Pith. sign in

REVIEW 6 major objections 10 minor 9 references

Self-Supervised Pretraining Improves Cross-Site and Cross-Scale Robustness of Point Cloud Leaf-Wood Segmentation

T0 review · 6 major / 10 minor · reviewed 2026-07-09 · glm-5.2

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

Pith's one-line read Pretraining on tree shapes halves wood volume error in forest scans

desk verdict SSL pretraining for TLS leaf-wood segmentation works, but the KR test-set gains are inflated by pretraining overlap with test trees read the letter →

arxiv 2607.06948 v1 pith:O7DQ2KNI submitted 2026-07-08 cs.CV cs.AI

Heeju Mun , Tackang Yang , Yunsoo Nam , Changhyun Choi This is my paper · ORCID
classification cs.CVcs.AI
keywords modelpointsegmentationacrosscloudsimprovedleaf-woodmethods
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

Separating leaf points from wood points in terrestrial laser scans of forests is a prerequisite for estimating tree volume and biomass, but existing methods lose accuracy when applied across different forest types, climates, and scanning densities. This paper shows that an encoder pretrained via masked reconstruction on a mix of synthetic 3D objects and 2,400 tree point clouds, then fine-tuned for leaf-wood segmentation, consistently outperforms both algorithmic and fully supervised baselines across six sites in four countries. The key finding is that even a small amount of domain-relevant data added to pretraining (roughly 4% of the pool) drives disproportionate gains in wood IoU and cross-site consistency, and that these segmentation gains propagate downstream: wood volume estimates from quantitative structure models fitted to the segmented wood points achieve the lowest mean absolute error (2.40 m³) among all tested methods, less than half that of algorithmic baselines.

What carries the argument

Point-M2AE, a masked autoencoder for point clouds that reconstructs masked regions using Chamfer Distance loss, pretrained on ShapeNet-55 augmented with tree point clouds, then fine-tuned with a PointNet++-style segmentation head. At inference, recursive voxel subdivision partitions large or dense point clouds into fixed-size inputs without architecture change, enabling the same model to operate on both individual trees and full forest plots.

What would settle it

If a model pretrained on ShapeNet-55 alone (without tree data) matched or exceeded the ShapeNet+KR-pretrained model on cross-site wood IoU and volumetric MAE, the central claim that domain-relevant pretraining data drives the improvement would be undermined.

Watch

Extended reading notes

Core claim

Adding 2,400 tree point clouds (a 4% increase) to a synthetic-object pretraining pool yields wood IoU gains exceeding half of the total improvement from pretraining itself, and these gains transfer to unseen sites and to downstream volumetric estimation. The mechanism is that masked-reconstruction pretraining shifts the encoder toward preserving fine woody structures in dense canopy regions where leaf and wood points intermingle, reducing the systematic wood-to-leaf misclassification that propagates into volume overestimation or underestimation.

Load-bearing premise

The claim that adding tree point clouds to pretraining drives the improvement rests on a confound: the tree data was always combined with synthetic objects during pretraining, never tested alone, so the independent contribution of tree geometry cannot be separated from the combined effect. Additionally, the same tree point clouds used for pretraining augmentation were also used for fine-tuning, which means the pretraining stage may have already seen the fine-tuning domain.

Editorial extensions

If this is right

  • Forest biomass inventories that rely on terrestrial laser scanning could adopt SSL-pretrained segmentation as a drop-in preprocessing step, reducing site-specific annotation requirements and improving volume estimates in tropical forests where destructive calibration data are scarce.
  • The disproportionate gain from a small domain-specific pretraining augmentation suggests that pretraining data curation for forestry 3D tasks should prioritize structural relevance over dataset volume, which has implications for other point cloud tasks in ecology and remote sensing.
  • Recursive voxel subdivision without retraining provides a practical pathway for deploying fixed-input-size architectures on operationally variable point clouds, a pattern that could extend to urban scanning, infrastructure inspection, or any domain where point density varies by orders of magnitude.
  • The finding that improved fine-branch recovery can increase volumetric error against destructive references that exclude small branches highlights a mismatch between segmentation quality and the ground truth used for evaluation, motivating reference standards that explicitly account for branch-size thresholds.
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

6 major / 10 minor

Summary. This manuscript investigates whether self-supervised pretraining (SSL) improves leaf-wood segmentation of terrestrial laser scanning (TLS) point clouds. The authors pretrain Point-M2AE on ShapeNet-55 augmented with 2,400 individual tree point clouds (KR dataset), then fine-tune for leaf-wood segmentation using voxel-based input partitioning. They evaluate on a multi-site benchmark (58 trees across 6 sites in 4 countries), at plot level, and on a downstream QSM-based volume estimation task (28 tropical trees). The pretrained model improves wood IoU, reduces cross-site variation, and achieves the lowest volume estimation error among compared methods. The central claim that SSL pretraining improves cross-site generalization is supported by the external benchmark, though the specific contribution of adding KR tree data to pretraining is confounded by data reuse between the pretraining pool and the fine-tuning/test split.

