{"id":"10a92278-c835-49b6-b009-5aee2e56be08","arxiv_id":"1908.09098","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A variational model using Beltrami coefficients and singular-value constraints jointly finds corresponding subregions and a bijective registration between partially overlapping surfaces.","lead":"This paper proposes an algorithm to register two 3D surfaces that only partially overlap, by simultaneously finding the matching regions and a smooth one-to-one map between them. The method uses quasiconformal geometry to control distortion and is tested on teeth, faces, and vertebrae.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Even granting maximality (1), the fidelity term in (11) is an unnormalized integral over f(X1)∩X2, so shrinking the overlap lowers E_fid; Algorithm 3 has no force to grow it, so the central claim of simultaneously finding the true overlap is unsubstantiated without a quantitative overlap test.","rationale":"The reader's CONDITIONAL verdict is reasonable, and I would not move it. The reader's weakest assumption, maximality (1), is a real scope restriction that the paper itself acknowledges. My concern is adjacent but distinct: even inside that scope, the optimization is degenerate with respect to overlap size. Equation (7) replaces the unknown Ω1 by f(X1)∩X2 using (1), but the energy (11) does not reward this intersection being large; the α and β terms penalize distortion, not small overlap. Landmarks are finite, and the admissible set S only bounds local singular values, so a smooth map can send most of X1 outside X2 while satisfying landmarks and keeping low distortion. Since the Demons update (20) is restricted to the current intersection, convergence to a small overlap is a plausible failure mode. The paper's experiments are qualitative and do not report overlap area, Dice, or ground-truth region error, so the claim that optimal regions are found is not empirically supported either. A synthetic overlap-recovery test with controlled initialization would settle this. This does not change the verdict: the paper is promising but conditional on such validation and on a justification or regularization of overlap size.","tokens_in":17781,"tokens_out":12694,"duration_ms":144742,"concrete_test":"Using the public GitHub code, run Algorithm 3 on the synthetic pair from Example 2 with a known ground-truth overlap under two initializations: identity and a smooth deformation that initially reduces f(X1)∩X2 by roughly 50% while respecting the same landmarks. Compare the Dice coefficient of the recovered region Ω1^*=(f*)^{-1}(f*(X1)∩X2) against the true Ω1, and record the final energy (11). If the shrunk-initialization run returns a markedly lower Dice with equal or lower final energy, the objective does not select the true corresponding regions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that Algorithm 3 simultaneously finds the optimal corresponding subregions and a bijective mapping. The weakest point is not only the scope restriction in Eq. (1); even granting Eq. (1), the relaxed objective (11) contains no term that rewards overlap area. E_fid is ∫_{f(X1)∩X2}(I1∘f^{-1}-I2)^2, an unnormalized integral. If the map f satisfies the landmark constraints but pushes almost all of X1 outside X2, this integral is essentially zero, and the regularization terms do not depend on where the intersection lies. The maximality condition f(Ω1)=Ω2=f(X1)∩X2 is an assumption about the unknown ground-truth map, not an objective term; the algorithm's output always satisfies it by defining Ω2 as the intersection, but nothing forces that intersection to be the true, large overlap. In Algorithm 3 (Section 5.4), the Demons displacement (20) is computed only on the current intersection g(X1)∩X2; points outside receive no intensity force, so the overlap has no expansion mechanism beyond landmark pull and smoothness. Consequently the variational model can converge to a small arbitrary overlap with low energy. The qualitative experiments do not report overlap area or Dice against ground truth, so they cannot rule out this degeneracy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses the inconsistent surface registration (ISR) problem, in which only subsets of two surfaces are in correspondence and the regions as well as the mapping between them must be found simultaneously. The authors propose a variational model whose fidelity term measures intensity mismatch on the current overlap f(X1)∩X2 and whose regularization controls distortion through the Beltrami coefficient and the singular values of the differential. A splitting algorithm (Algorithm 3) alternates among a projection onto singular-value bounds, Demons-style intensity matching, and geometric smoothing via thresholding of the Beltrami coefficient. The paper claims that bijectivity is easily enforced by thresholding the Beltrami coefficient and demonstrates the method on synthetic images, tooth surfaces, human faces, and vertebrae bones.","tokens_in":18188,"tokens_out":6175,"duration_ms":64544,"significance":"If the claims are substantiated, the paper would provide a practical method for partial surface registration without a pre-specified overlap, combining quasiconformal regularization with free-boundary deformation. The paper is generally clearly written, includes a public implementation, and presents experiments on several surface types, which are assets. However, the central technical claims