pith. sign in

arxiv: 2602.00470 · v2 · pith:HU4HKWPTnew · submitted 2026-01-31 · 💻 cs.CV

FG-TreeSeg: Flow-Guided Tree Crown Segmentation without Instance Annotations

Pengyu Chen , Fangzheng Lyu , Sicheng Wang , Cuizhen Wang This is my paper

Pith reviewed 2026-05-21 14:01 UTC · model grok-4.3

classification 💻 cs.CV
keywords tree crown segmentationinstance segmentationtraining-freeremote sensingflow-guided segmentationstar-convex objectsforest monitoringCellpose-SAM
0
0 comments X

The pith

Modeling tree crowns as star-convex objects in a flow field enables training-free instance segmentation of dense overlapping canopies.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper seeks to establish a training-free approach to individual tree crown segmentation that avoids the annotation burden of supervised deep learning and the under-segmentation issues of general foundation models. It transfers a flow-based technique from biomedical imaging to remote sensing by treating crowns as star-convex shapes inside a topological flow field. A sympathetic reader would care because accurate crown boundaries support large-scale biomass estimation and ecological monitoring without requiring new labeled datasets for each sensor or forest type. If the method holds, it supplies both direct segmentations and automatic label generation across varying canopy densities.

Core claim

By modeling tree crowns as star-convex objects within a topological flow field using Cellpose-SAM, the FG-TreeSeg framework forces the separation of touching tree crown instances based on vector convergence and thereby delivers a training-free solution that generalizes across sensor types and canopy densities on the NEON and BAMFOREST datasets.

What carries the argument

Cellpose-SAM vector convergence inside a topological flow field that separates star-convex tree crowns

If this is right

  • Accurate instance segmentation becomes possible in dense canopies without collecting instance-level labels.
  • The same framework supplies training labels for subsequent supervised models.
  • Performance holds across optical and other sensor types without retraining.
  • Direct support is provided for forest biomass and ecological monitoring pipelines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The star-convex flow assumption could be tested on other clustered natural objects such as shrub patches or building rooftops.
  • Combining the flow output with spectral indices might further refine crown height or health attributes.
  • Failure modes on extremely sparse or extremely dense stands would indicate the practical density range for deployment.

Load-bearing premise

Tree crowns can be reliably modeled as star-convex objects inside a topological flow field so that vector convergence forces separation of touching instances.

What would settle it

A dataset of irregular non-star-convex crowns where the flow field produces merged or fragmented boundaries comparable to unguided foundation models would disprove the separation mechanism.

Figures

Figures reproduced from arXiv: 2602.00470 by Cuizhen Wang, Fangzheng Lyu, Pengyu Chen, Sicheng Wang.

Figure 1
Figure 1. Figure 1: Overview of the Zero-Shot Tree Crown Segmentation Framework. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Effect of semantic masking on zero-shot instance segmentation. [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Conceptual framework of the proposed flow-based segmentation. (1) [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance comparison across varying canopy densities. While the [PITH_FULL_IMAGE:figures/full_fig_p003_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results of the zero-shot instance segmentation. The model demon [PITH_FULL_IMAGE:figures/full_fig_p003_4.png] view at source ↗
Figure 7
Figure 7. Figure 7: Generalization test on the BAMFORESTS dataset. [PITH_FULL_IMAGE:figures/full_fig_p004_7.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison with Mask R-CNN (TCD) and Detectree2. [PITH_FULL_IMAGE:figures/full_fig_p004_6.png] view at source ↗
Figure 8
Figure 8. Figure 8: The impact of the diameter prior. stands), whereas a larger diameter promotes the aggregation of continuous canopies typical of mature forests. This sensi￾tivity effectively allows for the injection of biological priors into the inference process. It enables foresters to tailor the segmentation granularity to specific stand ages or species characteristics without the need for computational retraining, offe… view at source ↗
read the original abstract

Individual tree crown segmentation is an important task in remote sensing for forest biomass estimation and ecological monitoring. However, accurate delineation in dense, overlapping canopies remains a bottleneck. While supervised deep learning methods suffer from high annotation costs and limited generalization, emerging foundation models (e.g., Segment Anything Model) often lack domain knowledge, leading to under-segmentation in dense clusters. To bridge this gap, we propose FG-TreeSeg, a training-free framework for tree crown instance segmentation that transfers flow-based delineation from biomedical imaging to remote sensing. By modeling tree crowns as star-convex objects within a topological flow field using Cellpose-SAM, the FG-TreeSeg framework forces the separation of touching tree crown instances based on vector convergence. Experiments on the NEON and BAMFOREST datasets and visual inspection demonstrate that our framework generalizes robustly across diverse sensor types and canopy densities, which can offer a training-free solution for tree crown instance segmentation and labels generation.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes FG-TreeSeg, a training-free framework for individual tree crown instance segmentation in remote sensing imagery. It adapts the Cellpose-SAM flow-based delineation method from biomedical imaging by modeling tree crowns as star-convex objects inside a topological flow field, relying on vector convergence to separate touching instances. Experiments on the NEON and BAMFOREST datasets, supported by visual inspection, are presented as evidence of robust generalization across sensor types and canopy densities, positioning the approach as an annotation-free solution for segmentation and label generation.

