Pith. sign in

REVIEW 3 major objections 5 minor 36 references

Graph-Based Deep Learning for Component Segmentation of Maize Plants

T0 review · 3 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 graph neural network enriched with PCA-based geometric features can segment maize plants into leaf, stem, and soil directly from LiDAR point clouds, achieving 93.20% mean IoU on the Pheno4D dataset.

desk verdict Useful engineering combination with clear empirical comparisons, but the evaluation relies on validation folds selected during tuning, so the headline accuracy claims are not yet established. read the letter →

arxiv 2507.00182 v2 pith:JUCAIAHM submitted 2025-06-30 cs.CV

J. I. Ruiz-Martinez , A. Mendez-Vazquez , E. Rodriguez-Tello This is my paper
classification cs.CV
keywords pointcloudsegmentationmaizeplantphenotypinggraphneuralnetworksEdgeConvattentionPCAgeometricfeaturesLiDARprecisionagriculture
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 proposes EdgeGAT, a graph neural network that labels every point in a 3D LiDAR scan of a maize plant as leaf, stem, or soil. The authors claim that combining PCA-derived geometric features with edge-based convolution and graph attention reaches above 80% average IoU and a 93.20% mean IoU on the Pheno4D dataset, outperforming PointNet, DGCNN, GAT, GCN, and GCN-UNet baselines. The motivation is that accurate component segmentation underpins biomass estimation, disease detection, and yield prediction in precision agriculture. The contribution is an extension of existing graph-based point-cloud segmentation methods, not a new problem.

What carries the argument

The load-bearing object is the EdgeGAT layer stack: K-nearest-neighbor graph construction with $k=16$ over spatial coordinates, followed by two EdgeConv blocks whose residual MLPs compute edge features $e_{ij} = \Phi([f_i \| f_j - f_i])$, where $\Phi$ is a residual MLP and $\|$ denotes concatenation. The original 13-dimensional features are concatenated with the second EdgeConv output to form 77-dimensional vertex features, which two multi-head graph attention layers then refine using attention coefficients $\alpha_{ij} = \mathrm{softmax}_j(\sigma(a^\top[Wh_i \| Wh_j]))$. This mechanism lets the model weight neighbors adaptively while retaining directional edge geometry, which the paper argues is essential for separating thin stems, broad leaves, and soil.

What would settle it

Re-run EdgeGAT on the same datasets at native resolution, or at 4,096 and 8,192 points per plant, with identical training; if mean IoU falls below the graph baselines, the reported advantage is an artifact of uniform downsampling. A second check is to evaluate on mature maize point clouds of age similar to the field dataset; if EdgeGAT's margin over GAT shrinks, the Pheno4D result does not transfer across developmental stages.

Watch

Extended reading notes

Core claim

The central discovery is that explicitly modeling local geometry twice improves component segmentation: first by enriching each point's raw xyz coordinates with ten PCA-derived descriptors (normals, curvature, linearity, planarity, scattering, omnivariance, anisotropy, and eigenentropy), and then by passing edge features through attention-weighted aggregation. On the Pheno4D maize dataset, EdgeGAT reaches 93.20% mean IoU and 96.40% accuracy, roughly 11 percentage points above the GAT and DGCNN baselines, and it also exceeds 73% mIoU on a separate field-collected maize dataset. The authors interpret this as evidence that combining edge convolutions with residual connections and graph attention captures both fine-grained local structure and broader contextual relationships in plant point clouds.

Load-bearing premise

The method is only tested on point clouds cut down to 1,024 points each, so the entire reported advantage rests on the assumption that this sparse sample still preserves the distinguishing shape of maize leaves, stems, and soil.

Editorial extensions

If this is right

  • If the claim is correct, LiDAR-based phenotyping can label maize components automatically, reducing the need for manual annotation of individual plants.
  • The method is accurate enough on young maize plants that the paper groups the immature spike with the stem class, suggesting a practical preprocessing rule for early-stage phenotyping.
  • Attention-based graph models outperform point-wise models such as PointNet on this task by a wide margin, pointing to graph structure as the key for fine-grained plant segmentation.
  • The ablation results indicate that PCA-derived geometric features are central: adding normals and curvature raises mIoU from roughly 60% to 67% on the field dataset, and the full 13-dimensional feature set pushes EdgeGAT to 73.44%.
  • EdgeGAT converges faster and reaches higher mean IoU than GAT and DGCNN, implying that the combination of EdgeConv and attention is more than the sum of its parts.

