acceptodds
Withdrawn from ICLR 2027

GIVF: Geometry-Guided Whole-Tree Reconstruction from Crop-Wise Predictions

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Whole-tree instance segmentation requires both assigning points to coherent objects and determining which reconstructed instances are valid. This becomes particularly challenging when large forest scenes are processed through overlapping crops, since reconciling local instance masks does not necessarily correct point membership errors within those predictions. We therefore formulate scene-level whole-tree decoding as two related problems: reconstructing point membership and assessing the validity of the resulting instances. To address them, we propose GIVF, a modular framework that integrates geometry-guided instance assembly and learned validity assessment into a single decoding pipeline. GIVF aggregates crop-wise offsets into spatial anchors and combines them with fine-scale geometric tree-membership and trunk-support evidence. Seeded geodesic propagation constructs disjoint whole-tree candidates on a scene-level graph using spatial connectivity and anchor consistency. A lightweight validity model subsequently determines whether each assembled candidate constitutes a valid tree instance, using its geometric structure, vertical profile, trunk support, and native prediction confidence. On four test plots of the Cherlet TLS benchmark, GIVF achieves a macro detection F1 of 0.7254 and a matched-tree mF1 of 0.8865.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
instance segmentation, 3D point cloud, individual tree segmentation, forest LiDAR

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_isGjOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the limitations of crop-wise predictions in reconstructing complete tree instances from forest LiDAR point clouds. The authors argue that merging local instance masks is insufficient to correct point-membership errors, and propose GIVF, which combines geometric evidence and learned spatial anchors for graph-based whole-tree reconstruction, followed by instance validity assessment. Experiments on 4 forest benchmarks demonstrate improved tree detection over existing methods, with additional evidence supporting supervised geometric adaptation across different LiDAR scenes.

Critical strengths
  1. GIVF achieves a macro detection F1 of 0.7254 on the four Cherlet benchmark plots, outperforming the strongest reported baseline (0.5825) by 14.3 percentage points. The consistent improvements across all four plots provide meaningful empirical evidence of the framework's effectiveness in challenging forest environments.

  2. The paper identifies an important limitation of crop-wise instance segmentation: merging local predictions does not necessarily correct erroneous point memberships. Its separation of geometry-guided membership reconstruction and instance validity assessment directly addresses this limitation and provides a coherent framework for whole-tree segmentation.

Critical weaknesses
  1. Appendix A.5 explicitly acknowledges that the 0.98 positive-candidate retention requirement originated from exploratory development during which test results had already been inspected. Although the final thresholds were selected using validation data, this raises concerns about test-set independence and weakens confidence in the reliability of the reported performance gains.

  2. The framework largely combines established techniques, while the ablation studies do not isolate the contributions of its key components. More importantly, despite substantial improvements in detection F1 (0.3170 to 0.7254), matched-tree mF1 increases only marginally (0.8778 to 0.8865), providing limited direct evidence that the proposed approach improves point-membership reconstruction, a central motivation of the paper.

Questions for authors
  1. Appendix A.5 states that test results had been inspected before adopting the 0.98 positive-candidate retention requirement. Could the authors clarify which experimental decisions were influenced by this inspection and provide evidence that the reported improvements remain robust under a strictly validation-only selection protocol?

  2. Table 4 shows substantial detection improvements but only marginal changes in matched-tree mF1, with different matched subsets across stages. Could the authors provide a controlled evaluation on the same ground-truth tree instances to demonstrate whether the proposed reconstruction genuinely improves point membership rather than primarily improving instance detection and filtering?

  3. Could the authors provide component-level ablations isolating geometric tree/trunk evidence, neural anchor consistency, and graph-based propagation? Such comparisons are important for determining whether the reported gains arise from the proposed mechanisms or could be achieved by simpler existing grouping strategies.

