion described in Section 2.1. This is load-bearing because the SOTA comparison in Section 4.4 depends on the reproduced baseline matching the published model. Please verify the citation and, if it is incorrect, replace it with the correct e2eET reference and re-check that the reproduced baseline is the actual published SOTA.
minor comments (5)
  1. [Section 3.2 vs. Table 1] The text in Section 3.2 says MixUp and CutMix result in 71.50% and 72.50%, while Table 1 reports 71.02% and 72.73%; please align the numbers.
  2. [Section 4.5 vs. Table 2] The full pipelined accuracy is reported as 98.20% in Table 3 and 98.21% in Table 2; please reconcile the discrepancy.
  3. [Section 4.5, Table 3] The ablation shows that image zoom alone decreases accuracy (96.55% vs. the 96.67% baseline), yet the full combination improves accuracy; please add a brief discussion of why a step that is harmful in isolation contributes positively in the combined pipeline.
  4. [Throughout] There are multiple typographical and grammatical errors, including 'This paper propose' (Section 1), 'showcast' (Section 2.1), '2eET' (Section 4.6), 'JH-MDB' (Figure 1), and the sentence fragment in the abstract beginning 'The e2eET model, recognized as...'.
  5. [Section 4.1] The sentence on DHG14/28 says it contains '2800 sequences across 14 gesture classes,' but DHG28 has 28 classes; please specify which subset has 14 vs. 28 classes.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: AugmentGest is an empirical augmentation study with fixed, pre-specified transformations and held-out evaluation.

full rationale

The paper's central claim is that fixed data-augmentation transformations improve gesture-recognition accuracy. The augmentation pipeline (random cropping to 0.9/0.95 size, rotation in [-15,15] degrees, zoom in [0.90,1.10], brightness/contrast in [0.8,1.2]) is defined independently of the evaluation results; none of these hyperparameters are fitted to any subset of the test data and no equation ties the reported accuracies back to the augmentation parameters by construction. The baseline models e2eET [31], FPPR-PCD [1], and DD-Net [3] are external prior works, not works by the present authors, so the main performance comparison does not reduce to a self-citation chain. The paper does cite several papers by Ignatov and Timofte in the related-work section, but these citations concern generalisation, AutoML, and hyperparameter-tuning contexts and are not load-bearing for the augmentation claim. The most substantive issue is that Section 4.2 explicitly states 'The published SOTA results are slightly higher' than the reproduced baselines, while Section 4.4 claims to have surpassed previously reported SOTA using those same reproduced baselines; this is a benchmark-comparison inconsistency, not a circular derivation. Since the evaluation is on held-out test sets with fixed augmentations and no posterior fitting, the circularity burden is very low and no circular step can be exhibited from the paper's text.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

The paper introduces no new physical or theoretical entities. Its free parameters are the augmentation hyperparameters, all chosen by hand. The main unstated assumptions concern the preservation of spatio-temporal information under cropping and the validity of comparing against reproduced baselines.

free parameters (5)
  • Crop size ratio = 0.9 or 0.95 (randomly chosen)
    Chosen by hand; the reported gains could depend on how much of the gesture image is retained.
  • Rotation angle range = [-15, +15] degrees
    Chosen by hand; the effectiveness of rotation on skeleton images is not analyzed.
  • Zoom factor range = [0.90, 1.10]
    Chosen by hand; the ablation shows zoom alone slightly hurts accuracy (96.55 vs 96.67 baseline).
  • Brightness factor range = [0.8, 1.2]
    Chosen by hand; brightness adjustment alone improves by 0.47% in the ablation.
  • Contrast factor range = [0.8, 1.2]
    Chosen by hand; listed together with brightness in the pipeline, no separate ablation.
assumptions (3)
  • domain assumption The 2D spatiotemporal image representation (depth plus joint coordinates, time encoded as color intensity) retains discriminative gesture information after cropping to 90-95%, rotation up to 15 degrees, zoom, and brightness/contrast changes.
    Invoked in Section 3.1 and 3.2; if cropping removes gesture start or end, the augmentation would destroy temporal information.
  • domain assumption The reproduced baselines for e2eET, FPPR-PCD, and DD-Net align with the original implementations, so relative improvements are meaningful.
    Section 4.2 states reproduced results align but are slightly lower than published SOTA; this assumption is needed for the 'surpassing SOTA' claim.
  • domain assumption Standard deep learning training (Adam, cross-entropy, fixed dataset splits) yields accuracies that are stable enough to compare without error bars.
    The paper provides single-run accuracies (Section 4.4) with no seeds or variance, yet uses small differences to claim improvements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of AugmentGest: Can Random Data Cropping Augmentation Boost Gesture Recognition Performance?." pith.science (2026). https://pith.science/paper/LYQOB5VA

