{"id":"5462c925-979c-46d1-80f5-1503a67d95b1","arxiv_id":"2411.13069","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A skeleton-based tetrahedron matching algorithm achieves faster and more accurate alignment of single-tree LiDAR scans than ICP and NDT on eight test trees.","lead":"This paper proposes a marker-free method to align multiple laser scans of individual trees by matching similar tetrahedra built from the trees' branch skeletons. The method is fast and reportedly more accurate than standard ICP and NDT registration, which could make forestry surveys cheaper and faster.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RMSE/HD evaluation in Section III.B lacks a defined correspondence or ground truth, making the central accuracy claim unverifiable.","rationale":"The reader's weakest assumption about skeleton-topology consistency is a plausible internal limitation, and the paper itself acknowledges it in Section IV. However, the paper's central claim is about outperforming ICP and NDT in accuracy, and that claim is only as strong as the evaluation metrics. The RMSE formula in Section III.B is presented in generic regression terms ('actual observation' vs. 'predicted value') without any mapping to the point-cloud setting. Neither the correspondence rule nor the existence of ground truth is described. This makes the reported numbers in Table 2 impossible to interpret or reproduce, which is a more load-bearing concern than the skeleton assumption because it undermines every quantitative comparison, not just generalization to difficult trees. I agree with the reader that the paper requires conditional acceptance, but I would emphasize the evaluation protocol as the primary condition. The reader already asks for 'an independent ground-truth evaluation,' so my concern does not change the verdict category; it refines the rationale. A concrete test of re-specifying and re-running the evaluation under a clear protocol would settle whether the accuracy numbers are trustworthy. If they are not, the central claim should be withdrawn. The verdict remains CONDITIONAL because the paper's method may still be sound, but the evidence currently presented is insufficient to assess it.","tokens_in":10239,"tokens_out":5525,"duration_ms":56699,"concrete_test":"Section III.B must specify: (i) the exact source and target point sets used for RMSE and Hausdorff evaluation; (ii) the correspondence rule (e.g., nearest-neighbor distance) and whether correspondences are symmetric; (iii) whether any independent ground truth (surveyed targets, known scanner poses, or manual alignment) was used. Then re-run the experiments on Tree01–Tree08 using a standardized protocol: align each scan pair with AMRST, ICP, and NDT under identical initial conditions, and report (a) mean/median/RMSE of nearest-neighbor distances from the transformed source to the target, (b) the symmetric Hausdorff distance, and (c) if ground truth is available, the absolute rotation/translation error. If the recomputed values do not match Table 2, or if no ground truth can be provided, the headline accuracy claim should be revised or removed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that AMRST 'significantly outperforms both ICP and NDT in registration accuracy' rests on the RMSE and Hausdorff values in Table 2. Section III.B provides only generic formulas for RMSE and Hausdorff distance. For RMSE, the paper defines y_i as 'actual observation' and y_i_hat as 'predicted value'—terms that have no direct meaning in point-cloud registration without a known correspondence or ground-truth transformation. The text never states which point sets are compared, how correspondences are obtained, or whether a marker-based or scanner-pose ground truth was used for evaluation. If RMSE is computed as the mean distance to the nearest neighbor after registration, then both ICP and NDT may perform worse simply because they lack a good initialization; but if it is computed using correspondences from the tetrahedra matching itself, the metric is circular and biased in favor of AMRST. Without this protocol, the reported RMSE and HD values cannot be reproduced or interpreted, and the claim of superior accuracy is unsupported. This is more fundamental than the skeleton-topology limitation (which the paper already acknowledges); it affects every accuracy number in the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes AMRST, a marker-free automatic registration method for single-tree terrestrial laser scanning point clouds. The pipeline separates wood and leaves, generates a tree skeleton from the branch points, extracts key points (branch points and endpoints) from the skeleton