Reviewer_q89zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GIVF, a geometry-guided framework for whole-tree instance segmentation from crop-wise LiDAR predictions. The method addresses inconsistencies in point membership across overlapping crops by integrating neural offset predictions, geometric tree/trunk evidence, and anchor-aware geodesic propagation to reconstruct scene-level tree instances. A subsequent gradient-boosted validity classifier filters unreliable candidates. Experiments on four Cherlet TLS benchmark plots demonstrate substantial improvements in detection performance, achieving a macro F1 of 0.7254 and a matched-tree mF1 of 0.8865. Although the results are promising, concerns remain regarding methodological novelty, experimental completeness, and evaluation rigor.

Critical strengths
  1. Practical problem formulation. The paper addresses an important challenge in large-scale forest point-cloud segmentation, where conventional crop-wise mask merging cannot adequately resolve incorrect point memberships and fragmented tree instances.
  2. Reasonable integration of geometric and learned cues. The combination of spatial anchors, tree/trunk evidence, and graph-based propagation provides a structured approach to reconstructing complete tree instances in complex forest environments.
  3. Substantial experimental improvements. GIVF achieves consistent detection improvements across all four benchmark plots, outperforming the reported TreeLearn and ForAINet configurations. The stage-wise evaluation also demonstrates the benefits of geometric assembly and subsequent validity filtering.
  4. Cross-domain applicability and implementation details. The supervised adaptation experiments demonstrate that lightweight geometric classifiers can improve performance while keeping the neural predictor frozen. The appendix provides relatively detailed implementation settings that support reproducibility.
Critical weaknesses
  1. The framework mainly combines existing techniques, including geometric descriptors, offset-based grouping, shortest-path propagation, and gradient-boosted classification. Its differences from closely related methods such as TreeLearn and SATree are insufficiently demonstrated (Lines 123–137, 202–242).
  2. Table 1 only compares against TreeLearn and ForAINet, omitting relevant methods such as ForestFormer3D (ICCV 2025), SATree (2026), and SegmentAnyTree (2024). These comparisons are important for establishing the method's competitiveness (Lines 270–280).
  3. The ablation only evaluates three overall stages without isolating key components, such as geometric tree scores, trunk-guided seeds, anchor consistency, and graph edge costs. Hyperparameter sensitivity analysis is also missing (Lines 418–446).
  4. The authors acknowledge that the retention requirement of 0.98 was inherited from exploratory development involving inspection of test results. This raises concerns about evaluation independence and possible selection bias (Lines 793–801).
  5. While detection F1 improves significantly, matched-tree mF1 increases only from 0.8778 to 0.8865. Additional point-level evaluations on fixed ground-truth instances are needed to support the claimed membership reconstruction improvements (Lines 432–446).
Questions for authors
  1. GIVF uses additional plot-specific geometric and validity classifiers, while baseline results are taken from prior work. Comparisons under identical backbones, supervision, and training conditions would be more convincing (Lines 250–256, 305–310).
  2. Cross-domain experiments still require target-domain supervision, while inference time, memory consumption, and scalability are not evaluated. Performance also remains limited on challenging scenes such as Robson Creek (Lines 317–323, 364–377).
  3. The manuscript contains several minor writing and formatting inconsistencies. For example, “Oneformer3D” should be “OneFormer3D” (Line 160). Additionally, Table 1 discusses macro F1 in its caption but does not explicitly report the corresponding values. A careful proofreading is recommended.
Reviewer_dKMKOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a geometry-informed framework GIVF for individual tree instance segmentation in TLS data. Addressing the limitations of existing methods—which struggle to effectively integrate predictions from overlapping local crops, often leading to fragmented instances and misassigned points—the authors propose reconstructing individual tree instances within the full scene, rather than simply merging instance masks derived from local predictions.

The key contributions are: (1) reconstructing complete tree instances via graph-based geodesic propagation; (2) incorporating an independent instance-validity classifier that filters candidate instances based on geometric structure, vertical distribution, trunk support, and prediction confidence; and (3) developing a modular geometric evidence fusion mechanism, which enables supervised cross-domain adaptation by keeping only a lightweight geometric classifier, while keeping the neural network unchanged.