Significance. If the central claims hold, the work could offer meaningful practical value by reducing the high annotation costs associated with supervised deep learning for tree crown delineation, a key task in forest biomass estimation and ecological monitoring. The transfer of flow-guided techniques and foundation models to remote sensing represents a potentially efficient alternative to domain-specific training, though its impact depends on verifiable performance gains over existing methods.

major comments (2)
  1. [Abstract] Abstract: The central claim of robust generalization across diverse sensor types and canopy densities rests on experiments on the NEON and BAMFOREST datasets, yet the manuscript supplies no quantitative metrics, baselines, error analysis, or implementation details to support this assertion, leaving the performance claims without verifiable support.
  2. [Framework description] Framework description: The approach explicitly models tree crowns as star-convex objects within the topological flow field so that Cellpose-SAM vector convergence separates instances; however, no quantitative checks or failure-case analysis are provided for irregular, overlapping, or non-radial canopies that violate star-convexity, which directly risks collapse of the separation mechanism under domain shift.
minor comments (2)
  1. [Abstract] The abstract would be strengthened by a brief mention of any specific quantitative indicators (e.g., IoU or F1 scores) even if preliminary, to allow readers to gauge the strength of the generalization claim.
  2. Notation for the flow field and vector convergence could be clarified with a short equation or diagram reference to improve reproducibility for readers unfamiliar with the Cellpose-SAM transfer.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review of our manuscript. We have carefully considered the major comments and provide point-by-point responses below. Where appropriate, we outline specific revisions that will be incorporated into the next version of the manuscript to address the concerns raised.

read point-by-point responses

  1. Referee: [Abstract] Abstract: The central claim of robust generalization across diverse sensor types and canopy densities rests on experiments on the NEON and BAMFOREST datasets, yet the manuscript supplies no quantitative metrics, baselines, error analysis, or implementation details to support this assertion, leaving the performance claims without verifiable support.

    Authors: We appreciate this observation. The current manuscript presents results primarily through visual inspection and qualitative comparisons on the NEON and BAMFOREST datasets to illustrate the training-free nature of FG-TreeSeg and its generalization across sensor types and canopy densities. While these results support the central claims, we agree that quantitative metrics, baselines, and error analysis would provide stronger verifiable evidence. In the revised manuscript, we will add quantitative segmentation metrics (such as IoU or F1 scores on available ground-truth subsets), comparisons against relevant baselines, and expanded implementation details in the methods and supplementary sections. revision: yes

  2. Referee: [Framework description] Framework description: The approach explicitly models tree crowns as star-convex objects within the topological flow field so that Cellpose-SAM vector convergence separates instances; however, no quantitative checks or failure-case analysis are provided for irregular, overlapping, or non-radial canopies that violate star-convexity, which directly risks collapse of the separation mechanism under domain shift.

    Authors: We thank the referee for this insightful comment. The star-convex modeling is inherited from the Cellpose-SAM framework as an effective approximation for separating instances via flow convergence. We acknowledge that this assumption may not hold perfectly for all irregular or non-radial canopies, which could impact performance in certain scenarios. In the revision, we will include a new analysis subsection that provides quantitative checks on the star-convexity assumption (e.g., measuring radial symmetry deviations across sampled crowns) and presents failure-case examples with visual and metric analysis from the datasets to better characterize limitations and domain-shift risks. revision: yes

Circularity Check

0 steps flagged

No circularity: direct transfer of Cellpose-SAM framework

full rationale

The paper presents FG-TreeSeg as a training-free adaptation that applies the existing Cellpose-SAM flow-based segmentation (originally from biomedical imaging) to remote-sensing tree crowns by explicitly modeling them as star-convex objects in a topological flow field. No equations, fitted parameters, or predictions are shown that reduce to the paper's own inputs by construction. The central mechanism relies on vector convergence for instance separation, which is an imported property of the prior method rather than a self-derived result. Experiments on NEON and BAMFOREST serve as external validation, not self-referential fits. No load-bearing self-citations, uniqueness theorems, or ansatz smuggling appear in the derivation chain. The framework is self-contained as an application of established tools without circular reduction.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim depends on the transferability of the biomedical flow model to tree crowns; the key untested premise is the star-convex flow-field representation of crowns.