Reading between the lines

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

  • The paper leaves implicit that uniform downsampling to 1,024 points may erase thin leaf tips and small stem fragments; a resolution-ablation study at 4,096 or 8,192 points would show whether the reported margins survive at higher density.
  • Because the Pheno4D plants are younger and structurally simpler than the field dataset, the 93.20% figure likely reflects both the architecture and the relative ease of the data; testing on age-matched plants would separate these factors.
  • The residual-style shortcut that concatenates the original 13-dimensional features to the 77-dimensional representation probably contributes to training stability, but the paper does not ablate this connection, so its exact contribution remains untested.
  • The attention weights from the GAT layers could be visualized to identify which neighbor relations carry the leaf-stem boundary signal, a step toward explaining the model's decisions and diagnosing failure cases.
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 / 5 minor

Summary. The paper proposes EdgeGAT, a graph neural network for semantic segmentation of maize plant point clouds into leaf, stem, and soil classes. The architecture enriches each point with PCA-derived geometric features, constructs a KNN graph, applies two EdgeConv blocks with residual MLPs, and classifies vertices with two GAT layers. The method is evaluated on two LiDAR datasets (Ao and Pheno4D) under five-fold cross-validation against GCN, GAT, DGCNN, PointNet, and GCN-UNet variants. The reported results are a validation mIoU of 73.35% on Ao and 93.20% on Pheno4D, with the abstract claiming an average IoU above 80%.

Significance. If the comparative results hold on truly independent data, the paper would provide an empirically useful architecture for plant organ segmentation from LiDAR, a task relevant to precision agriculture. Strengths include the systematic ablation of PCA-derived geometric features, comparison against several baselines on two real-world datasets, and convergence/training-time analyses. However, the significance is currently limited by the absence of an independent test set, the lack of error bars or repeated-seed results, and the absence of released code or data; the central comparative claim is plausible but not yet established.

major comments (3)
  1. [§4.2, Tables 4–7, Figures 8–13] All quantitative results are validation-fold metrics that were also used for hyperparameter tuning and, apparently, best-epoch selection. Section 4.2 states that the held-out fold is employed for hyperparameter tuning and overfitting detection, and performance is averaged over folds; no independent test split is described or evaluated. This makes the reported margins (e.g., 93.20% vs. 82.12% on Pheno4D in Table 7, and 73.35% vs. 70.54% on Ao in Table 6) optimistically biased estimates of generalization. The paper needs a properly held-out test set or nested cross-validation, reported separately from validation metrics, to support the claim that EdgeGAT outperforms the baselines.
  2. [§4.1, §4.4] The entire evaluation is performed on uniformly downsampled clouds of 1,024 points, although the original LiDAR files contain over 50,000 points. The paper never evaluates at native resolution and never analyzes how segmentation quality varies with the number of points. Since a practical phenotyping tool must process full-resolution scans, the method's utility depends on an unverified assumption that 1,024 points preserve sufficient leaf and stem structure; the authors should report performance at additional densities or at least justify this point count with a sensitivity study.
  3. [Abstract and §4.4] The abstract's 'percentages above 80% in the IoU average' is obtained by averaging the per-dataset mIoU values of 73.35% (Ao) and 93.20% (Pheno4D). This presentation masks the large performance gap between datasets and the fact that one dataset is below 80%. The per-dataset results should be the primary reporting units, with clear statements of the uncertainty across folds; Tables 6 and 7 additionally report single-point estimates with no standard deviation or repeated-seed variability, so the claimed margins over DGCNN and GAT cannot be assessed statistically.
minor comments (5)
  1. [§3.1] There are several typos and notation inconsistencies, including 'Scaterring' for scattering, '13-dimentional' in §3.2, 'clasification' in §3.1, and 'funge' in §2.1; a careful proofread is needed.
  2. [§4.3] The table numbering is inconsistent: Table 2 is captioned 'Table 1: Distribution of classes', and the text sometimes refers to Table 5 before Table 4 is introduced; renumber tables and fix all in-text references.
  3. [Figure 1 caption] The caption mentions 'pretraining' for the Graph Attention Networks stage, but no pretraining procedure is described anywhere in the paper; either remove the term or explain the pretraining step in Section 3.5/3.6.
  4. [§4.4, Figure 11] The text notes that the GAT baseline's validation loss increases during training, which suggests instability or overfitting, but this is not quantified or discussed; because hyperparameters were tuned on the validation folds, this behavior should be examined and reported.
  5. [§5 and §6] The discussion and conclusion claim 'faster convergence' and 'robustness' based mainly on a single run per model; without multiple seeds or error bars, these qualitative claims are not supported by the presented evidence.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the EdgeGAT comparison is an empirical composition of externally sourced modules; validation-set tuning is an evaluation limitation, not a circular derivation.