@misc{pith2026250607216,
  author       = {Pith},
  title        = {Pith review of: AugmentGest: Can Random Data Cropping Augmentation Boost Gesture Recognition Performance?},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LYQOB5VA}},
  note         = {Machine review of arXiv:2506.07216}
}
read the original abstract

Data augmentation is a crucial technique in deep learning, particularly for tasks with limited dataset diversity, such as skeleton-based datasets. This paper proposes a comprehensive data augmentation framework that integrates geometric transformations, random cropping, rotation, zooming and intensity-based transformations, brightness and contrast adjustments to simulate real-world variations. Random cropping ensures the preservation of spatio-temporal integrity while addressing challenges such as viewpoint bias and occlusions. The augmentation pipeline generates three augmented versions for each sample in addition to the data set sample, thus quadrupling the data set size and enriching the diversity of gesture representations. The proposed augmentation strategy is evaluated on three models: multi-stream e2eET, FPPR point cloud-based hand gesture recognition (HGR), and DD-Network. Experiments are conducted on benchmark datasets including DHG14/28, SHREC'17, and JHMDB. The e2eET model, recognized as the state-of-the-art for hand gesture recognition on DHG14/28 and SHREC'17. The FPPR-PCD model, the second-best performing model on SHREC'17, excels in point cloud-based gesture recognition. DD-Net, a lightweight and efficient architecture for skeleton-based action recognition, is evaluated on SHREC'17 and the Human Motion Data Base (JHMDB). The results underline the effectiveness and versatility of the proposed augmentation strategy, significantly improving model generalization and robustness across diverse datasets and architectures. This framework not only establishes state-of-the-art results on all three evaluated models but also offers a scalable solution to advance HGR and action recognition applications in real-world scenarios. The framework is available at https://github.com/NadaAbodeshish/Random-Cropping-augmentation-HGR

Figures

Figures reproduced from arXiv: 2506.07216 by the authors.

Figure 2
Figure 2. Results of Image Augmentation with the Proposed [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Framework overview: The proposed augmentation [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 3
Figure 3. DD-Net Validation accuracy trained on different [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. LEMUR 2: Unlocking Neural Network Diversity for AI

    cs.LG 2026-07 conditional novelty 5.5 of 10

    LEMUR 2 releases a multi-generator, multi-task neural-architecture corpus with real-device latency metadata intended as fuel for LLM-driven AutoML.

Reference graph

Works this paper leans on

34 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [31]

    Yusuf and M

    O. Yusuf and M. Habib. Development of a lightweight real- time application for dynamic hand gesture recognition. In 2023 IEEE International Conference on Mechatronics and Automation (ICMA), pages 543–548, 2023. 1, 2, 3, 4, 5

    work page 2023

  2. [1]

    Bigalke and M

    A. Bigalke and M. P. Heinrich. Fusing posture and position representations for point cloud-based hand gesture recogni- tion. In2021 International Conference on 3D Vision (3DV), pages 617–626. IEEE, 2021. 1, 2, 4

    work page 2021

  3. [2]