axioms (1)
  • domain assumption Tree crowns can be modeled as star-convex objects within a topological flow field analogous to cells.
    Invoked when describing how Cellpose-SAM forces separation of touching instances.

pith-pipeline@v0.9.0 · 5701 in / 1231 out tokens · 53628 ms · 2026-05-21T14:01:58.232517+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

22 extracted references · 22 canonical work pages · 1 internal anchor

  1. [1]

    Individual tree segmentation based on mean shift and crown shape model for temperate forest,

    E. Tusa, J.-M. Monnet, J.-B. Barr ´e, M. D. Mura, M. Dalponte, and J. Chanussot, “Individual tree segmentation based on mean shift and crown shape model for temperate forest,”IEEE Geoscience and Remote Sensing Letters, vol. 18, no. 12, pp. 2052–2056, 2021

    work page 2052

  2. [2]

    A dual- branch deep learning framework at the grid scale for individual tree segmentation,

    Z. Ding, H. Zhang, R. Wang, L. Zhang, H. Jiang, and T. Yun, “A dual- branch deep learning framework at the grid scale for individual tree segmentation,”IEEE Geoscience and Remote Sensing Letters, vol. 22, pp. 1–5, 2025

    work page 2025

  3. [3]

    The combined use of uav- based rgb and dem images for the detection and delineation of orange tree crowns with mask r-cnn: An approach of labeling and unified framework,

    F. Lucena, F. M. Breunig, and H. Kux, “The combined use of uav- based rgb and dem images for the detection and delineation of orange tree crowns with mask r-cnn: An approach of labeling and unified framework,”Future Internet, vol. 14, no. 10, p. 275, 2022

    work page 2022

  4. [4]

    Individual tree-crown detection in rgb imagery using semi-supervised deep learning neural networks,

    B. G. Weinstein, S. Marconi, S. Bohlman, A. Zare, and E. White, “Individual tree-crown detection in rgb imagery using semi-supervised deep learning neural networks,”Remote Sensing, vol. 11, no. 11, p. 1309, 2019

    work page 2019

  5. [5]

    Segment Anything

    A. Kirillov, E. Mintun, N. Ravi, H. Mao, C. Rolland, L. Gustafson, T. Xiao, S. Whitehead, A. C. Berg, W.-Y . Lo, P. Doll´ar, and R. Girshick, “Segment anything,”arXiv:2304.02643, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  6. [6]

    Assessing sam for tree crown instance segmentation from drone imagery,

    M. Teng, A. Ouaknine, E. Lalibert ´e, Y . Bengio, D. Rolnick, and H. Larochelle, “Assessing sam for tree crown instance segmentation from drone imagery,” 2025. [Online]. Available: https://arxiv.org/abs/ 2503.20199

    work page arXiv 2025

  7. [7]

    Instance segmentation of individual tree crowns with YOLOv5: A comparison of approaches using the ForInstance benchmark LiDAR dataset,

    A. Straker, S. Puliti, J. Breidenbach, C. Kleinn, G. Pearse, R. Astrup, and P. Magdon, “Instance segmentation of individual tree crowns with YOLOv5: A comparison of approaches using the ForInstance benchmark LiDAR dataset,”ISPRS Open Journal of Photogrammetry and Remote Sensing, vol. 9, p. 100045, 2023. [Online]. Available: https://www.sciencedirect.com/s...

    work page 2023

  8. [8]

    Bringing sam to new heights: Leveraging elevation data for tree crown segmentation from drone imagery,

    M. Teng, A. Ouaknine, E. Lalibert ´e, Y . Bengio, D. Rolnick, and H. Larochelle, “Bringing sam to new heights: Leveraging elevation data for tree crown segmentation from drone imagery,” 2025. [Online]. Available: https://arxiv.org/abs/2506.04970

    work page arXiv 2025

  9. [9]

    Towards complete tree crown delineation by instance segmentation with mask r–cnn and detr using uav-based multispectral imagery and lidar data,

    S. Dersch, A. Schoettl, P. Krzystek, and M. Heurich, “Towards complete tree crown delineation by instance segmentation with mask r–cnn and detr using uav-based multispectral imagery and lidar data,”ISPRS Open Journal of Photogrammetry and Remote Sensing, vol. 8, p. 100037, 2023

    work page 2023

  10. [10]

    Parallel fusion neural network considering local and global semantic information for citrus tree canopy segmentation,

    H. He, F. Zhou, Y . Xia, M. Chen, and T. Chen, “Parallel fusion neural network considering local and global semantic information for citrus tree canopy segmentation,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 17, pp. 1535–1549, 2024

    work page 2024

  11. [11]

    Multimodal uncertainty robust tree cover segmentation for high-resolution remote sensing images,

