{"id":"74002bca-3a80-498b-bc96-31268f4579dc","arxiv_id":"1908.04018","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A 3D point cloud method that filters away overlapping leaf edges, then grows labels from leaf centers, achieves 99.66% average F-measure on four plant species.","lead":"The paper presents a pipeline that separates overlapping leaves in 3D plant point clouds by first filtering away the crowded edges, then growing leaf labels back from the centers. It reports near-perfect leaf-level accuracy on four species scanned with three different 3D imaging setups.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported near-perfect accuracy is in-sample: all four test clouds are the same clouds used to set Table 7 parameters, so the generalization claim is not yet supported.","rationale":"The reader's weakest assumption is exactly the one I find most load-bearing: one sample plant per species with parameters tuned on those same clouds cannot establish generalization. The manuscript's own Section 6.4 and Section 8 concede that parameters are adjusted per cloud, so the concern is grounded in the text rather than an external standard. I considered whether a more technical flaw exists in the SBF boundary criterion or the TP>70% leaf-level metric, but those are permissive yet not fatal: the point-level cover rates and qualitative figures give some support that the method works on the tested clouds. The improvement over prior methods is credible on those instances. Therefore the verdict should remain CONDITIONAL, with the condition being a held-out evaluation or release of code/data that permits one. I do not see grounds to reject, because the method is internally consistent and the reported results are not contradicted by the evidence; I also do not see grounds to accept unconditionally, because the central generalization claim rests entirely on in-sample, single-instance data.","tokens_in":23373,"tokens_out":3187,"duration_ms":37399,"concrete_test":"Collect one additional plant per species with the same acquisition platforms described in Section 2.1.1. Set all parameters using only the Section 6.4 rules and each new cloud's average spacing (r about 4-5 times spacing, n_threshold below the Table 6-style spherical-neighborhood count, k=20, d1 2-4 times spacing, preprocessing set fixed per Figure 3), without any cloud-specific adjustment. Compare leaf-level Recall, Precision, F-measure and point-level cover rate against manual ground truth. If any species drops below, say, 95% F-measure or its cover rate falls by more than 5 points relative to Table 5, the generalization claim is unsupported; if all four held-out plants stay near the reported values, the in-sample tuning concern is resolved. This test requires the authors' implementation or a faithful reimplementation of Tables 1-3, since no code or data are currently provided.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the method generalizes across imaging platforms and plant species. What would have to be true is that the pipeline's tunable choices — the preprocessing filter set (Figure 3), r, n_threshold, k, d1, and the number of joint-filtering rounds — transfer to new plants without per-cloud re-tuning. The paper does not show this. Section 6.4 states that parameters are adjusted separately for each cloud, with r, n_threshold, and d1 derived from the average spacing and spherical-neighborhood counts of that same cloud (Tables 6 and 7), and Section 2.1.2 tests exactly one sample plant per species. Thus every headline number — 97%, 99%, 99%, 87% point-level cover rates and 99.66% average F-measure — comes from the same clouds that were used to select parameters. Section 8 explicitly concedes that several parameters need to be tuned for an optimal result. Because plant species, imaging platform, and tuning instance are perfectly confounded, the results cannot support the broad applicability claim. This is not an internal inconsistency: the individual segmentation outputs are plausible, and the comparison against [4] and [19] on these clouds is meaningful. The load-bearing gap is external validity: no held-out plant, no repeated acquisition, no error bars, and no code or data release. If the parameter-selection rules in Section 6.4 do not transfer, the near-perfect F-measure would not persist on new data.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes a three-stage pipeline for individual leaf segmentation in 3D plant point clouds: (i) species-specific concatenations of five preprocessing filters; (ii) a new 3D joint filtering operator that combines a radius-based outlier filter (RBOF) and a surface boundary filter (SBF) to erode overlapping regions and isolate leaf centers; and (iii) pre-segmentation of leaf centers by breadth-first 3D region growing, followed by facet over-segmentation and facet region growing to attach filtered boundary points back to leaves. The method is evaluated on one sample plant each of Epipremnum aureum, Monstera deliciosa, Calathea makoyana, and Hedera nepalensis, acquired with Kinect V2, binocular stereo, and multi-view systems. The authors report point-level cover rates of 97%, 99%, 99%, and 87% for these plants and leaf-level averages of Recall 100.00%, Precision 99.33%, and F-measure 99.66%, plus a leaf-area estimation application with average errors below 5% on nine Calathea leaves.","tokens_in":23666,"tokens_out":11262,"duration_ms":105900,"significance":"The proposed combination of erosion-like joint filtering and facet-based label growing is sensible, and the qualitative results are