acceptodds
Withdrawn from ICLR 2027

NHO: A Neural Hamiltonian Operator for Anchor-based Region Localization and Dense Correspondance

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Non-rigid partial-to-full shape correspondence from sparse anchors requires identifying the corresponding region on the full surface and recovering dense correspondences between the partial shape and that region. We present NHO, which combines sparse anchors with the intrinsic geometry of the partial shape to learn a neural Hamiltonian operator whose localized eigenspace encodes both the region support and intrinsic coordinates for dense correspondence. NHO parameterizes the Hamiltonian potential as an intrinsic neural field and optimizes it using anchor evidence together with spectral and geometric constraints. To resolve the spatial ambiguity left by sparse anchors, we introduce reciprocal refinement between operator estimation and correspondence recovery. At each round, the current eigenspace provides spectral coordinates and restricts matching to its induced support, while geometrically reliable correspondences provide additional evidence for updating the potential. After refinement, aggregated eigenfunction energy yields the final localization, and the recovered map initializes dense correspondence refinement. Experiments demonstrate competitive accuracy on both tasks and robustness to uniform scaling and rotation.

Primary area:
optimization
Keywords:
Non-rigid partial shape correspondence, Hamiltonian operators, 3D shape analysis.

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_oTN3OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This submission proposes a framework of joint region localization and dense correspondence recovery for partial-to-complete non-rigid shapes. Its key design is an intrinsic neural field that parameterizes a Hamiltonian potential, coupled with reciprocal refinement in which the localized operator guides correspondence estimation and reliable matches update the operator. Through evaluations on regarding benchmarks and ablations of the operator and refinement components, the paper demonstrates competitive localization and correspondence accuracy.

Critical strengths
  • The submission is overall well-written, I liked the visualization.

  • The methodology design is natural and indeed performs well under certain conditions.

Critical weaknesses
  • My biggest concern lies on the setting -- Selecting 25 anchor points from farthest-point sampling is indeed seems to be a very strong assumption. More concretely, to the best of my knowledge, this size of landmarks is often used in works on precise maps, e.g., [a, b], where the matching accuracy is more demanding. Furthermore, when considering the partial shape matching, the FPS scheme would benefit more as it can reach the boundary points on the partial shape, while the corresponding anchors on the full shapes might lie on smooth region where feature is weak (thus they are hard to identify by automatic methods).

Considering the above, I would expect something more than superior performance on regular benchmarks.

  • From the technical novelty, my feeling is that most of the tools have been explored in the existing methods. The strong assumption also makes it difficult to isolate the technical contribution.

[a] Lifted Bijections for Low Distortion Surface Mappings [b] Surface Maps via Adaptive Triangulations

Questions for authors
  • A comprehensive study on the relationship between supervision strength and the final performance could be useful.
Reviewer_ddPbOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

NHO jointly addresses region localization and dense partial-to-full shape correspondence from sparse anchors. An intrinsic neural field represents a Hamiltonian potential, optimized per shape pair using anchor, area, connectivity, and spectral losses. Its eigenfunctions provide both a region estimate and matching coordinates. Correspondences selected by local mapping distortion update the potential, and NAM-based Neural ZoomOut refines the final map. Experiments cover four benchmarks and four scale/rotation settings, with additional anchor and mesh-simplification analyses. The default configuration uses 25 annotated anchors and is stable across the tested transformations. Its correspondence error averaged over the four settings ranks first on three datasets and second on PFAUST-H.

Critical strengths

A coherent representation for two coupled tasks The learned Hamiltonian eigenspace supplies both spatial support and matching coordinates, giving a concrete mechanism for coupling localization with correspondence. The use of intrinsic geometry is well motivated. Table 3 supports the value of the localized representation within this pipeline: replacing the Hamiltonian with the full-shape LBO reduces IoU from 75.15 to 55.48 and increases final correspondence error from 6.53 to 9.84 on normalized, unrotated CUTS’24. This is evidence that adapting the representation to the partial support is useful. Consistent results under the tested transformations With annotated anchors, CUTS’24 localization IoU stays between 75.06 and 75.24 across the four input settings (Table 1). On PFARM, Table 2 reports mean correspondence error of 10.11, compared with 25.62 for EchoMatch, and NHO leads in each individual setting. The additional localization results and the anchor/discretization analyses give a broader picture of the method’s behavior. These results support its usefulness when reliable anchors are available.