are not fully established: the relaxed fidelity term does not reward overlap area, the bijectivity enforcement is heuristic, and the projection step lacks convergence guarantees. These gaps are load-bearing because the paper's main contribution is the simultaneous discovery of corresponding regions and a bijective, low-distortion mapping.","major_comments":[{"comment":"The relaxation in Eq. (7) makes E_fid an unnormalized integral over f(X1)∩X2, so the fidelity term alone is minimized by an empty intersection, as the authors themselves note. The regularization and landmark terms are not shown to prevent convergence to a small arbitrary overlap; none of the experiments (e.g., Figures 5, 9, 13, 17, 21, 25) report the area or Dice overlap of the recovered corresponding regions against ground truth, so the claim that Algorithm 3 simultaneously finds the optimal corresponding regions is not quantitatively supported.","section":"Section 4.1, Eq. (7)"},{"comment":"Thresholding the Beltrami coefficient to zero wherever |μ|≥1 yields a coefficient with L∞ norm <1, but the paper does not prove that the map reconstructed by the Linear Beltrami Solver is globally bijective or that its Beltrami coefficient obeys the bound; the experiments do not report the final |μ|∞ or any fold counts of the output maps. Since the abstract lists bijectivity enforcement as a key feature, a theoretical or at least numerical verification is needed.","section":"Section 5.3, Eq. (18) and Section 5.4, Eq. (21)"},{"comment":"The iterative projection onto S is not shown to converge to a map whose differential satisfies σ2≥K2 and σ1≤K1. Each Poisson solve (15) only approximates the target differential in a least-squares sense, and no fixed-point argument or convergence test (e.g., reporting max σ1 and min σ2 after each iteration) is provided; this gap affects Algorithms 2 and 3 because S is part of the admissible set A.","section":"Section 5.1, Algorithm 1"},{"comment":"The Demons displacement V(j) is defined only on the current intersection g_{N1,j}(X1)∩X2 (Eq. (20)), but the algorithm then applies V(j) to the whole map g to form \\tilde g without specifying how V is extended to all of X1. As a result, the update step is ambiguous, and points outside the current intersection receive no intensity force, so the overlap has no expansion mechanism beyond landmark pull and smoothness. This reinforces the degeneracy described in the first major comment and also hampers reproducibility.","section":"Section 5.4, Algorithm 3"}],"minor_comments":[{"comment":"The final paragraph states that experimental results are shown in Section 5, but the experiments actually appear in Section 6; Section 5 contains the algorithms.","section":"Section 1"},{"comment":"The caption says both 'The right shows the target static surface' and 'The right shows the difference of intensities on the registered surface'; these cannot both be correct and the caption should be corrected.","section":"Figure 18 caption"},{"comment":"The symbol f is used both for the registration map restricted to Ω1 and for its extension to all of X1; introducing a separate notation would reduce confusion.","section":"Section 4, Eq. (7)"},{"comment":"The claim that 'Bijectivity of the mapping is easily enforced' is stronger than what the thresholding step actually demonstrates; please qualify it as a heuristic or as local bijectivity in the abstract.","section":"Abstract and Section 5.4"}],"recommendation":"major_revision","confidential_remarks":"The stress-test concern about the unnormalized fidelity term is confirmed by reading Section 4.1. The absence of any overlap-area reward or quantitative overlap evaluation is the main obstacle to accepting the central claim; the remaining issues about bijectivity enforcement and the projection step are also load-bearing but could be addressed with additional numerical verification."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThis paper is worth a look, but the headline claim is softer than it appears. The genuinely new thing is a variational formulation for partial-overlap surface registration that treats the overlap as unknown and optimizes a map between the discovered subregions, using the Beltrami coefficient for smoothness and singular-value bounds for scale control. That is a sensible and nontrivial extension of the authors' earlier quasiconformal registration work, and the code is public. The free-boundary deformation part (Algorithm 2) is a reasonable engineering contribution.\n\nThe main problem is that the objective does not reward overlap. Equation (11) has fidelity term ∫_{f(X1)∩X2}(I1∘f^{-1}-I2)^2, an unnormalized integral. Any map that keeps the landmarks in X2 but pushes most of X1 outside X2 pays no fidelity penalty for the lost overlap. The Demons step in Algorithm 3 is only applied on the current intersection, so nothing forces the overlap to grow. Thus the claim of “simultaneously searching for the optimal corresponding regions” is not supported by the energy or the algorithm. The authors acknowledge the empty-intersection triviality, but only say that regularization and landmarks avoid it; that is not a proof, and as stated the fidelity term actually prefers smaller overlap. Without a quantitative overlap test against ground truth, the experiments cannot rule out convergence to a small arbitrary intersection.