    Y . Gui, W. Li, Y . Wang, X.-G. Xia, M. Marty, C. Ginzler, and Z. Wang, “Multimodal uncertainty robust tree cover segmentation for high-resolution remote sensing images,”IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, vol. 19, pp. 114– 128, 2026

    work page 2026

  12. [12]

    Cellvit: Vision transformers for precise cell segmentation and classification,

    F. H ¨orst, M. Rempe, L. Heine, C. Seibold, J. Keyl, G. Baldini, S. Ugurel, J. Siveke, B. Gr ¨unwald, J. Eggeret al., “Cellvit: Vision transformers for precise cell segmentation and classification,”Medical Image Analysis, vol. 94, p. 103143, 2024

    work page 2024

  13. [13]

    Cellpose-SAM: superhuman generalization for cellular segmentation,

    M. Pachitariu, M. Rariden, and C. Stringer, “Cellpose-SAM: superhuman generalization for cellular segmentation,”bioRxiv, 2025. [Online]. Available: https://www.biorxiv.org/content/early/2025/05/01/ 2025.04.28.651001

    work page 2025

  14. [14]

    Oam-tcd: a globally diverse dataset of high-resolution tree cover maps,

    J. Veitch-Michaelis, A. Cottam, D. Schweizer, E. N. Broadbent, D. Dao, C. Zhang, A. A. Zambrano, and S. Max, “Oam-tcd: a globally diverse dataset of high-resolution tree cover maps,” inProceedings of the 38th International Conference on Neural Information Processing Systems, ser. NIPS ’24. Red Hook, NY , USA: Curran Associates Inc., 2024

    work page 2024

  15. [15]

    Nuclei instance segmentation and classifi- cation in histopathology images with stardist,

    M. Weigert and U. Schmidt, “Nuclei instance segmentation and classifi- cation in histopathology images with stardist,” inThe IEEE International Symposium on Biomedical Imaging Challenges (ISBIC), 2022

    work page 2022

  16. [16]

    Individual tree crown delineation in high resolution aerial RGB imagery using StarDist-based model,

    F. Tong and Y . Zhang, “Individual tree crown delineation in high resolution aerial RGB imagery using StarDist-based model,”Remote Sensing of Environment, vol. 319, p. 114618, 2025. [Online]. Available: https://www.sciencedirect.com/science/article/pii/S0034425725000227

    work page 2025

  17. [17]

    Data for the neontreeevaluation benchmark,

    B. Weinstein, S. Marconi, and E. White, “Data for the neontreeevaluation benchmark,” Jan. 2022. [Online]. Available: https://doi.org/10.5281/ zenodo.5914554

    work page 2022

  18. [18]

    Bamforests: Bamberg benchmark forest dataset of individual tree crowns in very-high- resolution uav images,

    J. Troles, U. Schmid, W. Fan, and J. Tian, “Bamforests: Bamberg benchmark forest dataset of individual tree crowns in very-high- resolution uav images,”Remote Sensing, vol. 16, no. 11, 2024. [Online]. Available: https://www.mdpi.com/2072-4292/16/11/1935

    work page 2024

  19. [19]

    TreePseCo: Scaling individual tree crown segmentation using large vision models,

    J. Lungo Vaschetti, E. Arnaudo, and C. Rossi, “TreePseCo: Scaling individual tree crown segmentation using large vision models,”The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, vol. XLVIII-M-7-2025, pp. 275–282,

    work page 2025

  20. [20]

    Available: https://isprs-archives.copernicus.org/articles/ XLVIII-M-7-2025/275/2025/

    [Online]. Available: https://isprs-archives.copernicus.org/articles/ XLVIII-M-7-2025/275/2025/

    work page 2025

  21. [21]

    Comparing mask r-CNN and mask2former architectures for individual tree crown delineation,

    S. Ruschhaupt, J. Troles, and U. Schmid, “Comparing mask r-CNN and mask2former architectures for individual tree crown delineation,” in45. GIL-Jahrestagung, Digitale Infrastrukturen f ¨ur eine nachhaltige Land-, Forst-und Ern¨ahrungswirtschaft, 2025

    work page 2025

  22. [22]

    Accurate delineation of individual tree crowns in tropical forests from aerial rgb imagery using mask r-cnn,

    J. G. Ball, S. H. Hickman, T. D. Jackson, X. J. Koay, J. Hirst, W. Jay, M. Archer, M. Aubry-Kientz, G. Vincent, and D. A. Coomes, “Accurate delineation of individual tree crowns in tropical forests from aerial rgb imagery using mask r-cnn,”Remote Sensing in Ecology and Conservation, vol. 9, no. 5, pp. 641–655, 2023

    work page 2023