Significance. The paper addresses a practical and important problem: leaf-wood segmentation accuracy varies across forest types and sites, and existing methods rely on handcrafted geometric features with limited transferability. Applying SSL pretraining to this task is novel and well-motivated. The recursive voxel subdivision strategy enabling cross-scale inference without architecture change is a practical contribution. The multi-site benchmark spanning temperate, boreal, and tropical forests is broader than most prior evaluations. The public release of code, pretrained weights, and fine-tuned models is a notable strength that enhances reproducibility. The downstream volume estimation evaluation against destructive measurements provides end-to-end validation that is rare in this subfield.

major comments (6)
  1. §2.1, §2.2, Table 2: The KR dataset is used both for SSL pretraining augmentation (all 2,400 trees) and for fine-tuning/testing (7:2:1 split). The KR test set trees (10% of KR) therefore appear in the ShapeNet+KR pretraining pool but not in the ShapeNet-only pretraining pool. Since SSL pretraining learns representations from raw point coordinates without labels, the ShapeNet+KR encoder has encoded the test trees' geometry during pretraining, creating an asymmetric representation advantage on the KR test set. This inflates the primary quantitative evidence for adding tree data to pretraining (Table 2: wood IoU gains of 2.8-4.1 points on KR test set). The authors acknowledge in §4.3 that KR is always combined with ShapeNet, but do not discuss this specific test-set leakage from pretraining. The external benchmark (Table 3, 58 trees never seen during pretraining or fine-tuning) shows a much
  2. §2.1, §2.2, Table 2: The KR dataset is used both for SSL pretraining augmentation (all 2,400 trees) and for fine-tuning/testing (7:2:1 split). The KR test set trees (10% of KR) therefore appear in the ShapeNet+KR pretraining pool but not in the ShapeNet-only pretraining pool. Since SSL pretraining learns representations from raw point coordinates without labels, the ShapeNet+KR encoder has encoded the test trees' geometry during pretraining, creating an asymmetric representation advantage on the KR test set. This inflates the primary quantitative evidence for adding tree data to pretraining (Table 2: wood IoU gains of 2.8-4.1 points on KR test set). The authors acknowledge in §4.3 that KR is always combined with ShapeNet, but do not discuss this specific test-set leakage from pretraining. The external benchmark (Table 3, 58 trees never seen during pretraining or fine-tuning) shows a much
  3. smaller margin (wood IoU 78.5% vs 76.7%, a 1.8-point gap), and the volume estimation results (Table 5) show essentially no difference between ShapeNet-only and ShapeNet+KR (MAE 2.45 vs 2.40 m³). The authors should either (a) exclude KR test trees from the pretraining pool and re-evaluate, or (b) explicitly discuss this as a limitation and reframe the KR test set results accordingly, emphasizing the external benchmark as the primary evidence for cross-site generalization.
  4. §2.3, Tables 2-5: No statistical significance testing is reported for any comparison. The external benchmark differences between ShapeNet+KR and ShapeNet-only (Table 3: wood IoU 78.5% vs 76.7%) and between ShapeNet+KR and PointTransformer (Table 3: wood IoU 78.5% vs 74.4%) are based on 58 trees across 6 sites, with some sites having only 4-6 trees (Table 1: SWDEU-NL has 4 trees, SWDEU-BL has 6 trees, FIN-NL has 12 trees). With such small per-site sample sizes, it is unclear whether the reported differences are statistically meaningful or within noise. The authors should report confidence intervals or significance tests for the key cross-site comparisons, or at minimum acknowledge this limitation.
  5. §2.1, Table 1: The benchmark dataset is small and imbalanced across sites. The 58 individual trees are distributed across 6 sites with 4-15 trees per site, and the plot-level evaluation uses only 1 plot per leaf type (CAM-BL-PL: 15 trees, FIN-NL-PL: 15 trees). The plot-level results (Table 4) are thus based on n=1 plot per leaf type, making it difficult to assess whether the reported plot-level performance generalizes. The authors should acknowledge this as a limitation in §4.3 and clarify that plot-level conclusions are preliminary.
