acceptodds
Withdrawn from ICLR 2027

LOGOS: Residual-Gated Witness Refinement for Non-Rigid Shape Correspondence

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

A correspondence sends every vertex of one mesh to a vertex of another. Truncated spectral bases blur small parts and still swap left with right. An update that moves each vertex on its own can tear one region into islands. A network trained on one connectivity is accurate there and degrades when the mesh changes. is a training-free editor of a hard correspondence. A residual gate reassigns only high-residual vertices, cleanup freezes the largest piece of each cluster, and a polish moves points inside that cluster. From the same SHOT nearest-neighbor start it lowers mean geodesic error on all six splits, from ///// to ///// on SCAPE, SCAPE_r, FAUST_r, TOSCA_r, TOPKIDS, and SMAL_r, and it is the lowest stored mean on TOSCA_r (). Pushed labels on SCAPE_r are contiguous on of clusters. A FAUST-trained network is at home and on TOSCA_r. On TOPKIDS, 2D-GEM is and is .

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Laplacian, Shape Matching, Random Graph Matching, Functional Maps

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_nuFYOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

My understanding is that this paper aims to develop a “repair tool for correspondence tables.” Given a source mesh, a target mesh, and an initial correspondence, LOGOS modifies the target-to-source vertex index table to improve correspondence accuracy and keep the mapped regions as connected as possible.

Critical strengths
  1. As shown in Table 1, the method substantially reduces the error across all six datasets.
  2. The method is clearly defined and can be applied to correspondences produced by different initialization methods.
  3. The authors explicitly acknowledge the differences between the theoretical assumptions and the actual implementation, as well as limitations such as the stronger performance of in-domain networks and the shared-index setting.
Critical weaknesses
  1. The novelty of the paper is not entirely clear to me, particularly compared with the existing method HOPE.
  2. Although the proposed improvements are effective, it remains unclear whether they are substantial enough to meet the acceptance standard for ICLR. I therefore remain cautious and would like to consider the other reviewers’ assessments.
Questions for authors
  1. The method consists of three steps: reassignment, fragment cleanup, and within-cluster smoothing. Which step contributes the most to the improvements in correspondence accuracy and region connectivity? Are these contributions consistent across datasets?
  2. LOGOS makes the mapped regions more connected, but it also maps multiple target vertices to the same source vertex. Could this lead to a loss of detail or introduce distortion?
Reviewer_AUeHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LOGOS refines an existing hard vertex-to-vertex correspondence between two non-rigid meshes in three stages. First, a residual gate selects vertices whose 1-hop adjacency disagrees with the transported source adjacency, and these are reassigned by greedy matching on k-hop overlap scores (hops 1–8). Second, labels pushed from a source Voronoi partition are cleaned up so that each cluster keeps only its largest connected piece. Third, the most displaced vertices are snapped to the nearest source vertex inside their cluster. Starting from SHOT nearest neighbours, it lowers the mean geodesic error on six splits (SCAPE, SCAPE_r, FAUST_r, TOSCA_r, TOPKIDS, SMAL_r), and it is also evaluated as a refiner of trained ULRSSM maps and in terms of cluster contiguity. LOGOS is a training-free combinatorial refiner: it has no learned component and uses trained networks only as initializers and baselines.

Critical strengths

S1. The design is complete and the validation is thorough. Each stage targets one failure mode identified in the introduction (Sec. 3.6). The evaluation covers six splits, three initializers and nine method families (Table 1), measures contiguity separately (Table 3), and includes leave-one-stage ablations (Table 4), a parameter grid (Table 7) and seed sensitivity (Table 8). Settings where the method does not lead are also reported.

Critical weaknesses

W1. Limited methodological increment and limited learning insight. The method uses a k-hop overlap score similar to HOPE's and the same 1–8 hop range (App. I). The additions are a different gating criterion (a 1-hop adjacency residual), greedy instead of exact matching, a witness perturbation, and two post-processing steps (label cleanup and in-cluster projection). The leave-one-stage ablation shows small individual effects on SCAPE_r: 0.035 by default, 0.037–0.041 with one stage removed, and 0.033 with the perturbation removed (Table 4). As a training-free combinatorial refiner, the paper offers limited insight into representation learning for the ICLR audience.