full rationale

This paper makes no first-principles derivation; its central claim is an empirical comparison of trained segmentation architectures on the Ao and Pheno4D LiDAR datasets. The proposed EdgeGAT is assembled from published external building blocks: PCA-derived geometric features (refs. 20-25), KNN graph construction (refs. 26-27), residual MLPs (ref. 28), EdgeConv (ref. 32), and GAT layers (ref. 16). None of these components is justified by a self-citation, and no cited result is authored by the present authors, so no self-citation or imported-uniqueness chain is load-bearing. The reported mIoU figures (Tables 4-7, Figures 8-13) are validation-fold metrics, and Section 4.2 states that the held-out fold is used 'for hyperparameter tuning and to detect potential overfitting'; this means there is no independent test-set estimate and the comparative margin could be optimistically biased. That is a methodological/test-set limitation, not a definitional equivalence: the model weights are trained only on the training folds, and the validation metrics are measured rather than derived from the inputs. Section 4.1's downsampling to 1,024 points and Section 6's stated generalizability limits are acknowledged assumptions, not circular steps. Therefore the derivation chain is self-contained with respect to circularity, and the appropriate score is 0.

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

The central empirical result rests on several unverified modeling choices: the subsampling level, the graph neighborhood size, the PCA neighborhood definition, and the evaluation protocol. None of these are systematically varied, and the lack of a held-out test set is the most consequential assumption.

free parameters (5)
  • KNN graph neighborhood size k = 16
    Chosen fixed value for graph construction (Section 3.6); no sensitivity analysis is provided.
  • Downsampled point count per plant = 1,024
    Uniform random downsampling from >50k points (Section 4.1); the choice strongly affects the graph resolution and is not validated against full-resolution performance.
  • PCA neighborhood size = not specified
    The local neighborhood used to compute the covariance matrix C (Section 3.1) is never quantified, although all 10 geometric descriptors depend on it.
  • GAT heads and hidden units = 4 heads, 64 hidden units
    Architectural hyperparameters (Section 3.6) chosen without ablation or search.
  • Training hyperparameters (learning rate, epochs, batch size) = 0.001, 100, 8
    Standard choices (Section 3.6); no tuning or sensitivity analysis is reported.
assumptions (5)
  • domain assumption The validation-fold mIoU, averaged over five folds, estimates true generalization performance
    The paper uses the validation fold for hyperparameter tuning (Section 4.2), which typically inflates validation metrics; no independent test set is reported.
  • domain assumption Uniform random downsampling to 1,024 points preserves the discriminative geometric structure of maize plants
    All training and evaluation is on subsampled point clouds (Section 4.1); performance on native LiDAR resolution is not assessed.
  • domain assumption The train/validation partition does not mix points from the same plant across folds
    The split is described only as 'five equally sized subsets' (Section 4.2); if plants are split across folds, the reported accuracy is inflated by leakage.
  • domain assumption The ground-truth labels for the Ao dataset are correct and consistently defined
    The paper says points are 'manually selected and visualized using MeshLab' (Section 4.2) without detailing the labeling protocol or inter-labeler agreement; label provenance for Ao is unclear.
  • standard math Graph neural network training converges and the reported best epochs are representative
    Standard SGD/Adam behavior is assumed; no seeds or run-to-run variance are reported.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Graph-Based Deep Learning for Component Segmentation of Maize Plants." pith.science (2026). https://pith.science/paper/JUCAIAHM

@misc{pith2026250700182,
  author       = {Pith},
  title        = {Pith review of: Graph-Based Deep Learning for Component Segmentation of Maize Plants},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JUCAIAHM}},
  note         = {Machine review of arXiv:2507.00182}
}
read the original abstract

In precision agriculture, one of the most important tasks when exploring crop production is identifying individual plant components. There are several attempts to accomplish this task by the use of traditional 2D imaging, 3D reconstructions, and Convolutional Neural Networks (CNN). However, they have several drawbacks when processing 3D data and identifying individual plant components. Therefore, in this work, we propose a novel Deep Learning architecture to detect components of individual plants on Light Detection and Ranging (LiDAR) 3D Point Cloud (PC) data sets. This architecture is based on the concept of Graph Neural Networks (GNN), and feature enhancing with Principal Component Analysis (PCA). For this, each point is taken as a vertex and by the use of a K-Nearest Neighbors (KNN) layer, the edges are established, thus representing the 3D PC data set. Subsequently, Edge-Conv layers are used to further increase the features of each point. Finally, Graph Attention Networks (GAT) are applied to classify visible phenotypic components of the plant, such as the leaf, stem, and soil. This study demonstrates that our graph-based deep learning approach enhances segmentation accuracy for identifying individual plant components, achieving percentages above 80% in the IoU average, thus outperforming other existing models based on point clouds.