  6. §3.3, Table 4: For the needleleaf plot (FIN-NL-PL), the ShapeNet+KR-pretrained model performs worse than the ShapeNet-pretrained model across all metrics (e.g., mIoU 77.7% vs 78.5%, wood IoU 64.3% vs 65.2%). The authors attribute this to increased leaf-to-wood misclassification, but this result partially contradicts the central claim that adding KR tree data to pretraining improves segmentation. The abstract states 'mIoU of 84.7% for broadleaf and 77.7% for needleleaf plots' without noting that the ShapeNet-only model achieves higher mIoU (78.5%) on the needleleaf plot. The abstract should be amended to accurately reflect this mixed result.
minor comments (10)
  1. §2.2, Eq. (1): The Chamfer Distance loss is labeled as Equation (1), but the cross-entropy loss in the same section is also labeled as Equation (2). However, in §2.1, the truncated-cone formula is also labeled as Equation (1). The equation numbering is inconsistent across sections.
  2. Abstract: The abstract states the pretrained model improved wood IoU 'from 60.5% to 70.0% for needleleaf and from 69.7% to 76.3% for broadleaf trees.' These are KR test set results (Table 2), which are subject to the pretraining data leakage concern raised in Major Comment 1. The abstract should clarify that these are in-domain results, or lead with the external benchmark results.
  3. §2.1: The country codes in the benchmark description are introduced without a key. While Table 1 provides the mapping, the main text in §2.1 refers to 'SWDEU-BL,' 'DEU-BL,' 'FIN-NL,' etc. without immediately explaining the naming convention. A brief note on first use would improve readability.
  4. §2.3: The TreeQSM cover set parameters are listed as 'PatchDiam1 = 1, PatchDiam2Min = 1, PatchDiam2Max = 1.' It would help to briefly justify why the default parameters were used and whether any sensitivity analysis was performed.
  5. Figure 2: The figure caption refers to panels (a)-(l) but the description of zoomed panels is somewhat unclear. Consider explicitly stating which panels are zoomed views and which show the full tree.
  6. §4.1: The statement 'Deep learning models are generally expected to improve logarithmically with pretraining data volume (Sun et al., 2017)' is used to argue that a 4% data increase should yield small gains. This scaling law applies to supervised learning with labeled data; its applicability to SSL pretraining with unlabeled data is less established. The argument could be softened.
  7. Table 5: The relative MAE percentages are shown in parentheses but the column headers do not indicate these are relative MAE values. Adding '%' to the column headers or a footnote would improve clarity.
  8. §2.1: The reference to 'Seely et al. (2026)' appears to be a future-dated publication. If this is a preprint or forthcoming paper, the citation format should be clarified.
  9. Figure A2: The y-axis label 'patch coverage' is defined as 'the product of patch size and patch number' but the units or typical ranges are not immediately clear from the figure. A brief note in the caption would help.
  10. §3.4: One tree (MDD01_006) from PER was excluded due to unstable QSM reconstruction. The exclusion criteria (SD = 2.70 m³, 3.6× the mean SD) are reasonable, but it would be useful to report how the results change if this tree is included, to confirm the exclusion does not materially affect the conclusions.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found; the paper is an empirical study with standard metrics evaluated against external benchmarks

full rationale

This paper is an empirical study, not a derivation chain. The metrics (IoU, OA, MAE) are standard definitions computed independently of the model's training procedure. No 'prediction' reduces to a fitted input by construction. The reader's concern about KR data appearing in both the SSL pretraining pool and the fine-tuning/test split is a legitimate data-leakage concern (correctness risk), but it is not circularity: the test-set IoU is not defined in terms of the pretraining objective, and the model's predictions on held-out fine-tuning test data are not forced by construction. The paper also evaluates against fully external benchmarks (6 sites, 58 trees, and 28 tropical trees for volume estimation) that were never seen during pretraining or fine-tuning, providing independent evidence. No self-citation chain is load-bearing: Point-M2AE, TreeQSM, and all baselines are external. No uniqueness theorem or ansatz is invoked. The derivation is self-contained.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new mathematical entities, physical particles, or postulated objects. All architectural components (Point-M2AE, PointNet++ feature propagation, TreeQSM) are from prior work. The free parameters are engineering choices (thresholds, split ratios, epoch counts) rather than theoretically derived constants.

