{"id":"291b8562-af32-4e04-89d8-b1c0fc846af5","arxiv_id":"2507.20480","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"A new 3D-GS registration and fusion method using skeleton alignment, Mahalanobis-based KPConv features, and multi-factor Gaussian selection, with large but partly cherry-picked performance claims.","lead":"The paper proposes an automated pipeline for aligning and merging multiple 3D Gaussian Splatting scene maps using skeleton extraction, anisotropic convolution features, and a multi-factor fusion score. The claimed gains are large, but the central equations contain errors that make the method as written unreliable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (1) defines Gaussian-to-Skeleton distance independent of its left-hand argument, so the skeleton energy is not per-Gaussian and the core optimization is undefined as written.","rationale":"The reader's rejection is well-founded. The most load-bearing failure is Eq. (1). Because the right-hand side does not use the left-hand index except through the cluster G_j, D(h_i,S) is identical for all Gaussians in a cluster; the energy E(S) then counts each cluster's summed density multiple times, and the min over q_j is attracted to infinity as the exponential decays. No finite skeleton is selected by the data term. The text's claim that 'the first term aligns the skeleton with Gaussians' is therefore unsupported by the stated equation. This is an internal inconsistency in the definition of the main optimization objective, not merely a disagreement with a community baseline. The Algorithm 1 issue, where Stot uses S_ske twice instead of S_cen, is real but could be a typo; Eq. (1) cannot be dismissed as a typo without changing the entire skeleton-extraction method. The abstract's headline numbers are also single-scene best cases, but that is secondary to the undefined core formulation. I agree with the reader's weakest assumption and see no reason to alter the verdict.","tokens_in":12012,"tokens_out":4774,"duration_ms":49663,"concrete_test":"Re-implement the Sec. III-A skeleton optimization with Eq. (1) literally on a synthetic scene containing two well-separated Gaussian clusters and set λ=0; if the minimizing skeleton points do not converge to finite cluster centers but instead E(S) is minimized as q_j moves to infinity, the G2D distance is not a valid per-Gaussian data term and the method's skeleton extraction is undefined as written.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central skeleton-extraction step rests on Eq. (1), but as written D(h_i,S) = min_{q_j in S} sum_{i in G_j} exp(-0.5(q_j-µ_i)^T Σ_i^{-1}(q_j-µ_i)) does not depend on the Gaussian h_i on the left-hand side: the sum runs over all i in the cluster G_j, so every h_i in the same cluster receives the same value, and the index i is reused on both sides. Consequently, E(S) in Eq. (2) is not a sum of per-Gaussian distances but a cluster-level quantity. Worse, the data term is degenerate: since the exponential tends to 0 as q_j moves away from all cluster centers, the min over q_j is attained at infinity unless the curvature term alone prevents it. The G2D distance therefore cannot drive skeleton points toward the Gaussians. Because Sec. III-A and the skeleton-fusion scoring in Sec. III-C inherit this quantity, the core novelty is not mathematically defined as stated. The Algorithm 1 typo, where Stot uses S_ske twice and never S_cen, is secondary but reinforces the same impression that the formulas have not been checked against a working implementation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents an end-to-end framework for registering and fusing independently reconstructed 3D Gaussian Splatting (3D-GS) sub-maps. The method has three components: (i) a skeleton extraction and refinement procedure based on a proposed Gaussian-to-Skeleton distance and DBSCAN initialization; (ii) a Gaussian-adaptive KPConv feature extractor using Mahalanobis-distance neighborhoods and ellipsoid-aware kernels; and (iii) a multi-factor fusion score that balances skeleton adherence, local detail, and scene-center proximity. Experiments on the ScanNet-GSReg dataset and a self-collected Coord dataset compare the method against GaussReg and GaussReg-NF, reporting reduced rotation/translation/scale errors and improved PSNR, SSIM, and LPIPS.","tokens_in":12299,"tokens_out":6648,"duration_ms":68422,"significance":"If the technical content were correct, the framework would be a practically useful contribution to collaborative 3D-GS mapping: it removes manual reference-map selection and exploits Gaussian attributes beyond point positions, which are relevant for robotic perception. However, the central equations are internally inconsistent as written, the fusion algorithm does not match the described scoring scheme, and the empirical evaluation is limited to a single baseline. The paper does not provide code or machine-checked artifacts, so the reported gains cannot be independently verified. As submitted, the empirical claims are not supported.","major_comments":[{"comment":"The G2D distance D(h_i,S) as written does not depend on the left-hand Gaussian h_i: the right-hand side sums over all i in G_j, where G_j is the set of Gaussians associated with skeleton point q_j, so every Gaussian in the same cluster receives the same value. Moreover, the exponential is a similarity that tends to zero as