Critical strengths
  1. Improvement in individual tree detection performance

On the TLS forest test plots from Cherlet et al. (2026), the authors achieved detection performance better than the reported baseline configurations of TreeLearn and ForAINet, reaching a macro-averaged F1 score of 0.7254—an improvement of approximately 14.3 percentage points over the best-performing baseline (0.5825). Furthermore, a staged ablation study demonstrated that geometric reorganization boosted the detection F1 score from 0.3170 to 0.6413, while the application of instance validity filtering further increased it to 0.7254.

Critical weaknesses
  1. Insufficient baseline comparisons

The paper’s quantitative evaluation use TreeLearn and ForAINet, but there are many comparisons with other representative individual-tree instance segmentation methods—like TLS2trees, Treeiso, SATree etc. SATree, in particular, use the same mechanism as GIVF: it combines trunk localization, neural-network-predicted offsets, and a graph-based shortest-path algorithm to achieve individual-tree segmentation. Furthermore, the benchmark dataset introduced by Cherlet et al. (2026) already includes several traditional geometric segmentation algorithms; however, the paper fails to adequately incorporate these as baselines. Consequently, the current experimental results are insufficient to demonstrate the overall superiority of GIVF over existing geometry-guided segmentation methods, thereby weakening the soundness of the claimed performance gains and methodological innovation.

  1. Lack of clarity regarding the innovation of the core methodology.

In the Introduction, the authors argue that traditional "cross-crop instance-mask merging" fails to correct erroneous point assignments, positioning "scene-level membership reconstruction" as a key contribution. However, TreeLearn (Henrich et al., 2024) already performs clustering and point assignment across the entire scene by aggregating offset predictions from overlapping crops, rather than merely merging local instance masks. Thus, the problem the authors highlight is not a novel, previously unaddressed challenge. Moreover, SATree (Du et al., 2026) also integrates learned offsets, trunk localization, and graph-based shortest-path grouping. While GIVF employs a distinct design for graph edge weights and incorporates an independent validity classifier, the paper fails to sufficiently articulate how these design choices constitute a substantive innovation over existing approaches.

Furthermore, the current ablation studies modify multiple components—such as geometric classification, seed construction, and graph segmentation—simultaneously. This makes it impossible to isolate and verify the necessity or specific advantages of the proposed "anchor-aware geodesic assignment."

Questions for authors
  1. The differences compared to existing methods?

TreeLearn (Henrich et al., 2024) performs clustering and point assignment within the full scene by aggregating offset predictions from overlapping crops, rather than simply merging local instance masks; SATree (Du et al., 2026) incorporates trunk localization, learned offsets, and graph-based shortest-path segmentation.

Could the authors further clarify the core algorithmic innovations of GIVF relative to these existing methods—specifically, the fundamental distinction between "anchor-aware geodesic assignment" and current offset-guided grouping approaches?

  1. Lack comparisons with other geometric segmentation methods The paper primarily provides quantitative comparisons against TreeLearn and ForAINet; however, the original benchmark by Cherlet et al. (2026) also includes traditional geometric segmentation methods such as Rayextract, Treeiso, and SSSC. Could the authors provide comparisons against these methods using the same evaluation protocol, or explain why they were excluded as baselines?

Additionally, the current ablation—transitioning from "Native" to "Geometric assembly"—simultaneously alters geometric classification, seed construction, and graph segmentation, making it impossible to isolate the individual contribution of each module. Could the authors provide further controlled ablation studies—for instance, comparing graph propagation without the "anchor consistency penalty" against the full GIVF model, while keeping inputs, geometric predictions, and seed conditions constant—to demonstrate the specific contribution of the proposed graph-decoding mechanism?

  1. If the test set influence hyper-parameter selection during the method development process?

IN Appendix A.5, the paper mentions that the 0.98 positive-retention constraint used in "validity filtering" was derived from an exploratory development phase during which test results had already been examined. Could the authors clarify the extent to which test results influenced this constraint—as well as other design choices? Can the test set used for the final reported results still be considered an independent evaluation? This is crucial for assessing the reliability of the reported performance gains.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.