free parameters (6)
  • Voxel maximum point threshold = Empirically optimized (16,384 / 32,768 / 65,336 / 131,072 tested)
    Controls recursive voxel subdivision at inference; chosen based on segmentation accuracy (Appendix Figure A2).
  • Patch configuration (patch size × number of patches) = Empirically optimized
    Jointly optimized with voxel threshold; affects segmentation quality at inference.
  • Voxel partition count for fine-tuning = 8 (2×2×2)
    Fixed to meet batch size and GPU memory constraints during training.
  • TreeQSM cover set parameters = PatchDiam1=1, PatchDiam2Min=1, PatchDiam2Max=1
    Default parameters used for all QSM reconstructions; not tuned per method.
  • Pretraining epochs = 300
    Inherited from Point-M2AE original settings.
  • Fine-tuning train/val/test split = 7:2:1
    Standard split ratio chosen for KR dataset.
assumptions (5)
  • domain assumption Point-M2AE architecture and its reported hyperparameters are appropriate for tree point clouds without modification.
    §2.2: 'We retained the original Point-M2AE architecture and hyperparameters, which were reported to perform well across multiple downstream tasks.'
  • domain assumption Chamfer Distance loss on masked point reconstruction is a suitable pretext task for learning representations transferable to leaf-wood segmentation.
    §2.2: The MAE pretext task is assumed to learn representations useful for the downstream classification task.
  • domain assumption TreeQSM v2.0 with default parameters provides a fair and consistent volumetric reconstruction across different wood point inputs.
    §2.3: Default cover set parameters used for all methods; assumes QSM fitting does not introduce method-dependent bias.
  • domain assumption The 7:2:1 split of the KR dataset provides sufficient test set size (19 needleleaf, 221 broadleaf) for reliable performance estimation.
    §2.2: The split is used without power analysis or cross-validation.
  • domain assumption Visual inspection is a valid method for excluding trees with annotation quality issues from the benchmark.
    §2.1: Trees excluded based on visual inspection for sparsity, label inconsistency, or incompleteness.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Self-Supervised Pretraining Improves Cross-Site and Cross-Scale Robustness of Point Cloud Leaf-Wood Segmentation." pith.science (2026). https://pith.science/paper/O7DQ2KNI

@misc{pith2026260706948,
  author       = {Pith},
  title        = {Pith review of: Self-Supervised Pretraining Improves Cross-Site and Cross-Scale Robustness of Point Cloud Leaf-Wood Segmentation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/O7DQ2KNI}},
  note         = {Machine review of arXiv:2607.06948}
}
abstract

The accuracy of existing leaf-wood segmentation methods for tree point clouds varies across forest types and sites. Self-supervised learning (SSL) on point clouds has improved the generalization of deep learning models for forestry point cloud tasks, including biomass regression and individual tree segmentation, but its applicability to leaf-wood segmentation remains untested. In this study, we pretrained Point-M2AE, a widely used SSL architecture for point clouds, on ShapeNet-55 augmented with 2,400 individual tree point clouds. For fine-tuning and inference, we used recursive voxel subdivision to handle the wide variation in point density across inputs, allowing the same model to operate at both individual-tree and plot scales without architecture change. Compared to the model without pretraining, the pretrained model improved wood IoU from 60.5% to 70.0% for needleleaf and from 69.7% to 76.3% for broadleaf trees. On a benchmark spanning four countries across three climatic zones, the pretrained model achieved the smallest cross-site variation and highest overall performance among compared methods (LeWos, CWLS, and PointTransformer). Plot-level segmentation maintained accuracy comparable to individual-tree performance, with mIoU of 84.7% for broadleaf and 77.7% for needleleaf plots, showing that the model generalizes across scales without additional finetuning. As a downstream test in tropical forests, where dense canopies make segmentation challenging, we applied our model and a quantitative structure model to estimate wood volume for 28 trees from Guyana, Indonesia, and Peru to assess whether the segmentation improvements from SSL pretraining translate into improved downstream performance. The resulting volume estimates achieved the lowest error among all methods tested (MAE = 2.40 m$^3$), less than half that of algorithmic baselines (LeWos: 5.94 m$^3$; CWLS: 5.27 m$^3$).

