{"id":"f81dec43-e1a7-49c0-a901-22ee6bc78b3e","arxiv_id":"2412.18696","paper_version":2,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"STITCH augments Neural-Pull with a topological loss derived from persistent homology to encourage a single connected component in reconstructed surfaces.","lead":"The authors add a persistent-homology-based loss to a neural implicit surface reconstruction method, encouraging the output to be a single connected component. They show that optimizing this combined loss with SGD converges, but the guarantee of connectivity is not actually proven, and the evaluation measures the very loss being optimized.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper never proves that optimizing the proposed loss yields a single connected component: Section 4.2's theorems are point-set density statements, not statements about the zero-level set of the learned SDF.","rationale":"STITCH's practical contribution—adding a persistent-homology-based loss to Neural-Pull and showing competitive Chamfer distances on several datasets—is plausible and may be useful. The tables report lower significant-feature loss L_S than baselines, and the ablations suggest the combined loss helps on thin structures. Those empirical results deserve credit, though they are weakened by the fact that the topological metric evaluated is the same loss being optimized and by the absence of direct connected-component counts for the extracted meshes. The reader correctly identifies a serious flaw in Theorem 1: the proof in Appendix C.2 assumes L_c is continuously differentiable, but the persistence map is piecewise differentiable, and the coface-based gradient in Eqs. 8-9 is not shown to be a true subgradient. That alone invalidates the stated convergence guarantee. However, the single most load-bearing problem is the connectivity claim. Even a fully valid convergence theorem would not imply the abstract's promise, because Section 4.2's Theorems 2 and 3 are discrete density and separation results adapted from Hofer et al. [28]; they never connect a critical point of L to the number of connected components of the zero-level set or of the marching-cubes mesh. Definition 2 as written is trivially satisfied by any finite set, and the proofs in the appendix do not bridge the gap to topology. Therefore the central theoretical contribution—a provable single-connected-component guarantee—is unsupported. I agree with the reader's REJECT verdict, but for a slightly different primary reason: the missing implication from objective to level-set topology, rather than the differentiability of the coface gradient alone, is where the argument fundamentally breaks.","tokens_in":25625,"tokens_out":7159,"duration_ms":73941,"concrete_test":"Independently re-derive Theorem 2 from Definition 2 as printed: if alpha is the minimum and beta is the maximum of the pairwise distances, then every finite set with at least two points is alpha~beta-connected, making the theorem's hypothesis vacuous and its conclusion a statement about point density, not level-set topology. Then run the proposed topological loss on a signed distance field whose zero level set has two components, e.g., f(x) = min(||x-c1||-r, ||x-c2||-r) for two disjoint spheres, and check whether L_c is bounded and whether the configuration is stationary; if a two-component zero level set can have finite, non-increasing L_c, the claimed connectivity guarantee collapses.","verdict_should_be":"REJECT","load_bearing_attack":"The abstract's central claim is that the loss provably 'enables reconstructing shapes with a single connected component,' but Section 4.2 does not establish this. Definition 2 defines alpha~beta-connectedness of a finite set D via alpha = min_i xi_i and beta = max_i xi_i over pairwise distances; as printed, any finite set with at least two points satisfies this condition, so it carries no information about the number of connected components. Theorem 2 concludes only that M is (m-k+1)~beta-dense, a neighborhood-count bound, and Theorem 3 concludes that M is not epsilon-separated under a cardinality condition. Neither statement refers to the cubical complex C_K, the SDF f_theta, the persistence diagram, or the extracted marching-cubes mesh. No lemma shows that a critical point of L (Eq. 4) has small L_c, nor that small L_c implies a single H_0 feature at the zero level. The convergence proof in Appendix C.2 is also unsound: it asserts that L_c is continuously differentiable, whereas persistence maps are only piecewise differentiable, and the coface-based gradient in Eqs. 8-9 is asserted without proving it is a valid subgradient. But the connectivity gap is more load-bearing: even if the convergence argument were repaired, the abstract's guarantee of a single connected component would still be unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes STITCH, a neural implicit surface reconstruction method that augments the Neural-Pull