with a depth-filtering mechanism, and then finds pairs of similar tetrahedra in the two key-point sets to compute a coarse transformation. Fine registration is performed by applying Levenberg-Marquardt ICP to the leaf point clouds. The method is evaluated on eight trees of different species against ICP and NDT, reporting RMSE and Hausdorff distance as accuracy metrics and runtime as the efficiency metric. The paper claims that AMRST significantly outperforms both baselines in accuracy and is up to 593x and 113x faster than ICP and NDT, respectively.","tokens_in":10439,"tokens_out":3147,"duration_ms":31881,"significance":"If the reported results are reliable, the method would offer a practical, automatic, marker-free alternative for single-tree TLS registration, a task that is challenging due to the lack of distinct planar features and the structural complexity of tree canopies. The paper's main strengths are its coherent use of skeletal topology to create repeatable key points, the tetrahedron-based similarity matching for coarse alignment, and the explicit use of leaf point clouds for efficient fine registration. The speed advantages, if reproducible, would be valuable for forestry applications. However, the significance is substantially tempered by weaknesses in the evaluation: the accuracy metrics are undefined with respect to ground truth or correspondence, the method's free parameters are not reported, and the runtime comparison does not clarify which preprocessing steps are included. These issues currently prevent the central claims from being verified, so the paper's practical impact remains a promise rather than a demonstrated result.","major_comments":[{"comment":"The central accuracy claim that AMRST 'significantly outperforms both ICP and NDT' is not verifiable because the RMSE and Hausdorff distance are not defined with respect to any ground truth or correspondence protocol. The RMSE formula in Section II.F describes y_i as 'actual observation' and y_i_hat as 'predicted value', but in the context of point-cloud registration these terms have no operational meaning unless the text specifies which point sets are compared, how correspondences are established, and whether a known transformation (e.g., from scanner poses or artificial markers) was used as reference. If the RMSE is computed via nearest-neighbor distances after alignment, it may simply reflect initialization quality rather than true accuracy; if it is computed from the correspondences obtained by the tetrahedron matching itself, the metric is circular and biased in favor of AMRST. The Hausdorff distance is similarly ambiguous because no statement clarifies whether it is measured before or after fine registration and whether it uses the full point clouds or only the leaf/wood subsets. Without this protocol, the numerical values in Table 2 cannot be reproduced or interpreted, and the claimed superiority over ICP and NDT is unsupported.","section":"Section III.B, Table 2"},{"comment":"The free parameters of the method are not reported, making the experiments non-reproducible and the claimed robustness questionable. The edge-length tolerance epsilon, the volume-ratio tolerance beta, the SVD reconstruction threshold delta, and the depth threshold of five nodes for key-point selection are all introduced in Section II.D, and the paper states recommended ranges (epsilon between 0.01 and 0.1, beta less than 10%, delta less than 0.1) but never states the actual values used for the eight trees. If these parameters were tuned on the test data, as the phrasing 'based on experimental testing' suggests for the leaf-point-cloud choice, then the reported accuracy may reflect overfitting rather than generalizable performance. At minimum, the paper should report the parameter values and, ideally, a sensitivity analysis showing how the RMSE and HD vary with each parameter.","section":"Section II.D, Tables 2 and 3"},{"comment":"The runtime comparison lacks a clear accounting of which computational stages are included, so the speed advantage (up to 593x faster than ICP) may be inflated. The AMRST pipeline includes wood-leaf separation, skeleton generation, key-point extraction, tetrahedron matching, and fine registration; the table reports 'coarse registration time', 'fine registration time', and 'total time' for AMRST, but it is unclear whether the total time for ICP and NDT includes any preprocessing or is just the pure registration time on the full point cloud. Since AMRST operates on the skeleton and leaf point clouds, which are much smaller than the raw scans, the comparison is only meaningful if the cost of generating these reduced representations is included