Critical weaknesses

The comparisons do not isolate the benefit beyond reliable anchors The default results use 25 correspondences obtained from annotations, whereas the paper does not establish that competing methods receive equivalent test-time information. This is a legitimate anchor-assisted setting, but a comparison under the same anchor budget is needed to determine how much of the advantage comes from NHO’s design. In particular, sparse correct correspondences can remove substantial spatial and symmetry ambiguity before the proposed machinery is applied. The automatic-anchor variant does not reproduce the default advantage. On PFARM, Table 2 reports average errors of 10.11 for Ours, 41.73 for Ours (DPFM), and 25.62 for EchoMatch. On normalized, unrotated CUTS’24, Ours (DPFM) reaches 3.29, compared with 2.38 for DPFM and 2.17 for DPFM+ZoomOut. These results do not invalidate the stated anchor-assisted task, and they do not by themselves quantify the causal contribution of anchor quality. They do limit how far the default results support practical automatic matching. Appendix C also uses separate automatically selected anchor sets for operator optimization and alignment, without reporting their sizes and accuracy. The average-over-settings ranking further needs to be interpreted as a robustness result. On PFAUST-M, 12.01 versus 12.14 for EchoMatch is a small difference without uncertainty estimates. On PFAUST-H, NHO has higher average error than EchoMatch, 27.49 versus 19.72. These limitations make a comparison under the same supervision and appropriate baseline preprocessing important to my assessment. The specific neural and reciprocal contributions are insufficiently isolated The proposed combination builds closely on existing components: intrinsic neural partiality fields (Bensaïd et al., 2023), Hamiltonian spectral localization (Rampini et al., 2019), Hamiltonian-based partial correspondence and upsampling (Wu et al., 2020), and LMD-based iterative anchor selection and spectral alignment (Xiang et al., 2021). The manuscript acknowledges this literature. The remaining contribution is the particular neural potential and its reciprocal update, so direct controls for those choices are essential. The Hamiltonian-removal ablation establishes the benefit of a localized operator in this implementation, but does not isolate neural parameterization from direct vertex-wise potential optimization. Table 3 shows a positive but modest improvement from reciprocal refinement: IoU rises from 74.37 to 75.15 and final error falls from 6.97 to 6.53. NAM also changes the full method’s error from 8.37 to 6.53. These differences are not an additive decomposition of component contributions. Comparisons using the same anchors, post-processing, and optimization budget are needed to determine the incremental value of NHO. The reported 143.15 seconds per pair on an RTX 4090, excluding cached spectral preprocessing (Appendix G), makes the absence of an equal-budget comparison consequential.

Questions for authors

Questions For Authors

  1. Clarification of supervision and comparison conditions. Which baselines receive the same 25 annotated correspondences? Please specify the initialization, preprocessing, and post-processing used for each relevant baseline. For Ours (DPFM), what are the sizes, spatial coverage, and correctness of the two initial anchor sets across datasets and input settings? A controlled comparison using identical anchors would directly address my first concern.
  2. Evidence for the neural potential. With the same losses, anchors, spectral dimensions, and NAM refinement, how does NHO compare with direct vertex-wise potential optimization and with a suitable anchor-assisted Hamiltonian baseline? The Piecewise Smooth comparison would also be more informative with the same anchors. Such controls would distinguish the contribution of neural parameterization from that of localization itself.
  3. Evidence for reciprocal updates. Does the benefit of reciprocal refinement persist when the alternative spends the same total number of operator-optimization steps on the original anchors and uses the same final refinement? Please report variation across initializations for the Table 3 comparison and the small PFAUST-M advantage. Also clarify how region extraction is defined in the LBO replacement ablation, since Eq. 18 depends on the first eigenvalue, which is zero for the usual connected closed-surface LBO.
  4. Clarification of the discretization experiment. Why does Table 8 report IoU of 80.44–82.45 after simplification, above the 75.15 full-resolution result in Tables 3 and 6? Please provide the 100% resolution control under the identical evaluation protocol and describe how anchors, region labels, and area weights are transferred. This would clarify whether the gain reflects simplification or a change in evaluation. These requests concern evidence needed to assess the current contribution. Testing additional object classes or developing new guarantees would be useful future work, but is not a prerequisite for addressing the two main concerns above.
Reviewer_Qu2UOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes NHO for anchor-guided partial-to-full shape localization and dense correspondence. An intrinsic neural field parameterizes a Hamiltonian potential, optimized using anchor, area, topology, and spectral constraints. The resulting localized eigenspace supports region estimation and correspondence recovery, while reliable correspondences feed back into operator optimization. The final map is refined using NAM-based Neural ZoomOut.

Critical strengths

The formulation is well motivated: region localization and correspondence recovery are coupled through a shared localized spectral representation. The experiments demonstrate consistent performance across the tested rotations and scale settings. The additional analyses of anchor budgets, sampling strategies, and mesh discretization are useful.

Critical weaknesses

Limited robustness with automatically generated anchors. The primary strength of the paper relies on reliable anchor points. Although the version using automatically generated anchor points improves upon the original DPFM, a significant gap remains compared to the default version that utilizes ground-truth corresponding anchor points. On CUTS’24, replacing ground-truth anchors with automatic anchors increases average correspondence error from 6.83 to 27.86. Under normalized inputs, rotation increases the automatic variant’s error from 3.29 to 26.45. These results limit the practical scope of the reported robustness. An analysis of anchor quality and sensitivity to controlled anchor corruption would strengthen the evaluation.

Questions for authors
  1. Further evaluation with automatic anchors. Could the authors evaluate robustness to anchor errors and report the accuracy, number, and spatial coverage of automatically generated anchors across transformations? Controlled anchor-corruption experiments would help distinguish initialization failures from sensitivity in the subsequent refinement.

  2. Related-work coverage. Please consider discussing recent geodesic-supervised representations for shape correspondence, including From Extrinsic to Intrinsic: Geodesic-Guided Representation Learning for 3D Geometric Data (PRISM, ICML 2026) and SGSoft: Learning Fused Semantic-Geometric Features for 3D Shape Correspondence via Template-Guided Soft Signals (CVPR 2026). Clarifying how NHO’s pair-specific operator optimization differs from and complements these learned intrinsic representations would better position its contribution.

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

Related papers

Discussion (0)

Sign in to comment.