    DeVries and G

    T. DeVries and G. W. Taylor. Improved regularization of convolutional neural networks with cutout.arXiv preprint arXiv:1708.04552, 2017. 2

    arXiv 2017

  4. [3]

    Y . W. Fan Yang, Sakriani Sakti and S. Nakamura. Make skeleton-based action recognition model smaller, faster and better. InACM International Conference on Multimedia in Asia, 2019. 1, 2, 4

    work page 2019

  5. [4]

    C. Finn, P. Abbeel, and S. Levine. Model-agnostic meta- learning for fast adaptation of deep networks. InProceedings of the 34th International Conference on Machine Learning- Volume 70, pages 1126–1135. JMLR. org, 2017. 2

    work page 2017

  6. [5]

    M. Gado, T. Taliee, M. D. Memon, D. Ignatov, and R. Tim- ofte. Vist-gpt: Ushering in the era of visual storytelling with llms?arXiv preprint arXiv:2504.19267, 2025. 2 5

    arXiv 2025

  7. [6]

    Ganin and V

    Y . Ganin and V . Lempitsky. Unsupervised domain adaptation by backpropagation.International Conference on Machine Learning, 37:1180–1189, 2015. 2

    work page 2015

  8. [7]

    A. T. Goodarzi, R. Kochnev, W. Khalid, F. Qin, T. A. Uzun, Y . S. Dhameliya, Y . K. Kathiriya, Z. A. Bentyn, D. Igna- tov, and R. Timofte. Lemur neural network dataset: Towards seamless automl.arXiv preprint arXiv:2504.10552, 2025. 2

    arXiv 2025

Show all 34 references
  1. [8]

    I. J. Goodfellow, J. Shlens, and C. Szegedy. Explain- ing and harnessing adversarial examples.arXiv preprint arXiv:1412.6572, 2014. 2

    2014 arXiv

  2. [9]

    K. He, X. Zhang, S. Ren, and J. Sun. Identity mappings in deep residual networks, 2016. 2

    2016

  3. [10]

    H. Inoue. Data augmentation by pairing samples for images classification, 2018. 2, 3

    2018

  4. [11]

    Jhuang, J

    H. Jhuang, J. Gall, S. Zuffi, C. Schmid, and M. J. Black. Towards understanding action recognition. InInternational Conf. on Computer Vision (ICCV), pages 3192–3199, Dec

  5. [12]

    S. Ji, W. Xu, M. Yang, and K. Yu. 3d convolutional neural networks for human action recognition.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2013. 2

    2013

  6. [13]

    Kochnev et al

    R. Kochnev et al. Nngpt: Rethinking automl with large lan- guage models.arXiv preprint, 2025. 2

    2025

  7. [14]

    Kochnev, A

    R. Kochnev, A. T. Goodarzi, Z. A. Bentyn, D. Ignatov, and R. Timofte. Optuna vs code llama: Are llms a new paradigm for hyperparameter tuning?arXiv preprint arXiv:2504.06006, 2025. 2

    2025 arXiv

  8. [15]

    Krizhevsky and G

    A. Krizhevsky and G. Hinton. Learning multiple layers of features from tiny images. Technical Report TR-2009, Uni- versity of Toronto, 2009. 2

    2009

  9. [16]

    X. Li, T. Zhao, X. Zeng, M. Luo, K. Jiang, and Z. Sun. Edge computing: A survey on the hardware aspects.IEEE Access, 8:85454–85472, 2020. 2

    2020

  10. [17]

    J. Liu, X. Wang, C. Wang, Y . Gao, and M. Liu. Tem- poral decoupling graph convolutional network for skeleton- based gesture recognition.IEEE Transactions on Multimedia (TMM), 2024. 2

    2024

  11. [18]

    Y . Min, Y . Zhang, X. Chai, and X. Chen. An efficient pointl- stm for point clouds based gesture recognition. InProceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 5761–5770, 2020. 2

    2020

  12. [19]

    Molchanov, X

    P. Molchanov, X. Yang, S. Gupta, K. Kim, S. Tyree, and J. Kautz. Online detection and classification of dynamic hand gestures with recurrent 3d convolutional neural networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 4207–42...