W2. Limited robustness. The method is heuristic by the authors' own account: the score is a "local consensus heuristic" (Remark 2), and App. B states that the propositions give no recovery guarantee. Its failure modes are visible in the results:

  • Left–right flips are not resolved (App. K).
  • 22.5% of SMAL_r pairs have an error above 0.1 (Sec. 4.8).
  • On SMAL_r the method (0.056) trails both its own landmark start (0.033) and NAM-ULRSSM-smal (0.025).
  • Results are sensitive to fixed settings: on SCAPE, qlap=0.50q_{lap}=0.50 raises the error from 0.013 to 0.060, and on SMAL_r removing the perturbation raises it from 0.078 to 0.097 (Table 7).
Questions for authors

Q1. Recent shape-matching work (e.g., Hybrid Functional Maps, CVPR 2024) also evaluates on DT4D-H (Magnet et al., 3DV 2022) and SHREC'19, which contain non-isometric, cross-category and connectivity-varying pairs. How does LOGOS perform on these benchmarks? The DT4D-H inter-class setting directly tests the change-of-category scenario the paper emphasises.

Q2. TT denotes both the number of iterations (T=32T=32) and the witness perturbation operator (Algorithm 1; it appears twice in Table 5). Please disambiguate the notation.

Reviewer_dX5gOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper proposes LOGOS, a training-free method for refining non-rigid 3D shape correspondences. Starting from an existing correspondence, it combines residual-gated graph matching, geodesic cluster cleanup, and local smoothing to improve accuracy and spatial coherence.

Experiments on six benchmark datasets show substantial improvements over SHOT initialization and competitive performance against established methods, including the lowest reported geodesic error on TOSCA r. The method also improves spatial contiguity, although it does not guarantee bijectivity and is outperformed by specialized methods on some benchmarks.

Critical strengths

A key strength of the paper, highlighted by the authors, is that the training-freepropoerty. In particular, this is useful when dataset-specific training is unavailable. The method demonstrates substantial improvements over the initial SHOT correspondences across all six evaluated datasets and consistently outperforms NAM-SHOT when using the same initialization. The results are competitive with established approaches, including the lowest reported mean geodesic error on TOSCA r.

Another strength is the relatively broad experimental evaluation, covering several shape categories, initialization strategies, and comparisons with both classical and learning-based methods. The authors also consider spatial contiguity in addition to conventional geodesic error, which is a relevant aspect of correspondence quality. Although the reported contiguity improvements come with relatively high correspondence collision rates, this additional evaluation perspective is valuable.

Critical weaknesses

To the reviewer's understanding there is no guarantee for the convergence or monotonic improvement for the complete iterative procedure. In essence the different results in the proportions work for individual components. The method depends on several heuristic choices, and the role of the artificial witness-graph perturbation is not justified. The paper should explain why these operations are necessary and whether they generalize beyond the selected benchmarks. Beyond the presentation issues, I have several concerns regarding the methodological contributions and experimental validation. First, the proposed approach appears to combine established techniques for graph-based correspondence refinement, connected-component processing, and local smoothing. The fundamental novelty relative to closely related methods such as 2D-GEM and HOPE is not sufficiently clear.

A main concern is the following one. The paper suffers from significant presentation issues that make it unnecessarily difficult to read and assess. Although the overall idea of refining an initial correspondence through graph-based updates and local smoothing is relatively straightforward, the presentation is often fragmented and lacks a clear narrative. The authors frequently introduce algorithmic details, implementation choices, and empirical observations without sufficiently explaining their motivation or how they fit into the overall methodology. In particular, the distinction between the different stages of the algorithm, their respective objectives, and their actual effects on the correspondence is not always clear or communicated clearly. The theoretical results are also presented in a somewhat disconnected manner. The experimental section is similarly difficult to navigate. Rather than systematically establishing the benefits of the proposed contributions, the discussion frequently moves between individual datasets, initialization strategies, and particular numerical observations. Overall, the manuscript would benefit substantially from a more structured presentation, a clearer mathematical formulation of the underlying problem, a more concise description of the algorithm, and a systematic organization of the theoretical and experimental results. In its current form, the quality of the exposition significantly reduces the assessability.

Questions for authors
  1. The method LOGOS appears closely related to existing graph-based refinement methods, particularly 2D-GEM and HOPE. Could the authors clarify the fundamental novelty beyond combining residual gating, greedy matching, and cluster-constrained smoothing, and identify which components drive the improvements?

  2. LOGOS reports improved spatial contiguity but substantial correspondence collision rates (20.1% on SCAPE r and 18.5% on TOPKIDS). Could the authors clarify whether the improved contiguity comes at the expense of geometric correspondence quality?

  3. The propositions concern individual algorithmic steps, sometimes under assumptions differing from the implementation. Could the authors clarify which theoretical guarantees apply to the complete LOGOS algorithm and how they support its effectiveness?