signed-distance-function loss with two persistent-homology-based topological losses computed from a cubical complex filtration. The stated contributions are a differentiable topological loss term that enforces connectivity, a convergence guarantee for stochastic (sub)gradient descent on the unified loss, and a theoretical guarantee that the optimized reconstruction has a single connected component. The evaluation compares STITCH against Poisson surface reconstruction, Neural-Pull, IGR, DiGS, OG-INR, and NSH on SRB, DFAUST, plant, and thin-structure datasets, reporting Chamfer and Hausdorff distances together with a 'significant features' topological loss; ablations of the loss components and geometric initialization are provided in the supplement.","tokens_in":25941,"tokens_out":6892,"duration_ms":82245,"significance":"The application area is timely and the proposed framework is clean: combining a pull-based SDF loss with a persistence-based connectivity term is a natural idea, and the experimental study is broad, including thin structures and plant geometries where topological preservation is important. The paper also provides ablations and detailed supplementary tables. However, the central theoretical claim—a provable guarantee of a single connected component—is not established. Definition 2 is vacuous, Theorems 2 and 3 are point-set density statements that never refer to the learned SDF or its level sets, and the convergence theorem rests on an unjustified differentiability assumption. The primary topology metric is the same loss being optimized, so the reported topological improvements are expected and do not constitute independent verification. If the connectivity guarantee were removed and the evaluation corrected, the paper would still be a useful empirical contribution, but not at the level claimed.","major_comments":[{"comment":"The abstract and Section 1 claim that optimizing the loss 'provably' leads to 'reconstructing shapes with a single connected component.' This is not supported by Section 4.2. Definition 2 defines an alpha~beta-connected set D by taking alpha = min_i xi_i and beta = max_i xi_i over pairwise distances; as printed, every finite set with at least two points satisfies this condition, so the definition carries no information about the number of connected components. Theorem 2 concludes only that M is (m-k+1)~beta-dense, a neighborhood-count bound, and Theorem 3 concludes that M is not epsilon-separated under a cardinality condition. Neither theorem refers to the cubical complex C_K, the SDF f_theta, the persistence diagram, or the extracted marching-cubes mesh. No lemma shows that a critical point of L (Eq. 4) has small L_c, nor that small L_c implies a single H_0 feature at the zero level set. The Limitations section at the end of Section 6 acknowledges edge cases and restriction to H_0 features but does not disclose the absence of a theorem connecting the optimized loss to a single connected component. Therefore the central connectivity guarantee is unsupported.","section":"Section 4.2 / Definition 2 / Theorems 2–3"},{"comment":"The proof of Theorem 1 is not sound as written. The proof asserts that 'L = L_g + L_c continuously differentiable' and Section 4.1 states that 'L_c is definable since it is differentiable.' However, persistent homology maps are only piecewise differentiable, and the coface-based gradient in Eqs. (8)–(9) is asserted to be the partial derivative of the topological loss without proof that it is a valid (sub)gradient. If L_c is not differentiable, the chain rule in Eqs. (6)–(7) is not justified, and the appeal to Corollary 5.9 of Davis et al. [21] requires hypotheses—tameness/definability and a correct subgradient calculus for persistence functions—that are not established for the specific L_c in Eq. (3). Thus Theorem 1, even if its conclusion were accepted, does not deliver a valid convergence guarantee for the STITCH objective.","section":"Appendix C.2 / Theorem 1 / Section 4.1"},{"comment":"The primary topological evaluation metric is the 'significant features topological loss term' L_S, which is exactly the loss term being minimized in Eq. (3). The lower values reported for STITCH in Tables 3, 16, 21, and 26 are therefore expected from the optimization and do not independently verify the claim of a single connected component. The paper needs an independent topology measure—for instance, the number of connected components of the extracted 256^3 marching-cubes mesh, or the number of H_0 features in the persistence diagram of the final SDF at a fixed threshold—to support the experimental connectivity claims.","section":"Section D.4 / Tables 3, 16, 21, 26"},{"comment":"Definition 7 states that a set D is epsilon-separated if for all distinct z, z' in D, ||z - z'|| <= epsilon. This is the opposite of the standard separation condition, which requires distances at least epsilon (or greater than epsilon). The proof of Theorem 3 uses