  13. [20]

    Neverova, C

    N. Neverova, C. Wolf, G. W. Taylor, and F. Nebout. Hand pose estimation through semi-supervised and weakly- supervised learning.Computer Vision and Image Under- standing, 139:102–114, 2015. 2

    2015

  14. [21]

    Oikonomidis, N

    I. Oikonomidis, N. Kyriazis, and A. Argyros. Full dof track- ing of a hand interacting with an object by modeling occlu- sions and physical constraints. InIEEE International Con- ference on Computer Vision, pages 2088–2095, 2011. 2

    2011

  15. [22]

    Rupani, D

    B. Rupani, D. Ignatov, and R. Timofte. Exploring the collaboration between vision models and llms for en- hanced image classification.Dimensions, 27(1), 2025. doi:10.13140/RG.2.2.14615.69284. 2

    2025

  16. [23]

    Simonyan and A

    K. Simonyan and A. Zisserman. Two-stream convolutional networks for action recognition in videos.Advances in Neu- ral Information Processing Systems, 2014. 2

    2014

  17. [24]

    Q. D. Smedt, H. Wannous, and J.-P. Vandeborre. Skeleton- based dynamic hand gesture recognition. In2016 IEEE Con- ference on Computer Vision and Pattern Recognition Work- shops (CVPRW), pages 1206–1214, Las Vegas, NV , USA, June 2016. IEEE. 1, 2, 4

    2016

  18. [25]

    Q. D. Smedt, H. Wannous, J.-P. Vandeborre, J. Guerry, B. L. Saux, and D. Filliat. Shrec’17 track: 3d hand gesture recog- nition using a depth and skeletal dataset. In3DOR - 10th Eurographics Workshop on 3D Object Retrieval, pages 1–6, Lyon, France, Apr. 2017. 1, 2, 4

    2017

  19. [26]

    Szegedy, W

    C. Szegedy, W. Liu, Y . Jia, P. Sermanet, S. Reed, D. Anguelov, D. Erhan, V . Vanhoucke, and A. Rabinovich. Going deeper with convolutions. In2015 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 1–9, 2015. 2

    2015

  20. [27]

    Takahashi, T

    R. Takahashi, T. Matsubara, and K. Uehara. Data augmen- tation using random image cropping and patching for deep cnns.IEEE Transactions on Circuits and Systems for Video Technology, 30(9):2917–2931, 2020. 2

    2020

  21. [28]

    C. Wan, M. Proesmans, and L. Van Gool. Handnet: Deep learning for hand pose and hand shape estimation.arXiv preprint arXiv:1612.02401, 2016. 2

    2016 arXiv

  22. [29]

    H. Xiao, K. Rasul, and R. V ollgraf. Fashion-mnist: a novel image dataset for benchmarking machine learning al- gorithms, 2017. 2

    2017

  23. [30]

    Yun et al

    S. Yun et al. Cutmix: Regularization strategy to train strong classifiers with localizable features. 2019. 2, 3

    2019

  24. [32]

    Zagoruyko and N

    S. Zagoruyko and N. Komodakis. Wide residual networks,

  25. [33]

    Zhang, M

    H. Zhang, M. Cisse, Y . N. Dauphin, and D. Lopez-Paz. mixup: Beyond empirical risk minimization.ICLR, 2018. 2, 3

    2018

  26. [34]

    Zhong, L

    Z. Zhong, L. Zheng, G. Kang, S. Li, and Y . Yang. Random erasing data augmentation, 2017. 2 6

    2017

Pith tools

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