visually convincing. If the reported accuracies held on unseen plants, the method would be a practically valuable contribution to plant phenotyping, because it addresses the difficult problem of overlapping leaves and is demonstrated on multiple sensor types. The authors provide detailed point-level tables and compare against two relevant baselines ([4] and [19]); the per-plant F-measures in Table 5 are consistent with the stated average. The main value of the paper is the joint filtering plus add-back strategy, which is a reasonable and potentially transferable idea. However, the evidence currently supports only in-sample performance on four single-plant clouds, and the leaf-level metric definitions are too lenient to justify the near-perfect headline. No code or data are released, which further limits reproducibility.","major_comments":[{"comment":"The central generalization claim is not supported by the evaluation design. Only one sample plant per species is used, and Table 7 parameters (r, n_threshold, d1, number of joint filtering rounds) are set per point cloud using the average spacing and spherical-neighborhood statistics of that same cloud (Tables 6-7), with preprocessing filter concatenations also chosen per species (Figure 3). The reported point-level cover rates and the 99.66% average F-measure are therefore in-sample estimates. In addition, species and imaging platform are perfectly confounded (Epipremnum aureum with Kinect V2, Monstera deliciosa with binocular stereo, Calathea and Hedera with multi-view), so the claim that the method generalizes across platforms and species is not testable from these data. Section 8 concedes that several parameters require per-case tuning. The authors should add held-out plants, ideally multiple plants per species and repeated acquisitions, report error bars, and demonstrate that the Section 6.4 tuning rules transfer without per-cloud re-tuning; otherwise the broad applicability claim should be substantially weakened.","section":"Section 2.1.2, Section 6.4, Tables 6-7, Section 8"},{"comment":"The leaf-level metrics are inconsistent with the detailed point-level data. Table 4 lists a Hedera nepalensis leaf 37 with Num_gt=243, Rp=0, Mp=243, and cover_rate=0%, and leaf 11 with cover_rate=51.8%. Under the FN definition in Section 6.2, which states that a real leaf with more than 70% of its points not covered is an FN, leaf 37 should be counted as an FN, yet Table 5 reports FN=0 and Recall=100% for Hedera. The accompanying note to Figure 14 says leaves 24 and 37 were falsely merged into one segment, which should produce both an FP and an FN under the stated definitions. Moreover, the 70% coverage threshold means that a leaf with only 51.8% coverage is neither TP nor FN, so Recall can remain 100% despite substantial under-segmentation. The authors should align the definitions with the per-leaf data, report per-leaf TP/FP/FN, and supplement with stricter point-level or IoU-based metrics before claiming near-perfect leaf-level accuracy.","section":"Section 6.2, Table 4, Table 5"}],"minor_comments":[{"comment":"There are typographical errors such as 'Filer II' in Figure 3 and 'crowed' in Sections 1 and 8; these should be corrected.","section":"Section 3, Figure 3"},{"comment":"Equation (1) and several inline formulas are garbled by formatting; please use clean mathematical notation so the SBF angle computation is unambiguous.","section":"Equation (1), Section 4.2"},{"comment":"Table 4 is very hard to read because of the two-column layout and ambiguous row boundaries; please restructure it so each plant's leaves are clearly separated and all column headings are repeated for each block.","section":"Table 4"},{"comment":"The leaf area, length, and width validation is performed on only nine leaves from one Calathea makoyana plant; this limitation should be stated explicitly when the trait-estimation accuracy is reported.","section":"Section 7.2"},{"comment":"The conclusion acknowledges that parameters need tuning and preprocessing is species-specific, but the abstract overstates generality; the abstract should be tempered to reflect these limitations.","section":"Section 8, Abstract"},{"comment":"No data or code availability statement is provided; adding one would materially help reproducibility, especially given the parameter choices in Tables 6 and 7.","section":"Data availability"}],"recommendation":"major_revision","confidential_remarks":"The paper is borderline between major revision and rejection. The core idea is reasonable and the comparisons with [4] and [19] are meaningful, but the evaluation as designed cannot support the claimed generalization, and the leaf-level metric inconsistency affects the headline recall. I believe these deficiencies are addressable through additional held-out experiments and corrected metrics, so I recommend major revision rather than rejection. The journal may also wish to consider whether single-plant-per-species validation meets its standards for claims of cross-species and cross-platform applicability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know up front. First, this is a competent engineering paper: it combines known point-cloud filters (radius outlier removal, PCA-based boundary detection, facet over-segmentation) into a pipeline that segments overlapping leaves, and it reports very high accuracy on four sample plants. Second, the headline numbers are all in-sample: the same single plant per species that was used to set the pipeline's parameters is the plant on which accuracy is reported, so the generalization claim is not yet supported.