and if the baseline methods are given the same starting conditions. The paper should clarify the exact time breakdown and define the hardware/implementation details for all three methods.","section":"Section III.B, Table 3"},{"comment":"The paper's own discussion acknowledges that the method 'may struggle with trees whose branch structure is not clearly defined' and that dense leaves can obscure branches, but the experiments do not quantify the sensitivity to these factors. Given that the coarse registration is entirely dependent on the topological consistency of the skeleton between the two scans, and the wood-leaf separation method is only referenced (Section II.C, reference [34]) rather than described or evaluated, the paper should include a sensitivity analysis or at least report the number of key points extracted and matched per tree, the success/failure of tetrahedron matching, and the quality of the recovered transformation against a known ground truth. Without this information, the claimed robustness is not established by the presented evidence.","section":"Section IV"}],"minor_comments":[{"comment":"The phrase 'similar tetrahedras' uses an irregular plural; it should be 'similar tetrahedra'.","section":"Abstract and Title"},{"comment":"The reference to the wood-leaf separation method appears as '[34]' but the reference list only contains 29 entries, and item [28] corresponds to the Sun et al. paper on wood-leaf classification; the citation should be corrected.","section":"Section II.C"},{"comment":"The figure numbering is inconsistent: Figure 3 is described both as the demonstration of Tree04 preprocessing and as the skeleton generation, and Figure 7 is referenced twice in the text; the captions do not match the descriptions. The paper should be carefully proofread to fix figure references.","section":"Section III.A, Figures"},{"comment":"The equations for edge-length comparison, volume ratio, and the LM-ICP update are not numbered, making them awkward to reference; consider adding equation numbers for clarity.","section":"Section II.D"},{"comment":"The text states that for Tree01, the AMRST coarse registration RMSE is 0.08 while NDT is 3.83 and ICP is 0.65, and after fine registration RMSE is 0.06 and HD is 0.03; these values should be cross-checked against Table 2 because the table is not shown in the manuscript text and appears to have some formatting issues.","section":"Section III.B"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and the method description is coherent, but the evaluation is the weak point. The missing ground-truth protocol and unreported parameters are load-bearing issues that cannot be fixed by minor edits. I would encourage the authors to provide a clear evaluation protocol (e.g., using known scan positions or manually aligned reference point clouds), report the actual parameter values, and detail the runtime breakdown. If the authors can address these issues, the paper could become a valuable contribution to the forestry TLS registration literature."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the tetrahedron matching on skeleton key points is a reasonable new twist on skeleton-based registration, and the pipeline is clearly written. But the accuracy evaluation as reported is unverifiable: the RMSE and Hausdorff formulas in Section III.B have no defined point sets, correspondence, or ground-truth transformation, so every accuracy number in Table 2 floats on an unspecified protocol. That is the main soft spot, and it undermines the central claim.\n\nTo the paper's credit, the approach is concrete and the components are standard: wood-leaf separation, MST skeleton, key point extraction, tetrahedron congruence matching with volume/edge/SVD checks, then LM-ICP on leaf points. The speed advantage is plausible in direction because coarse registration gives ICP a good start; the reported runtimes are impressive, though implementation details could explain some of the gap.\n\nThe evaluation also never reports exact values for epsilon, beta, delta, or the depth threshold; only ranges/upper limits. That makes reproduction impossible and invites concern that they were tuned on the same eight trees. Only ICP and NDT are used as baselines; the natural comparison is against the skeleton-based methods of Bucksch and Zhou, which are cited but not evaluated. The paper does honestly note that the method may struggle when branch structure is unclear or leaves are dense, but that limitation is secondary to the measurement problem.