Figures

Figures reproduced from arXiv: 2507.00182 by the authors.

Figure 1
Figure 1. Our proposed model is a graph based EdgeGAT architecture. The first stage [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the EdgeConv module for feature enhancement. The Residual [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. Three-class point cloud representation of a maize plant in Ao dataset.(a) The [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Three-class point cloud representation of a maize plant used for model training [PITH_FULL_IMAGE:figures/full_fig_p019_4.png]
Figure 5
Figure 5. Figure 5: Validation mIoU results across different feature configurations for GCN, GAT, [PITH_FULL_IMAGE:figures/full_fig_p020_5.png]
Figure 6
Figure 6. Figure 6: Validation mIoU results across different feature configurations for GCN, GAT, [PITH_FULL_IMAGE:figures/full_fig_p021_6.png]
Figure 7
Figure 7. Figure 7: Validation Loss of GCN, GAT, UNet, UNet2, and EdgeGAT using 13 PCA [PITH_FULL_IMAGE:figures/full_fig_p022_7.png]
Figure 8
Figure 8. Figure 8: Validation mIoU of GCN, GAT, UNet, UNet2, and EdgeGAT using 13 PCA [PITH_FULL_IMAGE:figures/full_fig_p023_8.png]
Figure 9
Figure 9. Figure 9: Validation Loss of GCN, GAT, UNet, UNet2, and EdgeGAT using 13 PCA [PITH_FULL_IMAGE:figures/full_fig_p024_9.png]
Figure 10
Figure 10. Figure 10: Validation mIoU of GCN, GAT, UNet, UNet2, and EdgeGAT using 13 PCA [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 11
Figure 11. Figure 11: Validation Loss of point and graph-based models using 13-PCA features: Edge [PITH_FULL_IMAGE:figures/full_fig_p026_11.png]
Figure 12
Figure 12. Figure 12: Validation mIoU of point and graph-based models using 13-PCA features: [PITH_FULL_IMAGE:figures/full_fig_p027_12.png]
Figure 13
Figure 13. Figure 13: Validation mIoU curves for EdgeGAT, DGCNN, PointNet, and GAT on the [PITH_FULL_IMAGE:figures/full_fig_p029_13.png]
Figure 14
Figure 14. Figure 14: Visualization of point cloud classification: (a) Ground truth of maize plant, [PITH_FULL_IMAGE:figures/full_fig_p030_14.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

36 extracted references · 21 canonical work pages

  1. [1]

    Carroll, C

    C. Carroll, C. Carter, R. Goodhue, C.-Y. C. L. Lawell, Crop Disease and Agricultural Productivity, Tech. Rep. w23513, National Bureau of Eco- nomic Research, Cambridge, MA (Jun. 2017). doi:10.3386/w23513. URL http://www.nber.org/papers/w23513.pdf

    work page doi:10.3386/w23513 2017

  2. [2]

    Zhang, et al., Biomass and crop height estimation of different crops using uav-based 3d point cloud, Remote Sensing 12 (1) (2020) 17

    C. Zhang, et al., Biomass and crop height estimation of different crops using uav-based 3d point cloud, Remote Sensing 12 (1) (2020) 17. 36

    work page 2020

  3. [3]

    URL https://geopard.tech/blog/crop-identification/

    Geopard, The power of crop identification, Geopard Tech Blog (2022). URL https://geopard.tech/blog/crop-identification/

    work page 2022

  4. [4]

    URL https://www.picsellia.com/post/ precision-agriculture-computer-vision-crop-health-monitoring

    Picsellia, Precision agriculture: Using computer vision for crop health monitoring, Picsellia Blog (2024). URL https://www.picsellia.com/post/ precision-agriculture-computer-vision-crop-health-monitoring

    work page 2024

  5. [5]

