{"id":"06419f16-5e6a-4735-8e7a-9a34013dd33a","arxiv_id":"2509.06862","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"A topology-adaptive deformation model that edits a template's topology during ARAP alignment achieves better Chamfer distance on meshes with topological artifacts than trained shape-matching baselines, at the cost of worse geodesic correspondence accuracy.","lead":"This paper proposes a method to match 3D surface meshes that contain topological errors, such as glued or broken parts, by gradually editing a template mesh's topology while deforming it to fit both shapes. The approach uses a neural signed-distance field for topology edits and is tested on human and animal scans, showing improved 3D overlap on corrupted reconstructions despite using no training data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline CD advantage over trained baselines is not established as a matching-quality improvement: the paper's own results show CD can improve while MGE worsens (left-right flips, topology-edit failure cases).","rationale":"The reader's conditional verdict is appropriate. The reader's weakest assumption focuses on the topology-update step: silhouette loss can improve while correspondence quality degrades. My concern is closely related but points at the evaluation rather than the optimization: the headline claim of 'better 3D alignment quality' is supported only by Chamfer distance, and the paper itself provides concrete examples where lower CD is accompanied by worse MGE. Therefore the CD advantage is not sufficient evidence for the stronger interpretation that the method improves shape matching. This is not a formal inconsistency in the algorithm, but it is a genuine soft spot in the argument as presented. The proposed per-pair delta analysis and correspondence-induced CD would settle whether the CD advantage is an artifact of optimizing geometric coverage independently of correct correspondences. I do not recommend changing the verdict because the reader already made the central claim conditional; I would keep the conditional verdict and ask the authors to add the correspondence-aware analysis and re-frame the claim accordingly.","tokens_in":20606,"tokens_out":9459,"duration_ms":109290,"concrete_test":"Use the public code to recompute Table 2 per shape pair. For every pair, record ours and Merrouche et al. CD and MGE. Count pairs where ours has lower CD but higher MGE; if more than ~20% of pairs show this divergence, the aggregate CD advantage is confounded. More direct: warp each source vertex to the target by the displacement of its nearest template vertex between \\tilde T_M and \\tilde T_N, compute CD between this warped source and the target, and compare with Merrouche et al.'s correspondences under the same protocol. If the correspondence-induced CD is not lower for ours, the proposed method's alignment advantage does not transfer to the matching task.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim rests on Table 2: CD 0.3342 on ExtFAUST and 0.1597 on 4DHumanOutfit vs 0.5795/0.2283 for Merrouche et al. But CD is not the objective being optimized; the deformation objective (Eq. 8) and topology-update objective (Eq. 10) use silhouette loss, a surface-overlap proxy. The paper's own Fig. 3 shows a left-right flip that yields smaller CD and significantly larger MGE, and Fig. 12 documents accepted topology edits that create disconnected components or suboptimal cuts. Thus lower CD can coexist with worse correspondences. In Table 2 MGE is worse for the proposed method (ExtFAUST 8.74 vs 7.32; 4DHumanOutfit 13.31 vs 6.15), so the CD advantage does not establish better matching. The load-bearing premise is that CD is a meaningful '3D alignment quality' measure in this matching context; the paper's own failure analysis undermines that premise.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper targets dense non-rigid mesh matching when the input meshes contain topological artifacts. It proposes a topology-adaptive deformation model that alternates between (i) finding bijective patch-level associations between a template and both input meshes, (ii) deforming the template with a patch-wise ARAP model, and (iii) editing the template's topology through its neural SDF representation using a silhouette-based objective. Correspondences are then extracted from the two deformed templates. The method is evaluated on FAUST, SCAPE, SMAL, a topologically corrupted subset of ExtFAUST, and 4DHumanOutfit, and compared with functional-map and deformation-guided baselines. The paper's central claim is that, without any data-driven prior, the method achieves better 3D alignment quality (measured by Chamfer distance) than learning-based methods, while being competitive on correspondence quality.","tokens_in":20964,"tokens_out":5218,"duration_ms":63326,"significance":"If the main claim holds, this is a useful contribution: it offers a purely optimization-based alternative to learning-based mesh matching that can explicitly adapt topology, which is relevant for real-world multi-view reconstructions. The paper is clearly written, the ablation study is informative, and the code is publicly announced. The idea of using neural SDF edits to resolve topological obstructions inside an ARAP alignment loop is original and worth publishing. However, the headline advantage over learned methods is metric-dependent and currently overstated: the Chamfer distance