\n\nWho should read it: forestry TLS people working on single-tree registration. It's a plausible preprocessing method, but the evidence is conditional. A serious referee should ask for the evaluation protocol to be spelled out—what exactly is compared, whether any ground truth (marker or scanner pose) exists—and for exact parameters, sensitivity, and skeleton-based baselines. I'd send it to review rather than desk reject, because the core idea is worth a proper look, but I would not cite the accuracy numbers until they are backed by a reproducible metric.","headline":"Plausible skeleton-based tetrahedron registration for single-tree point clouds, but the accuracy evaluation is unverifiable as reported.","tokens_in":11002,"tokens_out":2593,"would_cite":false,"duration_ms":24846,"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":"This paper proposes AMRST, a marker-free method that registers single-tree laser scans by matching similar tetrahedra built from branch skeletons, and reports substantially higher accuracy and speed than ICP and NDT.","keywords":["point cloud registration","single-tree","terrestrial laser scanning","marker-free","tetrahedra matching","skeleton key points","wood-leaf separation","LM-ICP"],"falsifier":"On a tree with dense foliage, scan the same tree from two positions and run the method: if coarse-registration RMSE jumps to the meter level (the range where ICP and NDT fail) instead of staying near the decimeter level reported for the eight test trees, the skeleton-consistency assumption is not met. Scanning the same tree in leaf and leaf-off would isolate how much of the registration quality depends on visible branch structure.","tokens_in":9988,"feed_emoji":"🌳","tokens_out":7111,"duration_ms":65056,"temperature":0.7,"pith_summary":"The paper introduces AMRST, a fully automatic, marker-free method for registering two terrestrial laser scans of the same tree. It separates each scan into wood and leaves, extracts a minimum-spanning-tree skeleton from the wood points, and selects key points where branches fork or end. It then searches both scans for pairs of similar tetrahedra whose vertices are those key points; matching four corresponding vertices supplies the coarse transformation, and a Levenberg-Marquardt ICP on the leaf points refines it. On eight trees of different species the method reports RMSE and Hausdorff distances far below those of ICP and NDT, with total runtimes of about a second to a few seconds per tree (up to 593x faster than ICP and 113x faster than NDT). If these results hold, complete single-tree point clouds for forestry measurements can be assembled without placing artificial targets in the field.","feed_headline":"Matching tetrahedra registers tree scans 593x faster","feed_subtitle":"A skeleton-based, marker-free method aligns single-tree point clouds in seconds with sub-0.1 m error","key_machinery":"The load-bearing object is the congruent tetrahedron pair defined on skeleton key points. The pipeline extracts an MST skeleton from the wood point cloud, labels branch points (nodes with more than two children) and endpoints (branch tips and the root), and keeps a depth-filtered set of the first five nodes along each root-to-tip path. It then forms tetrahedra from these key points, compares edge lengths with a relative tolerance epsilon, ranks tetrahedra by volume, checks that volume ratios are within beta, and verifies congruence by SVD with threshold delta. The four matched vertices of a congruent pair provide the coarse transformation; the fine stage then uses LM-ICP on leaf point clouds, minimizing point-to-point distance with a Levenberg-Marquardt update.","core_discovery":"On the paper's own terms, the central discovery is that branch skeletons give viewpoint-invariant key points for single-tree registration, and four such points define tetrahedra that are congruent across scans up to a small tolerance. The method builds all tetrahedra from each key-point set, filters edges by relative length error, sorts them by volume, and accepts a match when volume ratio is near one, sorted edges agree, and an SVD reconstruction error stays below threshold; the matched vertices then give the rigid transformation for coarse registration. The paper further claims that refining with leaf points rather than trunk points avoids the tight-fit problem of trunk geometry and converges quickly because the coarse alignment is already good. The reported outcome is coarse registration already more accurate than the full ICP and NDT baselines, and fine registration reducing error further (e.g., RMSE 0.06 m and Hausdorff distance 0.03 m on Tree01).","pith_inferences":["The tetrahedron-matching core is not tree-specific: any object scanned from two viewpoints that yields a stable skeleton (root systems, river deltas, piping) could be registered