the metric entropy N_epsilon with this definition, and the paragraph following Theorem 3 states the conclusion as 'there exist at least two points with distance smaller than epsilon,' which is the negation of the standard definition but not of the printed one. The misstatement makes Theorem 3 vacuous or ambiguous as written and needs correction.","section":"Definition 7 / Theorem 3 (Appendix C.2)"}],"minor_comments":[{"comment":"Assumption 1(a) contains a typesetting error: 'Σ_{t=1}^∞ = +∞' should read 'Σ_{t=1}^∞ α_t = +∞'.","section":"Section 4.1, Assumption 1(a)"},{"comment":"The sentence 'Set |N|= N and |S| = S' overloads N and S as both set names and cardinalities, which is confusing; consider using cardinalities n_N and n_S.","section":"Section 3.3"},{"comment":"The symbols 'a' in Eq. (8) and 'ϑ' in Eq. (9) are never defined; the magnitudes of the gradients should be given explicitly or their derivation described.","section":"Section 3.4, Eqs. (8)–(9)"},{"comment":"The sentence 'due to the new CVPR policy on presenting additional dataset results in the supplement' is irrelevant to the technical content and should be removed from the arXiv supplement.","section":"Supplement F"},{"comment":"The notation 'ξ_i' is introduced as the increasing sequence of pairwise distances, but the definition of 'α ~ β-connected' then simply restates the definitions of α and β; the condition should be reformulated to express a nontrivial connectivity requirement, or the concept should be replaced with a standard connectedness notion.","section":"Definition 2"}],"recommendation":"reject","confidential_remarks":"I concur with the stress-test assessment. The paper's advertised central contribution—a provable guarantee of a single connected component—is not established: the connectivity theorems concern finite point-set density and do not connect to the learned SDF or the extracted mesh, and the convergence proof relies on unjustified differentiability of the topological loss. The empirical topology metric is circular. The manuscript would need a substantially new theoretical argument connecting the optimized loss to the topology of the output surface, together with an independent evaluation metric, before it could be reconsidered."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the empirical recipe is plausible and the supplementary material is thorough, but the paper's headline guarantee—provable convergence to a single connected component—is not supported by the math as written.\n\nWhat is new: the paper attaches a persistent-homology-based loss (borrowed from Dong et al. 2022) to Neural-Pull, and evaluates it on several datasets with careful ablations. The experiments on SRB, DFAUST, plants, and thin structures, including the loss-component and level-set ablations in the supplement, are useful if you want to see how a topology-regularized SDF behaves on real sparse point clouds. That is legitimate work.\n\nWhere it falls apart: Section 4 does not prove what the abstract promises. Theorem 1 is a restatement of Davis et al. 2020, and its proof in Appendix C.2 simply asserts that Lc is continuously differentiable. That is not true for persistence-based losses; they are piecewise differentiable at best, and the coface gradient in Eqs. 8-9 is a heuristic with no proof that it is a valid subgradient. So the convergence guarantee, as presented, is not established.\n\nThe connectivity theorems are worse. Definition 2 defines alpha~beta-connectedness via alpha = min and beta = max of pairwise distances, which any finite set with at least two points satisfies vacuously. Theorem 2 then concludes that M is (m-k+1)~beta-dense, and Theorem 3 concludes M is not epsilon-separated under a cardinality condition. Neither statement refers to the cubical complex, the SDF, the persistence diagram, or the extracted mesh. Nothing connects a critical point of L to small Lc, and nothing connects small Lc to a single H0 feature at the zero level. The central claim of the paper is simply not proven.\n\nThe main topological evaluation metric is also problematic: the significant-feature loss LS in Table 3 is exactly the loss term being optimized (Eq. 3, Section D.4). Lower values are expected and do not independently verify connectivity. Meanwhile, the Chamfer and Hausdorff tables show STITCH mostly underperforms OG-INR and NSH on geometric accuracy. That might be acceptable if the topology evidence were strong, but it is circular.