\n\nWhat's genuinely new is the 3D joint filtering operator (RBOF + SBF) and the way it erodes boundary/overlap points to expose leaf centers, then grows facets back. The authors are explicit that this builds on their earlier work [4] and on [33], so the novelty is mostly in the combination, not the components. That is a legitimate extension. The experiments are carefully presented: per-leaf point-level cover rates, leaf-level TP/FP/FN, comparisons against [4] and [19], and the numbers are internally consistent. The trait estimation demo on Calathea (area, length, width) is a nice bonus.\n\nThe critical weakness is external validity. Section 6.4 shows that r, n_threshold, and d1 are derived from the average spacing and neighborhood counts of the same point cloud being segmented, and Section 2.1.2 uses one plant per species. So the 97%/99%/99%/87% cover rates and 99.66% F-measure are not held-out estimates. Section 8 concedes several parameters need tuning. There are no error bars, no repeated acquisitions, no independent test plants, and no code or data release. That doesn't make the results false—they are plausible, and the comparisons against [4] and [19] are meaningful on these clouds—but the paper's claim of wide applicability across species and platforms is not established. If the parameter-selection rules transfer to new plants, the method is likely useful; we just don't know yet.\n\nThis is a paper for someone working on 3D plant phenotyping or point cloud foliage segmentation. It deserves a serious referee, but the referee should send it back for a held-out evaluation, not desk-reject it. I'd recommend peer review with the clear expectation that the authors add new plants or repeated scans, demonstrate parameter transfer, and ideally release code and data.","headline":"A solid engineering pipeline for overlapping-leaf segmentation in 3D point clouds, but the near-perfect accuracy is measured on the same clouds used to tune parameters, so the generality claim is not yet supported.","tokens_in":24274,"tokens_out":1922,"would_cite":false,"duration_ms":19810,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a two-stage 3D filtering-and-growing pipeline can segment densely overlapping leaves in plant point clouds at near-perfect accuracy, by first eroding shared boundaries and then expanding labeled leaf centers back…","keywords":["leaf segmentation","point cloud","3D joint filtering","facet region growing","plant phenotyping","overlapping leaves","leaf area estimation"],"falsifier":"Scan a new plant of any of the four tested species, apply the parameter values from Tables 6 and 7 without any retuning, and compute the leaf-level F-measure; if the result falls well below the reported 99.66% average, the generalization claim fails.","tokens_in":1511,"feed_emoji":"🍃","tokens_out":4364,"duration_ms":65417,"temperature":0.7,"pith_summary":"This paper tries to establish that individual leaves can be segmented from crowded, overlapping plant canopies in 3D point clouds by first eroding away the shared boundaries between leaves and then growing labeled leaf centers back into the filtered regions. The proposed pipeline begins with a new 3D joint filtering operator that combines a radius-based outlier filter for cross-overlapping leaves and a surface boundary filter, based on PCA normals, for coplanar overlapping leaves. After filtering, the remaining leaf-center points are labeled by region growing, then the filtered boundary points are over-segmented into facets and assigned to those labels by facet region growing. If the claim holds, dense canopy point clouds from different sensors and species can be turned into per-leaf point sets, enabling automatic estimation of leaf area, length, and width.","feed_headline":"3D filter separates overlapping leaves at 99.66% F-measure","feed_subtitle":"A two-part filter erodes shared leaf edges, then facet growing labels each leaf in dense canopy scans.","key_machinery":"The load-bearing mechanism is the 3D joint filtering operator, composed of a Radius-based Outlier Filter (RBOF) and a Surface Boundary Filter (SBF). RBOF removes sparse points in cross-overlap areas by counting neighbors within a radius, while SBF detects boundary points by computing PCA normals in a k-neighborhood and measuring the maximum angular gap between projected neighbor directions, acting like morphological erosion on a 3D surface. After one or more rounds of joint filtering, the remaining leaf-center points are separated by breadth-first 3D region growing, and the filtered boundary points are over-segmented into facets, which are then grown into full leaves by assigning them to adjacent labeled centers from inside outward.","core_discovery":"The central discovery is that an erosion-like 3D filtering stage can break the connectivity between heavily overlapped leaves in plant point clouds, after which a facet-based region growing stage can restore complete leaves by expanding labels from leaf centers outward. The paper reports point-level cover rates of 97% for Epipremnum aureum, 99% for Monstera deliciosa, 99% for Calathea makoyana, and 87% for Hedera nepalensis, with leaf-level averages of 100.00% Recall, 99.33% Precision, and 99.66% F-measure across those four sample plants. It also claims the method works on point clouds from binocular stereo, a consumer depth camera, and multi-view structure-from-motion, and that the segmented leaves allow automatic trait estimation with average errors of 0.47% for area, 2.89% for length, and 4.64% for width