improvements are accompanied by worse mean geodesic error on the same benchmarks, and the topology-update acceptance rule is based on silhouette overlap, which the paper's own failure analysis shows can improve while correspondence quality degrades. The central mechanism is plausible but the evidence for the strongest claim needs to be rebalanced and quantified.","major_comments":[{"comment":"The claim that the method 'even outperform[s] methods trained on large datasets in 3D alignment quality' rests entirely on Chamfer distance. On the same topology-corrupted benchmarks, the proposed method has worse MGE than Merrouche et al. (ExtFAUST: 8.74 vs. 7.32; 4DHumanOutfit: 13.31 vs. 6.15). MGE is the standard correspondence-quality metric, and Chamfer distance is not the objective being optimized. Please qualify the abstract/conclusion claim, report both metrics together, state the number of test pairs, and provide error bars or a significance test. It would also be informative to report the CD/MGE numbers after excluding left-right flip cases (Fig. 3), since those flips are acknowledged to lower CD while increasing MGE.","section":"Abstract; Table 2"},{"comment":"Topology updates are accepted based solely on the silhouette loss. The paper's own examples show that this criterion can improve geometric overlap while degrading correspondences: Fig. 3 shows a left-right flip with smaller Chamfer distance and significantly larger MGE, and Fig. 12 documents accepted topology edits that create disconnected components or suboptimal cuts. Since the topology-adaptive deformation is the main novelty, the acceptance criterion should either incorporate a correspondence-aware or symmetry-aware term, or the paper should quantify how often such flips/failure cases occur in the reported averages and show that the CD advantage is not driven by them.","section":"Sec. 3.1.3; Algorithm 1 (lines 24-35); Fig. 3; Fig. 12"},{"comment":"The text says the template is initialized as either M or N, but Algorithm 1 and Sec. 8.6 fix T0 = M, and topology updates use only l_topo(N, \\tilde T_N^i). This asymmetry is not discussed or justified. Because the template choice and the one-sided topology editing can bias the extracted correspondences, the paper should either evaluate both initializations or explicitly justify the fixed choice. As written, the claim that the method is initialization-agnostic is not supported.","section":"Sec. 3.2; Algorithm 1; Sec. 8.6"}],"minor_comments":[{"comment":"Typo: 'neural filed representation' should be 'neural field representation'. Also, 'm-cubes' is used as a shorthand for marching cubes without defining it at first use.","section":"Sec. 3.1.3"},{"comment":"The caption and surrounding text contain missing numbers ('with   vertices', 'with   vertices'), apparently due to a compilation issue. Please fix so the discretization experiment is readable.","section":"Supplementary Fig. 6"},{"comment":"The displayed equation for l_sil has a line break and extra parentheses that make it hard to parse. Please re-set it cleanly.","section":"Eq. (9)"},{"comment":"The runtime (54 min per pair on ExtFAUST/4DHumanOutfit) is mentioned only in the supplementary. Since the method is optimization-based, a comparison of runtime with the learning-based baselines (including their test-time optimization) would help calibrate the practical contribution.","section":"Sec. 4"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the core idea is interesting. The main concern is that the 'outperforming trained methods' headline is not supported once MGE is considered, and the topology-update acceptance rule can improve CD while degrading correspondences. The authors should be asked to rebalance the claims and add quantitative evidence about flips/failure cases. The use of SMPL fittings from the authors' own prior work [52] for 4DHumanOutfit is disclosed but worth probing: it may favor the proposed method in ways that are not discussed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. The core idea is genuinely new: it edits a template's topology on the fly during ARAP deformation by updating a neural SDF, letting it match shapes with topological artifacts without any training data. That is worth a look. But the headline claim—outperforming trained methods in 3D alignment quality—rests on Chamfer distance, not on the field's standard correspondence metric. On MGE, the proposed method is worse than the learning-based baseline on both artifact datasets (8.74 vs 7.32 on ExtFAUST; 13.31 vs 6.15 on 4DHumanOutfit), and the paper's own left-right flip example shows CD improving while correspondence quality collapses.\n\nWhat the paper does well: the method is well-engineered, the ablations are informative, and the authors are unusually honest about failure modes. They document disconnected components and suboptimal cuts from the greedy topology editing, and they show that removing the ARAP constraint lowers CD but raises MGE—so they clearly understand that CD gains do not track correspondence quality. The code is public. On FAUST/SCAPE/SMAL it is the only non-learning method that does not completely fail on SMAL, which is a real result even if MGE is higher than learned methods.