\n\nWho this is for: practitioners who want a drop-in topological regularizer for Neural-Pull may get value from the method and the ablation tables. The paper in its current form does not deserve publication, but it is substantive enough to deserve a serious referee. I would recommend sending it to review with a request for major revision: replace Section 4 with an honest statement of what is and is not shown, add a non-circular topology evaluation (e.g., counting components of the extracted mesh against ground truth), and tone down the abstract.","headline":"An incremental but well-executed application of Dong et al.'s topological loss to Neural-Pull; the theoretical claims don't survive contact with the paper.","tokens_in":26484,"tokens_out":3188,"would_cite":false,"duration_ms":30451,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["55N31","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A persistent-homology loss added to a signed-distance-field reconstruction makes sparse, irregular point clouds resolve into a single connected 2-manifold surface, with a convergence proof for stochastic gradient descent.","keywords":["neural implicit representation","signed distance function","persistent homology","topological loss","surface reconstruction","point cloud","connected components","stochastic gradient descent"],"falsifier":"Numerically compare the analytic topological-loss gradient from Equations 8 and 9 against central finite differences of the persistence loss along random SDF perturbations; a substantial mismatch at any non-degenerate filtration point would show that the coface route is not the true derivative and would break the differentiability premise of Theorem 1.","tokens_in":25452,"feed_emoji":"🔗","tokens_out":6710,"duration_ms":64102,"temperature":0.7,"pith_summary":"The paper proposes STITCH, a neural implicit surface reconstruction method that adds topological penalties to the usual signed-distance-field loss. The penalties come from persistent homology computed on a cubical complex built from the predicted SDF, and they push the zero-level set to have exactly one connected component. The authors prove that stochastic gradient descent on the combined loss converges to a critical point, and they argue that the resulting solution is a single connected 2-manifold. They demonstrate the method on engineering shapes, human scans, plant geometries, and thin structures, reporting lower topological-loss values than the baselines while keeping geometric error competitive. The intended payoff is watertight, simulation-ready meshes from sparse and irregularly spaced point clouds.","feed_headline":"Topology-aware loss forces neural surfaces into one connected piece","feed_subtitle":"Sparse, irregular point clouds become a single 2-manifold surface once persistent-homology penalties join the SDF loss.","key_machinery":"The central object is the persistence diagram of 0-dimensional homology (connected components) of a cubical complex filtered by the predicted SDF. The mechanism that carries the argument is coface-based gradient routing: a coface of a cell is a higher-dimensional cell containing it, and for each birth or death event the coface stores the grid indices whose SDF values define the feature's birth and death times. The topological loss gradients are assembled by looking up those coface indices, which makes the persistent-homology loss end-to-end differentiable. The other load-bearing piece is the definability assumption, stated through o-minimal structures, that lets the paper invoke a tame-function convergence theorem for SGD.","core_discovery":"On its own terms, the paper's central claim is that topology can be controlled during neural implicit surface reconstruction by making persistent homology differentiable through the coface structure of a cubical complex. Using the predicted SDF as the filtration function, every birth and death event in a persistence diagram is assigned to the cell whose addition created or destroyed the feature, and that cell's SDF value carries the gradient. This yields two loss terms, one suppressing noisy features near the diagonal and one preserving significant features, which are added to the Neural-Pull pull-loss. The paper further claims that the combined loss is definable and continuously differentiable under stated assumptions, so SGD converges almost surely to a critical point, and that the converged solution yields a single connected component rather than islands or gaps.","pith_inferences":["If the coface-gradient construction truly captures the persistence derivatives, the same mechanism could be applied to 1-dimensional and 2-dimensional persistent features, allowing a user to prescribe a full topological signature rather than just connectivity.","The convergence and connectivity claims are stated for the trained SDF, yet the final mesh is extracted at a much higher resolution than the coarse cubical complex used during training; testing whether the one-component guarantee survives that resolution gap would bound the practical claim.","The bias toward a single connected component means genuinely disconnected objects could be merged, so a natural testable extension is a user-specified component-count prior instead of always one.","The authors empirically use a negative level set to reduce inflation on some shapes, which suggests the connectivity guarantee may depend on the level-set value; probing that