by the same congruent-tetrahedra search.","Because the matching stage is deliberately decoupled from wood-leaf separation, improvements in foliage removal should translate directly into better registration for dense canopies without redesigning the matching step.","The method leaves the question of heavy leaf occlusion open; a direct test would compare registration of the same tree scanned in leaf and after leaf-off, isolating the contribution of visible branching structure to registration success."],"forward_implications":["Forestry workflows can register single-tree scans without physical markers, removing a time-consuming field step.","The reported sub-0.1 m RMSE after fine registration is in the range useful for diameter-at-breast-height and biomass estimation.","Because total runtime is on the order of seconds per tree, the method can be applied to plot-level data in reasonable time.","The coarse registration's independence from the final ICP initialization means the pipeline does not need a user-provided initial alignment."],"supporting_citations":[{"why":"Supplies the automatic wood-leaf separation that produces the branch point cloud used for skeletonization.","marker":"[34]"},{"why":"Supplies the MST-based skeleton extraction whose branch points and endpoints become the key points.","marker":"[35]"},{"why":"Provides the tetrahedral congruence theorem used to justify edge-length-based tetrahedra matching.","marker":"[36]"},{"why":"Establishes the prior skeleton-based local registration approach this method builds on.","marker":"[26]"},{"why":"Establishes the prior automatic skeleton-based coarse registration approach.","marker":"[27]"},{"why":"Provides the Point Cloud Library and ICP baseline used in the method's implementation and comparisons.","marker":"[15]"},{"why":"Characterizes ICP convergence's dependence on initialization, motivating the coarse-to-fine design.","marker":"[19]"}],"fun_headline_variants":["Tetrahedra from skeletons align tree scans 593x faster","Marker-free tetrahedra registration for single-tree scans","Similar tetrahedra pair tree scans without markers","Skeleton key points form tetrahedra for quick tree registration","593x faster tree scan registration via tetrahedra"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the two scans' branch skeletons are topologically similar enough that branch points and endpoints can be matched; when dense foliage hides branches or branching is ambiguous, those key points and the tetrahedra built from them lose correspondence.","fun_headline_variants_meta":{"raw":{"variants":["Tetrahedra from skeletons align tree scans 593x faster","Marker-free tetrahedra registration for single-tree scans","Similar tetrahedra pair tree scans without markers","Skeleton key points form tetrahedra for quick tree registration","593x faster tree scan registration via tetrahedra"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000467,"raw_usage":{"total_tokens":2367,"prompt_tokens":1020,"completion_tokens":1347,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":636,"completion_tokens_details":{"reasoning_tokens":1267}},"tokens_in":636,"tokens_out":1347,"duration_ms":10751,"temperature":1.0,"reasoning_tokens":1267,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T16:52:22.944947+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a tree with dense foliage, scan the same tree from two positions and run the method: if coarse-registration RMSE jumps to the meter level (the range where ICP and NDT fail) instead of staying near the decimeter level reported for the eight test trees, the skeleton-consistency assumption is not met. Scanning the same tree in leaf and leaf-off would isolate how much of the registration quality depends on visible branch structure.","supporting_citations":[{"cited_title":"Localized registration of point clouds of botanic trees,","cited_arxiv_id":null,"evidence_quote":"Establishes the prior skeleton-based local registration approach this method builds on."},{"cited_title":"Automatic registration of tree point clouds from terrestrial LiDAR scanning for reconstructing the ground scene of vegetated surfaces,","cited_arxiv_id":null,"evidence_quote":"Establishes the prior automatic skeleton-based coarse registration approach."},{"cited_title":"3D is here: Point cloud library (PCL),","cited_arxiv_id":null,"evidence_quote":"Provides the Point Cloud Library and ICP baseline used in the method's implementation and comparisons."},{"cited_title":"Method for registration of 3-D shapes,","cited_arxiv_id":null,"evidence_quote":"Characterizes ICP convergence's dependence on initialization, motivating the coarse-to-fine design."}],"review_version":1}