    Mahlein, Plant disease detection by imaging sensors – parallels and specific demands for precision agriculture and plant phenotyping, Plant Disease 100 (2) (2016) 241–251

    A.-K. Mahlein, Plant disease detection by imaging sensors – parallels and specific demands for precision agriculture and plant phenotyping, Plant Disease 100 (2) (2016) 241–251

    work page 2016

  6. [6]

    URL https://geopard.tech/blog/predicting-crop-yield-with-remote-sensing-data/

    Geopard, Crop yield prediction with remote sensing data in precision agriculture, Geopard Tech Blog (2022). URL https://geopard.tech/blog/predicting-crop-yield-with-remote-sensing-data/

    work page 2022

  7. [7]

    Bertolini, D

    M. Bertolini, D. Mezzogori, M. Neroni, F. Zammori, Machine learning for industrial applications: A comprehensive literature review, Expert Systems with Applications 175 (2021) 114820. doi:10.1016/j.eswa.2021.114820. URL https://www.sciencedirect.com/science/article/pii/ S095741742100261X

    work page arXiv 2021

  8. [8]

    S. O. Ara´ ujo, R. S. Peres, J. C. Ramalho, F. Lidon, J. Barata, Ma- chine learning applications in agriculture: Current trends, challenges, and future perspectives, Agronomy 13 (12) (2023). doi:10.3390/ agronomy13122976. URL https://www.mdpi.com/2073-4395/13/12/2976

    work page 2023

Show all 36 references
  1. [9]

    O’Shea, R

    K. O’Shea, R. Nash, An introduction to convolutional neural networks (2015). arXiv:1511.08458. URL https://arxiv.org/abs/1511.08458

    2015 arXiv

  2. [10]

    Japes, J

    B. Japes, J. Mack, F. Rist, K. Herzog, R. T¨ opfer, V. Steinhage, Multi- View Semantic Labeling of 3D Point Clouds for Automated Plant Phe- notyping, arXiv:1805.03994 [cs] (May 2018). URL http://arxiv.org/abs/1805.03994

    2018 arXiv

  3. [11]

    Y. Guo, H. Wang, Q. Hu, H. Liu, L. Liu, M. Bennamoun, Deep learning for 3d point clouds: A survey, IEEE Transactions on Pat- 37 tern Analysis and Machine Intelligence 43 (12) (2021) 4338–4364. doi: 10.1109/TPAMI.2020.3005434

    2021

  4. [12]

    C. R. Qi, H. Su, K. Mo, L. J. Guibas, Pointnet: Deep learning on point sets for 3d classification and segmentation, arXiv preprint arXiv:1612.00593 (2016)

    2016 arXiv

  5. [14]

    Y. Guo, H. Wang, Q. Hu, Q. Liu, L. J. Guibas, Deep learning on point clouds: A survey, IEEE Transactions on Neural Networks and Learning Systems 32 (11) (2020) 4333–4352

    2020

  6. [15]

    Y. Wang, Y. Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, J. M. Solomon, Dynamic graph cnn for learning on point clouds (2019). arXiv:1801. 07829. URL https://arxiv.org/abs/1801.07829

    2019 arXiv

  7. [16]

    Veliˇ ckovi´ c, G

    P. Veliˇ ckovi´ c, G. Cucurull, A. Casanova, A. Romero, P. Lio, Y. Bengio, Graph attention networks, arXiv preprint arXiv:1710.10903 (2017). URL https://arxiv.org/abs/1710.10903

    2017 arXiv

  8. [17]

    J. A. Bondy, U. S. R. Murty, Graph Theory with Applications, Elsevier, New York, 1976

    1976

  9. [18]

    T. N. Kipf, M. Welling, Semi-supervised classification with graph con- volutional networks, arXiv preprint arXiv:1609.02907 (2016)

    2016 arXiv

  10. [19]

    H. Gao, S. Ji, Graph u-nets, in: Proceedings of the 36th International Conference on Machine Learning, PMLR, 2019, pp. 2083–2092

    2019

  11. [20]

    Hotelling, Analysis of a complex of statistical variables into principal components, Journal of Educational Psychology 24 (6) (1933) 417–441

    H. Hotelling, Analysis of a complex of statistical variables into principal components, Journal of Educational Psychology 24 (6) (1933) 417–441. doi:10.1037/h0071325

    1933 doi

  12. [21]

    Pearson, On lines and planes of closest fit to systems of points in space, Philosophical Magazine 2 (11) (1901) 559–572