dependence would clarify how robust the topological conclusion is."],"forward_implications":["Reconstructed surfaces from sparse point clouds become watertight single-component 2-manifolds, so downstream meshing and physical simulation can use them directly.","The topological loss can be added to any SDF-based implicit reconstruction method, not only the Neural-Pull architecture used in the experiments.","The convergence proof extends persistence-optimization results beyond subgradient descent to stochastic gradient descent for a unified reconstruction-plus-topology loss.","The significant-feature topological loss is reduced by large margins relative to tested baselines on thin structures and plant geometries, which the authors attribute to the connectivity constraint.","Since only 0-dimensional features are used, the same framework is naturally extendable to loops and voids, at higher computational cost, as the paper notes."],"supporting_citations":[{"why":"It supplies the Neural-Pull signed-distance-field formulation whose pull-loss is the geometric component of the unified loss.","marker":"[37]"},{"why":"It introduces the noisy-feature and significant-feature persistence losses that the paper adapts into its topological loss terms.","marker":"[22]"},{"why":"It provides the connectedness definition and the density and separatedness lemmas used in Theorems 2 and 3.","marker":"[28]"},{"why":"It establishes that persistence maps are definable and locally Lipschitz, which the proof of Theorem 1 relies on.","marker":"[17]"},{"why":"Its tame-function stochastic subgradient convergence result is the proof template the paper follows for SGD convergence.","marker":"[21]"},{"why":"It supplies the implicit geometric regularization context and the geometric network initialization used in the architecture.","marker":"[24]"},{"why":"Poisson surface reconstruction is the classical implicit baseline against which the method's geometry and topology are compared.","marker":"[31]"}],"fun_headline_variants":["Topology loss keeps neural surfaces connected","Persistent homology steers implicit surface reconstruction","Differentiable topology for one-piece neural shapes","Neural surface reconstruction with enforced connectivity","Single connected surfaces via persistent homology loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the combined loss is smooth enough everywhere for the coface-based gradient to be a true descent direction, even though the persistence-diagram part of the loss is only piecewise differentiable and the coface assignment is a heuristic.","fun_headline_variants_meta":{"raw":{"variants":["Topology loss keeps neural surfaces connected","Persistent homology steers implicit surface reconstruction","Differentiable topology for one-piece neural shapes","Neural surface reconstruction with enforced connectivity","Single connected surfaces via persistent homology loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000142,"raw_usage":{"total_tokens":1105,"prompt_tokens":821,"completion_tokens":284,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":437,"completion_tokens_details":{"reasoning_tokens":221}},"tokens_in":437,"tokens_out":284,"duration_ms":3525,"temperature":1.0,"reasoning_tokens":221,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T04:33:00.371409+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Numerically compare the analytic topological-loss gradient from Equations 8 and 9 against central finite differences of the persistence loss along random SDF perturbations; a substantial mismatch at any non-degenerate filtration point would show that the coface route is not the true derivative and would break the differentiability premise of Theorem 1.","supporting_citations":[{"cited_title":"Topology- controllable implicit surface reconstruction based on persis- tent homology","cited_arxiv_id":null,"evidence_quote":"It introduces the noisy-feature and significant-feature persistence losses that the paper adapts into its topological loss terms."},{"cited_title":"Connectivity-optimized representation learning via persistent homology","cited_arxiv_id":null,"evidence_quote":"It provides the connectedness definition and the density and separatedness lemmas used in Theorems 2 and 3."},{"cited_title":"Optimizing persistent homology based functions","cited_arxiv_id":null,"evidence_quote":"It establishes that persistence maps are definable and locally Lipschitz, which the proof of Theorem 1 relies on."},{"cited_title":"Stochastic subgradient method converges on tame functions","cited_arxiv_id":null,"evidence_quote":"Its tame-function stochastic subgradient convergence result is the proof template the paper follows for SGD convergence."},{"cited_title":"Pois- son surface reconstruction","cited_arxiv_id":null,"evidence_quote":"Poisson surface reconstruction is the classical implicit baseline against which the method's geometry and topology are compared."}],"review_version":1}