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

9 extracted references · 9 canonical work pages

  1. [1]

    Introduction Terrestrial Laser Scanning (TLS) provides three-dimensional point clouds of forest, which is widely used to characterize tree structure and estimate aboveground biomass (AGB) (Å kerblom & Kaitaniemi, 2021; Dassot et al., 2011, 2012; Liang et al., 2016) . At the individual -tree level, basic parameters such as diameter at breast height (DBH) a...

    work page 2021

  2. [2]

    Data Scale Dataset Size Source Purpose - ShapeNet-55 57,448 objects Chang et al

    Method 2.1. Data Scale Dataset Size Source Purpose - ShapeNet-55 57,448 objects Chang et al. (2015) • SSL pretraining Individual Tree-level KR-BL 2,199 trees AI-Hub (2023) • SSL pretraining • Leaf-wood segmentation fine - tuning KR-NL 201 trees SWDEU-BL 6 trees Weiser et al. (2024) • Individual tree -level leaf -wood segmentation evaluation SWDEU-NL 4 tre...

    work page 2015

  3. [3]

    Buttresses were modelled as triangular plates with V = 1 2 𝐿 ∙ 𝑊 ∙ 𝐻 from the recorded length, width, and height

    (1) Where L is the segment length and D 1, D 2 are the diameters at the two endpoints. Buttresses were modelled as triangular plates with V = 1 2 𝐿 ∙ 𝑊 ∙ 𝐻 from the recorded length, width, and height. Branches were calculated as frustums from the recorded base and top diameters and segment length. For and IND, the original data provides volumes of each co...

    work page 2021

  4. [4]

    Results 3.1. Effect of SSL Pretraining on Leaf-Wood Segmentation Self-supervised pretraining improved leaf -wood segmentation performance compared to Randomly-initialized model on the KR test set , with the largest gains appearing in wood prediction (Table 2). All three encoder configurations achieved overall accuracy above 90% for both needleleaf and bro...

    work page

  5. [5]

    Discussion 4.1. The Pretrained Encoder Transfers to Leaf-Wood Segmentation Across Forest Sites The encoder pretrained through the masked reconstruction pretext task transferred successfully to leaf-wood segmentation. Wood IoU increased according to the pretraining data configuration on both the KR test set and the cross -site benchmark (Tables 2 and 3), a...

    work page doi:10.5061/dryad.rfj6q5799 2017

  6. [6]

    ShapeNet: An Information-Rich 3D Model Repository

    https://doi.org/10.1111/2041-210X.12301 Chang, A. X., Funkhouser, T., Guibas, L., Hanrahan, P., Huang, Q., Li, Z., Savarese, S., Savva, M., Song, S., Su, H., Xiao, J., Yi, L., & Yu, F. (2015). ShapeNet: An Information-Rich 3D Model Repository (arXiv:1512.03012). arXiv. https://doi.org/10.48550/arXiv.1512.03012 Chen, S., Verbeeck, H., Terryn, L., Van den B...

    work page Pith review arXiv doi:10.1111/2041-210x.12301

  7. [7]

    State -of- the-Art in Photogrammetry, Remote Sensing and Spatial Information Science,

    https://doi.org/10.1080/17538947.2023.2198261 Dassot, M., Colin, A., Santenoise, P., Fournier, M., & Constant, T. (2012). Terrestrial laser scanning for measuring the solid wood volume, including branches, of adult standi ng trees in the forest environment. Computers and Electronics in Agriculture , 89, 86 –93. https://doi.org/10.1016/j.compag.2012.08.005...

    work page doi:10.1080/17538947.2023.2198261 2023

  8. [8]

    https://doi.org/10.1038/s41467-025-63946-6 Mei, J., Zhang, L., Wu, S., Wang, Z., & Zhang, L. (2017). 3D tree modeling from incomplete point clouds via optimization and L1 -MST. International Journal of Geographical Information Science, 31(5), 999–1021. https://doi.org/10.1080/13658816.2016.1264075 Momo Takoudjou, S., Ploton, P., Sonké, B., Hackenberg, J.,...

    work page doi:10.1038/s41467-025-63946-6 2017

Show all 9 references
  1. [9]

    https://doi.org/10.1109/TGRS.2022.3218603 Van den Broeck, W

    IEEE Transactions on Geoscience and Remote Sensing. https://doi.org/10.1109/TGRS.2022.3218603 Van den Broeck, W. A. J., Terryn, L., Chen, S., Cherlet, W., Cooper, Z. T., & Calders, K. (2025). Pointwise deep learning for leaf-wood segmentation of tropical tree point clouds from...

    2022 doi

Pith tools

Reviewed July 9, 2026 · model on record in the stance chip above.