\n\nThe soft spots are mostly around claims, not engineering. CD is not the objective being optimized; silhouette overlap is, and the paper itself shows that optimizing that proxy can degrade correspondences (the left-right flip). There are no error bars or significance tests. The 4DHumanOutfit evaluation uses SMPL fittings from the authors' prior work, which is a reasonable benchmark but not independent. The topology update step is a heuristic with a greedy acceptance rule that can get stuck in local minima, as Fig. 12 shows.\n\nOverall: this deserves a serious referee. The contribution is real and the paper is honest, but the authors should re-frame the claimed contribution as better 3D alignment, not better matching, and add an analysis of when CD improvements correspond to genuine correspondence gains. Anyone working on non-rigid matching of multi-view reconstructions should read it.","headline":"The topology-adaptive deformation model is a genuine new contribution, but the paper's headline claim of beating trained methods rests on Chamfer distance alone, while the standard correspondence metric (MGE) shows the opposite.","tokens_in":21421,"tokens_out":3054,"would_cite":true,"duration_ms":30667,"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":"A topology-adaptive deformation model can match 3D meshes with topological artifacts without any learned prior, and reports tighter 3D alignment than trained baselines on noisy benchmarks.","keywords":["3D shape matching","topological artifacts","as-rigid-as-possible deformation","neural signed distance field","bijective associations","silhouette loss","non-rigid registration","mesh topology adaptation"],"falsifier":"Take two meshes with known ground-truth correspondences where one has a glued hand, run the optimization with the template initialized as the glue-affected mesh, and check whether the accepted topology edits remove the glue and reduce geodesic error. If the edits reduce silhouette loss but leave geodesic error unchanged or worse, or if no edit is accepted, the claimed topology-adaptive mechanism is not what produces the alignment.","tokens_in":20522,"feed_emoji":"🧩","tokens_out":11704,"duration_ms":122994,"temperature":0.7,"pith_summary":"Non-rigid mesh matching usually assumes the deformation between two shapes is near-isometric or as-rigid-as-possible, but real scans from multi-view reconstruction often have topological artifacts—glued hands, closed ankles—that break both assumptions. The paper proposes a topology-adaptive deformation model that removes the assumption by allowing the template mesh itself to change topology while it is being deformed to match two shapes. In the proposed optimization, the template is represented twice: as a mesh for patch-wise ARAP deformations under bijective associations, and as a neural signed-distance field whose zero level set can be edited and re-meshed when alignment fails. The central claim is that this training-free strategy handles highly non-isometric shapes and noisy per-frame reconstructions, and reports tighter 3D alignment (lower Chamfer distance) than learning-based baselines on topological benchmarks, although its ground-truth correspondence error is higher on clean benchmarks.","feed_headline":"Silhouette-guided topology edits align 3D meshes with glued parts","feed_subtitle":"On noisy scan benchmarks it reports tighter 3D alignment than trained baselines, without any data-driven prior.","key_machinery":"The central machinery is a dual representation of the template: a triangle mesh with a patch-wise rigid deformation model (ARAP, with rotations and translations per patch blended by distances), and a neural signed-distance field (a feature volume plus an MLP) whose zero level set is re-meshed after each topology update. The two are connected by differentiable iso-surface extraction, which converts vertex-level silhouette gradients into SDF parameter updates, so a topology edit is a gradient step in the space of implicit surfaces. Associations are kept bijective by representing them as a doubly stochastic patch-association matrix whose rows and columns sum to one, produced by a Sinkhorn proje","core_discovery":"The core claim is that topological differences between two shapes can be resolved during matching rather than repaired beforehand: the optimization grows, cuts, or reconnects regions of a template when the ARAP deformation under bijective associations cannot align them with the target. Topology edits are driven by a silhouette loss between the deformed template and the target; gradients with respect to template vertices are pushed through differentiable iso-surface extraction into the parameters of a neural signed-distance field, and the edited surface is extracted by marching cubes. An edit is accepted only when it lowers the total silhouette mismatch, which prevents the local patch-neighbo","pith_inferences":["Editorial inference: using a symmetry-aware or geodesic-consistency term to score topology edits could suppress the left-right flips the paper reports, since those flips lower Chamfer distance but greatly increase geodesic error.","Editorial inference: the greedy hill-climbing acceptance rule implies the result depends on the initial template and on the order of viewpoints; running the optimization with both input meshes as template and merging the two alignments could escape some local topology minima.","Editorial inference: the same topology-adaptive deformation model could be used beyond matching, for example to evolve a template through a sequence of per-frame reconstructions, which the authors name as future work."],"forward_implications":["Mesh