Reviewer_SPkSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents LOGOS, a training-free procedure for refining hard non-rigid shape correspondences. Starting from a pointwise map, it constructs perturbed auxiliary mesh adjacencies ("witness graphs"), uses an adjacency residual to select movable vertices, and reassigns them through greedy matching on reindexed k-hop neighborhood-overlap scores. It then cleans pushed geodesic-Voronoi cluster labels by retaining their largest components and performs a within-cluster averaging and snapping step. The pipeline is evaluated on six mesh-correspondence splits. From a common SHOT initialization, it lowers normalized geodesic error on every split and obtains the best stored result on TOSCA_r. The paper also evaluates learned initializers, label contiguity and collisions, ablations, parameter sensitivity, and several theoretical properties of idealized suboperations.

Critical strengths

The reported refinement from a weak common initialization is empirically strong across six datasets, generally reducing the SHOT error by a large margin. Ablations and parameter sweeps are also reported, and I appreciated the lack of overlclaims (e.g. the paper explicitly mentions the final result not being necessarily bijective).

Critical weaknesses

In my view, the conceptual contribution is too incremental and insufficiently significant for ICLR.

The pipeline combines familiar ingredients: local graph-consistency residuals, k-hop neighborhood overlap, greedy matching, connected-component cleanup, geodesic Voronoi labels, neighborhood averaging, and nearest-vertex snapping. The paper does not identify a new learning principle, optimization formulation, representation, or general mechanism arising from this combination. Relative to local structural methods such as 2D-GEM and HOPE, the main contribution appears to be a carefully engineered sequence of heuristics and thresholds rather than a clear conceptual advance. Such an engineering component could be valuable inside a broader geometry-processing system or as a focused workshop contribution, but I do not see sufficient novelty or general interest for the main track of ICLR.

The significance is further limited by the problem formulation and evaluation regime. The experiments concern full pointwise correspondence on established small mesh benchmarks, while more difficult and currently consequential aspects of shape correspondence (including substantial partiality, topology or genus changes, severe cross-category variation, and robust correspondence under realistic geometric corruption) are either excluded or only weakly represented. The paper itself states that large partiality and genus changes are outside scope. As a result, the work does not establish why this restricted refinement problem should be considered a timely ICLR contribution in 2027 rather than an incremental geometry-processing utility.

One important aspect I noticed is that the theoretical section does not validate the implemented algorithm, and the paper/code specification contains several inconsistencies. I took my time going through the code and theorems, and whereas the propositions are mostly correct, they concern elementary or idealized facts: row equality makes a residual vanish; nearest-anchor assignment has a pointwise geodesic property but does not prove label contiguity; greedy matching is a standard 1/2-approximation for score weight; and symmetric averaging lowers a Dirichlet energy. The reported method, however, uses asymmetric directed weights followed by discrete snapping, so the Dirichlet result explicitly does not apply. Cleanup uses approximate heat distances and does not modify the correspondence, and its proposition explicitly does not establish a connected output label. The matching result controls only the score of the selected matching within the movable set B, not geodesic accuracy, residual decrease, global injectivity, or the collision-containing completed map. Collectively, the results provide no recovery, convergence, monotonicity, stability, or final-contiguity guarantee for LOGOS.

There are also concrete specification mismatches. Appendix B.7 says zero-score edges are included in the greedy scan, whereas both the Python and CUDA implementations filter them out. A positive-score column can consequently remain unmatched after its positive rows have been consumed, after which the implementation's argmax maps the all-zero assignment column to source index 0. The implementation uses a strict greater-than test where Algorithm 1 says "at or above" the polishing quantile, implements the residual denominator with a clamp rather than the stated addition of delta, and uses max(1,T//2) polishing steps rather than ceil(T/2) for nondefault odd T. The hop-shell definition also needs an explicit zero-diagonal convention to exclude two-step return walks. Most of these discrepancies are individually repairable and do not necessarily invalidate the stored results at T=32, but together they weaken confidence that Algorithm 1 is a precise, self-contained description of the released method.

Questions for authors
  1. What do the authors regard as the central new idea, beyond combining local k-hop agreement, greedy matching, component relabeling, and within-cluster averaging? In particular, what capability or principle does LOGOS introduce that is absent from 2D-GEM, HOPE, and standard graph/component post-processing?

  2. Can the authors reconcile Algorithm 1 and Appendix B.7 with the released implementation?

  3. Can the authors provide evidence that the method remains useful on a contemporary challenging setting not structurally close to the full-correspondence benchmarks used here. For example, substantial partial overlap, topology change, or a controlled cross-category protocol. And explain why the randomized witness perturbation is preferable to directly using the clean graph?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.