q_j moves away from the cluster means, so minimizing E(S) in Eq. (2) does not attract skeleton points toward the Gaussians; without a bounded domain, the minimum is approached at infinity. Because the skeleton refinement in Sec. III-A and the fusion scoring in Sec. III-C both inherit this quantity, the core optimization is undefined as stated.","section":"Sec. III-A, Eq. (1) and Eq. (2)"},{"comment":"In the fusion loop, Stot(mu) is computed as alpha*Sske(mu) + beta*Sdeta(mu) + gamma*Sske(mu), so the skeleton-adherence term appears twice and the scene-center proximity score Scen, which the text in Sec. III-C describes as one of the three factors, never enters. The final selection and the soft-thresholding condition therefore do not implement the described multi-factor fusion strategy, and the reported fusion results cannot be attributed to the method as written.","section":"Algorithm 1"},{"comment":"The headline improvements in the abstract are not representative of the average results: the 41.9% RRE reduction is the improvement on scene 0089-01 only, while the average over the displayed ScanNet-GSReg scenes is approximately 24%, and the 10.11 dB PSNR gain is from scene 0715-00 only. The authors should report aggregate statistics over the full test set, including the number of scenes and variance, and should not present single-scene gains as general improvements.","section":"Abstract and Sec. IV-B, Table I and Table III"},{"comment":"The registration and fusion evaluations compare the proposed method only with GaussReg-NF and GaussReg, respectively. Given the paper's claim of a general framework, a single baseline is insufficient to establish competitiveness. The authors should compare with other recent 3D-GS registration and fusion methods and provide ablations for the contributions of GA-KPConv, the skeleton term, and each fusion factor. This is especially important because the method has many free parameters (alpha, beta, gamma, tau, lambda, DBSCAN epsilon and minPts, epsilon_skel, epsilon_overlap, sigma, M) whose sensitivity is not analyzed.","section":"Sec. IV-B and Sec. IV-C"}],"minor_comments":[{"comment":"The text contains a duplicated sentence: 'This weighted center weights denser Gaussians more in cluster formation...' appears twice, and the sentence about the first and second terms in Eq. (2) is repeated with identical wording, suggesting an editing error.","section":"Sec. III-A"},{"comment":"The subscript notation is inconsistent: the kernel weighting is written w_m, but the exponent uses d_k and x_k, and the relationship between the m and k indices is not defined.","section":"Sec. III-B, Eq. (4)"},{"comment":"The caption says 'Quantitative comparison of registration results' but the table reports rendering metrics (PSNR, SSIM, LPIPS) after fusion; the caption should say 'fusion results' instead.","section":"Table III caption"},{"comment":"The Avg. rows cannot be verified from the printed rows because of the ellipsis; the authors should state the total number of test scenes and report mean and standard deviation for each metric.","section":"Tables I and III"},{"comment":"The qualitative comparisons would be easier to assess with error maps or quantitative overlays; as printed, the visual differences between methods are difficult to judge.","section":"Figs. 3-6"}],"recommendation":"reject","confidential_remarks":"The manuscript's two central algorithmic components, Eq. (1) and Algorithm 1, are internally inconsistent as written: the G2D distance is not a per-Gaussian distance, and the fusion score uses S_ske twice instead of Scen. These are not minor typographical issues because they affect the definition of the method and the interpretation of the experimental results. The authors may have a working implementation, but the submitted text does not describe it accurately. The narrow comparison against a single baseline further weakens the empirical claims. I recommend rejection, though a substantially revised manuscript with corrected formulations, a broader set of baselines, and full ablations could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper tackles a real problem and the ingredients are sensible, but the central math as written doesn't define the proposed skeleton objective, and the abstract's headline gains are single-scene bests. I'd send it back for a serious rewrite before trusting any of the numbers.\n\nThe genuinely new bit is the combination: L1-medial skeleton extraction adapted to 3D-GS via covariance-aware distance, a Mahalanobis-distance KPConv variant for feature extraction, and a score-based fusion that balances skeleton adherence, local detail, and scene-center proximity. That combination is not in the prior literature, and the problem—automated registration/fusion of independently built 3D-GS sub-maps—matters for multi-robot mapping. The experiments cover a meaningful benchmark (ScanNet-GSReg) and show consistent average improvements over the stripped GaussReg baseline in Table I, with a clear qualitative difference in Fig. 3.