\n\nOther soft spots are real but less severe. The thresholding of the Beltrami coefficient (Eqs. 18 and 21) sets μ=0 where |μ|≥1, but there is no proof that the reconstructed map is globally bijective or that the singular-value bounds are preserved. Algorithm 1 is an iterative projection with no convergence proof. Hyperparameters are tuned per dataset, and the evaluation is almost entirely visual: no Dice score, no overlap area, no sensitivity analysis. These are not fatal if the paper is read as a heuristic method paper, but they do limit how strongly the central claims can be stated.\n\nThe citation pattern is fine: the authors cite their own Beltrami solver and prior quasiconformal registration work, which is appropriate given the method builds directly on those tools. No red flag there.\n\nWho is this for? Researchers working on partial shape registration who want a practical quasiconformal pipeline and are willing to supply their own overlap initialization or verification. I would not cite this paper as evidence that joint overlap discovery works, but I might cite the free-boundary distortion-controlled deformation formulation. It deserves peer review because the problem is important and the method is plausibly useful, but it needs major revision: add a term or mechanism that actually rewards overlap, prove or at least carefully test the thresholding step, and report overlap quality quantitatively.","headline":"A useful quasiconformal pipeline for partial-overlap registration, but the central claim that the algorithm discovers the true overlap is not backed by the objective or the experiments.","tokens_in":18624,"tokens_out":2764,"would_cite":true,"duration_ms":31263,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["30C62","65D18"],"pacs":[],"model":"deepseek-v4-flash","headline":"Partial surfaces with unknown overlap can be registered in one pass by minimizing a single variational energy that couples intensity matching with quasiconformal distortion control.","keywords":["inconsistent surface registration","Beltrami coefficient","quasiconformal mapping","variational model","mapping distortion","free boundary deformation","surface registration","bijective mapping"],"falsifier":"Run the algorithm on a synthetic pair where the true correspondence consists of two separated patches with matching intensities; if the recovered overlap collapses to one patch, shrinks to empty, or the fidelity term cannot detect the second patch, the maximality assumption is violated and the model's fidelity term is not measuring the true registration error.","tokens_in":17569,"feed_emoji":"🧩","tokens_out":5739,"duration_ms":56162,"temperature":0.7,"pith_summary":"Two surfaces often correspond only on unknown partial regions, and standard registration fails because it assumes a global bijection. This paper claims that the inconsistent surface registration problem can be solved as a single variational problem: minimize an energy that combines intensity mismatch on the current overlap, smoothness of the Beltrami coefficient, and control of the differential's singular values, while the overlap itself is found as the intersection of the deformed source with the target. A splitting algorithm alternates between updating the map and updating the Beltrami coefficient, which keeps the map locally bijective by simple thresholding. If this works as demonstrated, partial-surface registration in graphics and medical imaging would no longer require the user to specify corresponding regions in advance.","feed_headline":"Partial surfaces registered without pre-specified overlap","feed_subtitle":"A quasiconformal energy and splitting scheme recover the overlap and a low-distortion bijection in one pass.","key_machinery":"The machinery is the Beltrami coefficient $\\mu(f)$ from quasiconformal theory, a complex-valued field that encodes infinitesimal angle distortion, together with the singular values $\\sigma_1(Df),\\sigma_2(Df)$ encoding scale distortion. The paper splits the energy by introducing an auxiliary field $\\nu$, and alternates: for fixed $f$, smooth $\\nu$ by gradient descent with thresholding $|\\mu|<1$; for fixed $\\nu$, reconstruct $f$ by solving the Beltrami equation with a linear elliptic solver and enforce singular-value bounds by projection. This lets the algorithm evolve the boundary freely rather than fixing it in advance.","core_discovery":"The central claim is that the optimal corresponding subregions and the optimal bijection between them can be recovered together by minimizing the relaxed energy (11), in which the fidelity term integrates intensity mismatch over $f(X_1)\\cap X_2$ and the regularization separates into a term tying the map's Beltrami coefficient $\\mu(f)$ to an auxiliary field $\\nu$ and a smoothing term on $\\nu$. The authors show that the bijectivity constraint reduces to thresholding $|\\mu|<1$, and the scale constraint reduces to projecting each differential $Df$ onto matrices with singular values between $K_2$ and $K_1$ via an explicit singular-value formula. They then obtain the corresponding regions as $\\Omega_2^* = f^*(X_1)\\cap X_2$ and $\\Omega_1^*=(f^*)^{-1}(\\Omega_2^*)$.","pith_inferences":["A natural extension is to drop or relax the maximality property so that multiple disconnected corresponding regions are allowed; this would require a different fidelity term than $\\int_{f(X_1)\\cap X_2}$.","The dependence on Demons-style displacement updates suggests the method inherits sensitivity to intensity noise at high-gradient regions, so a