on Calathea makoyana.","pith_inferences":["The erosion-then-regrow scheme is not inherently leaf-specific; the same RBOF-plus-SBF boundary removal followed by facet growing could plausibly separate other touching or overlapping objects in point clouds, such as fruit clusters, rock fragments, or stacked industrial parts, whenever the objects have locally distinguishable surface orientation.","If the parameters of the joint filter were set automatically from local point spacing and density statistics, rather than tuned per point cloud, the method could become a turnkey preprocessing step for high-throughput phenotyping pipelines.","The number of joint-filtering rounds functions like a scale-space erosion parameter; a natural extension would be to stop filtering when the connected-component count stabilizes, making the depth of erosion adaptive to canopy density.","The reported accuracy values come from the same clouds used for parameter tuning, so a direct test on new plants with fixed parameters would tell whether the method generalizes or needs per-species recalibration."],"forward_implications":["If the central claim is correct, per-leaf phenotyping of crowded canopies becomes feasible without manual separation, enabling automatic leaf area, length, and width measurement in plant research and agricultural applications.","The method is claimed to generalize across three types of 3D imaging platforms, so a single segmentation pipeline could be applied to point clouds from binocular stereo, depth sensors, and multi-view reconstruction without changing the core algorithm.","The reported near-perfect leaf-level scores suggest that the combination of boundary erosion and facet-based regrowth resolves both cross-overlapping and coplanar-overlapping leaf configurations, the two failure modes cited for existing methods.","Because the pipeline also performs trait estimation from the segmented point clouds, it offers an end-to-end path from raw scans to phenotypic trait values."],"supporting_citations":[{"why":"Supplies the facet over-segmentation and facet region growing machinery that this method extends, including parameters for the over-segmentation stage.","marker":"[4]"},{"why":"Provides the facet segmentation approach used for clustering filtered boundary points into facets.","marker":"[13]"},{"why":"Serves as the state-of-the-art comparison baseline for automated leaf segmentation from terrestrial laser scanning point clouds.","marker":"[19]"},{"why":"Inspires the PCA-based boundary point detection that underlies the Surface Boundary Filter, giving the method its edge-erosion capability.","marker":"[33]"},{"why":"Supplies the greedy projection triangulation technique used to turn a segmented leaf point cloud into a mesh for leaf area estimation.","marker":"[35]"}],"fun_headline_variants":["3D filter untangles overlapping leaves at 99.66% F-measure","99.66% F-measure for separating overlapping plant leaves","New 3D method separates overlapping leaves in point clouds","Overlapping-free leaf segmentation hits 99.66% F-measure"],"cache_read_input_tokens":26240,"weakest_assumption_plain":"The evaluation assumes that one sample plant per species, with parameters tuned separately for each point cloud using statistics from that same cloud, is enough to prove the method generalizes to other plants of the same species and to new species.","fun_headline_variants_meta":{"raw":{"variants":["3D filter untangles overlapping leaves at 99.66% F-measure","99.66% F-measure for separating overlapping plant leaves","New 3D method separates overlapping leaves in point clouds","Overlapping-free leaf segmentation hits 99.66% F-measure"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000793,"raw_usage":{"total_tokens":3562,"prompt_tokens":1083,"completion_tokens":2479,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":699,"completion_tokens_details":{"reasoning_tokens":2403}},"tokens_in":699,"tokens_out":2479,"duration_ms":16676,"temperature":1.0,"reasoning_tokens":2403,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:54:19.931992+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Scan a new plant of any of the four tested species, apply the parameter values from Tables 6 and 7 without any retuning, and compute the leaf-level F-measure; if the result falls well below the reported 99.66% average, the generalization claim fails.","supporting_citations":[{"cited_title":"- S.; Yan, S.; Cai, X","cited_arxiv_id":null,"evidence_quote":"Supplies the facet over-segmentation and facet region growing machinery that this method extends, including parameters for the over-segmentation stage."},{"cited_title":"Facet Segmentation - Based Line Segment Extraction for Large - Scale Point Clouds","cited_arxiv_id":null,"evidence_quote":"Provides the facet segmentation approach used for clustering filtered boundary points into facets."},{"cited_title":"Automated Segmentation of Leaves from Deciduous Trees in Terrestrial Laser Scanning Point Clouds","cited_arxiv_id":null,"evidence_quote":"Serves as the state-of-the-art comparison baseline for automated leaf segmentation from terrestrial laser scanning point clouds."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inspires the PCA-based boundary point detection that underlies the Surface Boundary Filter, giving the method its edge-erosion capability."},{"cited_title":"C.; Rusu, R","cited_arxiv_id":null,"evidence_quote":"Supplies the greedy projection triangulation technique used to turn a segmented leaf point cloud into a mesh for leaf area estimation."}],"review_version":1}