\n\nBut the soft spots are serious. Eq. (1) defines D(h_i,S) with a right-hand side that sums over all Gaussians in a cluster and doesn't use i on the right. As a result, every Gaussian in a cluster gets the same distance, the energy E(S) is not a per-Gaussian fit, and the min over q_j is attained at infinity because the exponential decays to zero—so the skeleton extraction is undefined as written. That's not a nitpick; it's the foundation of both the initialization and the fusion scoring. If there's a typo, it needs to be stated explicitly with the correct equation. Algorithm 1 has a second, smaller typo: Stot uses S_ske twice and never Scen. The abstract's 41.9% RRE reduction and 10.11 dB PSNR gain are cherry-picked best scenes from Tables I and III; the averages are roughly 24% and 5 dB, which are still decent but not as advertised. The registration baseline is GaussReg without fine-tuning, so the comparison is weaker than the text implies. No code or data is released, and the Coord dataset appears to be private.\n\nOverall the idea is plausible and the direction is worth pursuing, but the paper as written is unreliable. Who's it for? Researchers working on 3D-GS map merging. They'll find the fusion scoring idea useful, and they should treat the numbers with caution. If I were the editor, I'd send it to review only with a strong warning about the equations, or desk-reject and invite a resubmission after major revision—either way, it needs a careful rewrite before it can be accepted.","headline":"A plausible 3D-GS registration/fusion idea undercut by a garbled core equation and cherry-picked headline numbers; worth a major-revision round, not acceptance as-is.","tokens_in":12868,"tokens_out":3588,"would_cite":false,"duration_ms":33568,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Independently built 3D-GS sub-maps can be aligned and fused automatically via skeleton alignment and covariance-aware features, cutting rotation error 41.9% and lifting PSNR by 10.11 dB.","keywords":["3D Gaussian Splatting","sub-map registration","skeleton extraction","Gaussian fusion","ellipsoid-aware convolution","multi-view mapping","point cloud registration"],"falsifier":"Take two Gaussians with different means or covariances that are assigned to the same skeleton point, evaluate Eq. (1) for both, and check whether the computed values differ; if they do not, the skeleton-refinement objective is not actually per-Gaussian, so the method's alignment and fusion scores would lose their stated geometric grounding.","tokens_in":11759,"feed_emoji":"🗺️","tokens_out":9804,"duration_ms":85278,"temperature":0.7,"pith_summary":"The paper tries to establish that 3D Gaussian Splatting sub-maps built from different viewpoints, such as a drone view and a ground vehicle view, can be aligned and fused without a human choosing a reference map and without a hard distance cutoff that deletes scene geometry. Its pipeline extracts a compact skeleton from each sub-map, registers the maps with features that adapt to the anisotropic shape of each Gaussian, and then merges overlapping Gaussians with a multi-factor score. If the claim holds, multi-robot or multi-view mapping could produce a single high-fidelity 3D-GS map automatically, with the reported gains of a 41.9% reduction in rotation registration error on a complex scene and a 10.11 dB PSNR improvement on a challenging fusion scene.","feed_headline":"Skeleton alignment cuts 3D-map registration error by 41.9%","feed_subtitle":"Automatic drone-vehicle Gaussian sub-map fusion also gains 10.11 dB PSNR.","key_machinery":"The load-bearing machinery is the pair formed by the Gaussian-to-Skeleton (G2D) distance and the Gaussian-Adaptive KPConv (GA-KPConv). G2D is intended to score how well each Gaussian fits a skeleton by evaluating a covariance-scaled Gaussian density at nearby skeleton points, with DBSCAN giving the initial skeleton and a curvature-regularized energy refining it. GA-KPConv selects neighborhoods by Mahalanobis distance using each Gaussian's covariance $\\Sigma_i$ and morphs spherical kernel points into ellipsoidal receptive fields through $\\Sigma_i^{-1}$, so local features respond to anisotropic structure; these features feed a geometric transformer that estimates the rigid transformation. The fusion step then computes, for each overlapping Gaussian, a weighted score $S_{tot} = \\alpha S_{ske} + \\beta S_{deta} + \\gamma S_{cen}$ (with the paper's Algorithm 1 also mixing in a max condition) to retain the better primitive instead of applying a hard threshold.","core_discovery":"The central claim is that two independently reconstructed 3D-GS sub-maps can be registered and fused automatically by exploiting the Gaussian geometry itself. A Gaussian-to-Skeleton distance, initialized by DBSCAN clustering and refined by an L1-style curvature-regularized energy, produces a structural skeleton for each sub-map; a Gaussian-adaptive KPConv uses Mahalanobis-distance neighborhoods and covariance-morphed ellipsoidal kernels to extract local features; and a soft fusion score balances skeleton adherence, local detail response, and scene-center proximity to decide which Gaussians survive in the overlap. The paper argues that this removes the manual reference-map selection and the rendering-quality loss caused by hard-threshold primitive filtering, and it reports reduced rotation and translation registration errors together with higher PSNR, SSIM, and lower LPIPS on the ScanNet-GSReg benchmark and its own Coord dataset.","pith_inferences":["If Eq. (1) is repaired so that the distance genuinely depends on the queried Gaussian $h_i$ (for instance, by taking the minimum over skeleton points of that Gaussian's density), the skeleton objective becomes well-defined and the