multiscale or robust cost could stabilize convergence.","One testable extension is to replace curvature with learned or semantic descriptors and check whether the same energy recovers partial correspondences without manual landmarks.","The free-boundary deformation algorithm itself could be reused outside registration, for example in shape deformation with controlled area and angle distortion."],"forward_implications":["The overlap region need not be prescribed: it is recovered as $f^*(X_1)\\cap X_2$ after optimization.","Bijectivity is maintained by thresholding the Beltrami coefficient, so fold-free registration is obtained without complicated injectivity constraints.","The same pipeline handles large, non-rigid deformations on disk-type and genus-one surfaces after conformal flattening.","The per-triangle projection and per-vertex intensity updates parallelize, so the algorithm is practical on meshes with thousands of vertices.","In cases with no true overlap, the fidelity term alone would vanish, but the landmark and regularization terms keep the iteration from collapsing to the trivial empty solution."],"supporting_citations":[{"why":"Supplies the quasiconformal theory: the Beltrami equation, measurable Riemann mapping theorem, and regularity estimates used to justify the smoothing energy.","marker":"[3]"},{"why":"Establishes the landmark- and intensity-based quasiconformal registration formulation this work extends to inconsistent domains.","marker":"[10]"},{"why":"Provides the least-squares conformal flattening used to reduce surface registration to two planar domains.","marker":"[11]"},{"why":"Gives the Linear Beltrami Solver used to reconstruct the map from its Beltrami coefficient at each splitting step.","marker":"[19]"},{"why":"Supplies the Demons displacement update used for the intensity-matching subproblem in equation (20).","marker":"[25]"},{"why":"Provides the non-rigid ICP baseline compared against in the mammalian tooth experiment.","marker":"[2]"},{"why":"Provides the rigid ICP baseline whose failure motivates the need for deformable inconsistent registration.","marker":"[22]"}],"fun_headline_variants":["Surface registration finds overlap and mapping together","No pre-specified overlap needed for surface registration","Quasiconformal method jointly detects overlap and maps surfaces","Simultaneous overlap detection and bijective surface mapping","Recover partial surface overlap and low-distortion bijection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the true corresponding region is exactly the intersection of the deformed whole source surface with the target surface, so that one connected overlap captures all meaningful correspondence.","fun_headline_variants_meta":{"raw":{"variants":["Surface registration finds overlap and mapping together","No pre-specified overlap needed for surface registration","Quasiconformal method jointly detects overlap and maps surfaces","Simultaneous overlap detection and bijective surface mapping","Recover partial surface overlap and low-distortion bijection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000178,"raw_usage":{"total_tokens":1288,"prompt_tokens":927,"completion_tokens":361,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":543,"completion_tokens_details":{"reasoning_tokens":287}},"tokens_in":543,"tokens_out":361,"duration_ms":4149,"temperature":1.0,"reasoning_tokens":287,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:21:31.699392+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the algorithm on a synthetic pair where the true correspondence consists of two separated patches with matching intensities; if the recovered overlap collapses to one patch, shrinks to empty, or the fidelity term cannot detect the second patch, the maximality assumption is violated and the model's fidelity term is not measuring the true registration error.","supporting_citations":[{"cited_title":"Princeton University Press (2008) 3, 4, 5","cited_arxiv_id":null,"evidence_quote":"Supplies the quasiconformal theory: the Beltrami equation, measurable Riemann mapping theorem, and regularity estimates used to justify the smoothing energy."},{"cited_title":"SIAM Journal on Imaging Sciences 7(4), 2364–2392 (2014) 3, 7","cited_arxiv_id":null,"evidence_quote":"Establishes the landmark- and intensity-based quasiconformal registration formulation this work extends to inconsistent domains."},{"cited_title":"In: ACM transactions on graphics (TOG), vol","cited_arxiv_id":null,"evidence_quote":"Provides the least-squares conformal flattening used to reduce surface registration to two planar domains."},{"cited_title":"Computing quasiconformal folds","cited_arxiv_id":"1804.03936","evidence_quote":"Gives the Linear Beltrami Solver used to reconstruct the map from its Beltrami coefficient at each splitting step."},{"cited_title":"Medical image analysis 2(3), 243–260 (1998) 11, 12","cited_arxiv_id":null,"evidence_quote":"Supplies the Demons displacement update used for the intensity-matching subproblem in equation (20)."},{"cited_title":"In: 2007 IEEE Conference on Computer Vision and Pattern Recognition, pp","cited_arxiv_id":null,"evidence_quote":"Provides the non-rigid ICP baseline compared against in the mammalian tooth experiment."},{"cited_title":"In: 3dim, vol","cited_arxiv_id":null,"evidence_quote":"Provides the rigid ICP baseline whose failure motivates the need for deformable inconsistent registration."}],"review_version":1}