matching with topological artifacts no longer needs a separate preprocessing repair step; the optimiser can split or reattach regions when the ARAP deformation under bijective associations is infeasible.","Because no training data is used, the method is applicable to shape classes or acquisition settings for which no pre-trained correspondence model exists, such as per-frame multi-view reconstructions of moving people.","The template found by the optimization is itself an output: it shows which regions of the source had to be topologically edited to achieve the alignments, which can be read as a form of automatic artifact detection.","The method's 3D alignment quality on topological benchmarks is reported as better than learning-based baselines, while its geodesic error remains higher; this places the contribution in alignment quality rather than in semantic correspondence accuracy."],"supporting_citations":[{"why":"Supplies the patch-based rigid deformation representation used to parameterize ARAP deformations of the template.","marker":"[17]"},{"why":"Provides the differentiable rasterizer used to compute silhouette loss between the deformed template and the target.","marker":"[39]"},{"why":"Defines the iso-surface extraction step that recovers an edited mesh from the neural field's zero level set.","marker":"[46]"},{"why":"The deformation-guided unsupervised mesh-matching baseline the paper extends and compares against on topological benchmarks.","marker":"[52]"},{"why":"Supplies the differentiable iso-surface extraction that propagates vertex-level silhouette gradients into the neural SDF parameters, enabling topology edits.","marker":"[58]"},{"why":"Defines the as-rigid-as-possible deformation energy that the rigidity constraint is based on.","marker":"[65]"},{"why":"Provides the Sinkhorn projection that turns inter-patch feature similarity scores into a doubly stochastic bijective association matrix.","marker":"[3, 51]"}],"fun_headline_variants":["Topology-adaptive deformation resolves mesh mismatches with silhouette-guided edits","Silhouette loss drives topology changes for robust 3D shape matching","Edit topology on the fly: ARAP alignment with silhouette-guided surface changes","Matching shapes with holes and glues: topology-adaptive deformation wins","No data prior needed: topology changes guided by silhouette errors"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The load-bearing premise is that the silhouette loss points to the exact places where the template's topology is wrong, so editing the neural field along that gradient improves the true correspondences—and the greedy acceptance rule does not trap the search in a worse local topology.","fun_headline_variants_meta":{"raw":{"variants":["Topology-adaptive deformation resolves mesh mismatches with silhouette-guided edits","Silhouette loss drives topology changes for robust 3D shape matching","Edit topology on the fly: ARAP alignment with silhouette-guided surface changes","Matching shapes with holes and glues: topology-adaptive deformation wins","No data prior needed: topology changes guided by silhouette errors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000255,"raw_usage":{"total_tokens":1390,"prompt_tokens":706,"completion_tokens":684,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":450,"completion_tokens_details":{"reasoning_tokens":592}},"tokens_in":450,"tokens_out":684,"duration_ms":6848,"temperature":1.0,"reasoning_tokens":592,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T22:57:48.134812+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take two meshes with known ground-truth correspondences where one has a glued hand, run the optimization with the template initialized as the glue-affected mesh, and check whether the accepted topology edits remove the glue and reduce geodesic error. If the edits reduce silhouette loss but leave geodesic error unchanged or worse, or if no edit is accepted, the claimed topology-adaptive mechanism is not what produces the alignment.","supporting_citations":[{"cited_title":"Free- form mesh tracking: a patch-based approach","cited_arxiv_id":null,"evidence_quote":"Supplies the patch-based rigid deformation representation used to parameterize ARAP deformations of the template."},{"cited_title":"Modular primitives for high-performance differentiable rendering.ACM Transac- tions on Graphics (ToG), 2020","cited_arxiv_id":null,"evidence_quote":"Provides the differentiable rasterizer used to compute silhouette loss between the deformed template and the target."},{"cited_title":"Lorensen and Harvey E","cited_arxiv_id":null,"evidence_quote":"Defines the iso-surface extraction step that recovers an edited mesh from the neural field's zero level set."},{"cited_title":"Deformation-guided unsuper- vised non-rigid shape matching","cited_arxiv_id":null,"evidence_quote":"The deformation-guided unsupervised mesh-matching baseline the paper extends and compares against on topological benchmarks."},{"cited_title":"Meshsdf: Differentiable iso-surface extraction","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable iso-surface extraction that propagates vertex-level silhouette gradients into the neural SDF parameters, enabling topology edits."},{"cited_title":"As-rigid-as-possible surface modeling","cited_arxiv_id":null,"evidence_quote":"Defines the as-rigid-as-possible deformation energy that the rigidity constraint is based on."}],"review_version":1}