reported fusion gains might be reproducible; as printed, the formula is ambiguous.","The same skeleton-plus-covariance machinery could plausibly serve as a loop-closure cue in incremental mapping, since a stable skeleton identity could signal when a robot re-enters a previously mapped region; the paper does not develop this use.","The covariance-shaped kernels suggest a natural extension to multi-sensor fusion where each primitive's covariance encodes sensor uncertainty, so LiDAR and vision maps could be registered with the same anisotropic features; this is an inference, not a claim in the paper.","The authors state in the conclusion that the feature extractor is geometry-only and lacks semantic awareness; adding semantic cues is a natural next step that would not require changing the skeleton or fusion machinery."],"forward_implications":["Registration no longer requires a manually selected reference sub-map, because either sub-map's skeleton can anchor the alignment and the fusion skeleton is built by merging both.","Complex and anisotropic scenes should register more accurately: the paper reports a 41.9% reduction in RRE and an 82.3% reduction in RTE on scene 0089-01, at the cost of about 21% more computation time.","Fusion should preserve structural detail that hard-threshold filtering destroys, with a reported 10.11 dB PSNR gain and 22.3% LPIPS reduction on an occluded scene, and higher average PSNR than the GaussReg baseline.","Skeleton quality improves in the same pipeline: at $\\epsilon=1.0$, minPts $=8$, the GA-L1 skeleton reaches 3.5% curvature deviation and 0.974 connectivity versus 7.9% and 0.920 for L1 extraction.","Ground-and-aerial mapping can be automated end to end, since the two sub-maps are fused into one coordinate system without manual scale or template selection."],"supporting_citations":[{"why":"Defines the 3D Gaussian Splatting representation and rendering model that the registration and fusion operate on.","marker":"[5]"},{"why":"GaussReg supplies the baseline registration and hard-threshold fusion approach that the paper compares against and extends.","marker":"[13]"},{"why":"GeoTransformer is the feature-matching backend that estimates the rigid transformation from the extracted Gaussian-adaptive features.","marker":"[22]"},{"why":"The Mahalanobis distance definition underlies both the anisotropic neighborhood selection and the ellipsoidal kernel weighting.","marker":"[26]"},{"why":"The L1-medial skeleton method is the skeleton-extraction approach that the proposed G2D distance generalizes to Gaussian splats.","marker":"[46]"},{"why":"DBSCAN clustering provides the density-based skeleton initialization without requiring a preset number of skeleton points.","marker":"[47]"}],"fun_headline_variants":["Automatic 3D-GS map fusion boosts PSNR by 10.11 dB","Skeleton-based 3D-GS registration cuts errors by 41.9%","Gaussian-adaptive features enable hands-free 3D map merging","Soft fusion preserves detail in 3D Gaussian splatting maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire skeleton pipeline assumes that Eq. (1) really measures each Gaussian's distance to the skeleton, but as printed the right-hand side sums over all Gaussians assigned to a skeleton point and never uses the index $i$, so the energy in Eq. (2) would not, taken literally, know which Gaussian is being fitted.","fun_headline_variants_meta":{"raw":{"variants":["Automatic 3D-GS map fusion boosts PSNR by 10.11 dB","Skeleton-based 3D-GS registration cuts errors by 41.9%","Gaussian-adaptive features enable hands-free 3D map merging","Soft fusion preserves detail in 3D Gaussian splatting maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000192,"raw_usage":{"total_tokens":1373,"prompt_tokens":996,"completion_tokens":377,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":292}},"tokens_in":612,"tokens_out":377,"duration_ms":3759,"temperature":1.0,"reasoning_tokens":292,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T17:43:34.693433+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two Gaussians with different means or covariances that are assigned to the same skeleton point, evaluate Eq. (1) for both, and check whether the computed values differ; if they do not, the skeleton-refinement objective is not actually per-Gaussian, so the method's alignment and fusion scores would lose their stated geometric grounding.","supporting_citations":[{"cited_title":"Kerbl, G","cited_arxiv_id":null,"evidence_quote":"Defines the 3D Gaussian Splatting representation and rendering model that the registration and fusion operate on."},{"cited_title":"Chang, Y","cited_arxiv_id":null,"evidence_quote":"GaussReg supplies the baseline registration and hard-threshold fusion approach that the paper compares against and extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"GeoTransformer is the feature-matching backend that estimates the rigid transformation from the extracted Gaussian-adaptive features."},{"cited_title":"On the generalized distance in statistics","cited_arxiv_id":null,"evidence_quote":"The Mahalanobis distance definition underlies both the anisotropic neighborhood selection and the ellipsoidal kernel weighting."},{"cited_title":"Huang, S","cited_arxiv_id":null,"evidence_quote":"The L1-medial skeleton method is the skeleton-extraction approach that the proposed G2D distance generalizes to Gaussian splats."}],"review_version":2}