    K. Pearson, On lines and planes of closest fit to systems of points in space, Philosophical Magazine 2 (11) (1901) 559–572. doi:10.1080/ 14786440109462720. 38

    1901

  13. [22]

    I. T. Jolliffe, Principal Component Analysis, 2nd Edition, Springer Series in Statistics, Springer, New York, 2002. doi:10.1007/b98835

    2002 doi

  14. [23]

    Weinmann, B

    M. Weinmann, B. Jutzi, C. Mallet, Feature extraction and classification of 3D point clouds for mapping of urban facades, ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences II- 3/W4 (2015) 57–64. doi:10.5194/isprsannals-II-3-W4-57-2015

    2015 doi

  15. [24]

    Pauly, M

    M. Pauly, M. Gross, L. P. Kobbelt, Efficient simplification of point- sampled surfaces, Proceedings of the conference on Visualization (2002) 163–170doi:10.1109/VISUAL.2002.1183788

    2002 arXiv

  16. [25]

    Hackel, J

    T. Hackel, J. D. Wegner, K. Schindler, Fast semantic segmentation of 3d point clouds with strongly varying density, in: ISPRS annals of the pho- togrammetry, remote sensing and spatial information sciences, Vol. 3, 2016, pp. 177–184. doi:10.5194/isprs-annals-III-3-177-2016

    2016 doi

  17. [26]

    T. M. Cover, P. E. Hart, Nearest neighbor pattern classification, IEEE Transactions on Information Theory 13 (1) (1967) 21–27

    1967

  18. [27]

    C. R. Qi, L. Yi, H. Su, L. J. Guibas, Pointnet++: Deep hierarchical feature learning on point sets in a metric space (2017). arXiv:1706. 02413. URL https://arxiv.org/abs/1706.02413

    2017 arXiv

  19. [28]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, Proceedings of the IEEE conference on computer vision and pattern recognition (CVPR) (2016) 770–778

    2016

  20. [29]

    Y. Rong, W. Huang, T. Xu, J. Huang, Dropedge: Towards deep graph convolutional networks on node classification, in: International Confer- ence on Learning Representations (ICLR), 2020. URL https://arxiv.org/abs/1907.10903

    2020 arXiv

  21. [30]

    K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recog- nition, in: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2016, pp. 770–778

    2016

  22. [31]

    Tolstikhin, N

    I. Tolstikhin, N. Houlsby, A. Kolesnikov, L. Beyer, X. Zhai, T. Un- terthiner, J. Yung, A. Steiner, F. Heide, et al., Mlp-mixer: An all-mlp architecture for vision, arXiv preprint arXiv:2105.01601 (2021). 39

    2021 arXiv

  23. [32]

    Y. Wang, Y. Sun, Z. Liu, S. E. Sarma, M. M. Bronstein, J. M. Solomon, Dynamic graph cnn for learning on point clouds, in: ACM Transactions on Graphics (TOG), Vol. 38, ACM, 2019, pp. 1–12

    2019

  24. [33]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, I. Polosukhin, Attention is all you need, in: Advances in Neural Information Processing Systems, 2017

    2017

  25. [34]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization (2017). arXiv:1412.6980. URL https://arxiv.org/abs/1412.6980

    2017 arXiv

  26. [35]

    Z. Ao, F. Wu, S. Hu, Y. Sun, Y. Su, Q. Guo, Q. Xin, Automatic segmentation of stem and leaf components and individual maize plants in field terrestrial lidar data using convolutional neural networks, The Crop Journal 10 (5) (2022) 1239–1250, crop phenotyping studies with appli...

    2022 doi

  27. [36]

    Schunck, F

    D. Schunck, F. Magistri, R. A. Rosu, A. Cornelißen, N. Chebrolu, S. Paulus, J. L´ eon, S. Behnke, C. Stachniss, H. Kuhlmann, L. Kling- beil, Pheno4d: A spatio-temporal dataset of maize and tomato plant point clouds for phenotyping and advanced plant analysis, PLOS ONE 16 (8) (...

    2021 doi

  28. [37]

    Cignoni, M

    P. Cignoni, M. Callieri, M. Corsini, M. Dellepiane, F. Ganovelli, G. Ranzuglia, MeshLab: an Open-Source Mesh Processing Tool, in: V. Scarano, R. D. Chiara, U. Erra (Eds.), Eurographics Italian Chap- ter Conference, The Eurographics Association, 2008. doi:10.2312/ LocalChapterE...

    2008

Pith tools

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