acceptodds
Under review as a conference paper at ICLR 2027

TOPOAUDIT3D: GEOMETRY IS NOT TOPOLOGY IN IMPLICIT SURFACE RECONSTRUCTION

Abstract

Implicit surface reconstruction is usually ranked by overlap or surface distance, although a geometrically accurate level set can contain the wrong number of components, tunnels, or cavities. We present TopoAudit3D, a source-audited framework with two explicitly separated components: a model-agnostic Betti-stable output audit and TopoAudit3D-SDF, a controlled localized-residual SDF reconstruction probe. The audit maps heterogeneous outputs to one physical coordinate system, evaluates topology across reference-stable offsets, and binds aggregates to case identities, evaluator replay, and provenance. Cardiac surfaces are used only as a demanding topology stress test for closed implicit surfaces. The frozen archive contains 2,793 instances, five reconstruction tracks, and 13,965 method–instance records. We reserve the 1,920-instance, repair-free M&Ms track as primary evidence. NKSR obtains lower ASSD than TopoAudit3D-SDF (0.4479 vs. 0.5200 mm) but only 7.76% topology success versus 68.70%. Among 1,203 cases where TopoAudit3D-SDF passes the complete stable-offset criterion and NKSR fails, NKSR still has lower ASSD in 784. A controlled 1,080-run study shows why evaluation and causal attribution must remain separate: residual gating and hard-example sampling help within TopoAudit3D-SDF, yet a capacity-matched flat MLP is stronger on every reported endpoint of the 60-case subset, while removing gradient detachment slightly improves geometry. We therefore make neither an identical-input claim nor an architecture-superiority claim. Instead, we establish that geometric leadership does not certify topological reliability and derive sufficient conditions under which a bounded residual deformation preserves the isotopy class of a zero level set.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
implicit surface reconstruction, topology, Betti numbers, signed distance functions, 3D reconstruction, topological evaluation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_3gUxOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents TopoAudit3D, a framework for evaluating the topology of reconstructed 3D objects. The work is motivated by the observation that 3D reconstructions can achieve low geometric errors in terms of Chamfer Distance (CD) and Hausdorff Distance (HD) while still exhibiting incorrect topology. The authors propose an evaluation method that converts different 3D representations into a common signed distance field (SDF) and compares their Betti numbers with those of the ground truth at multiple offsets. They also introduce a neural SDF model that refines coarse predictions through local corrections. Experiments on three cardiac datasets highlight discrepancies between geometric and topological evaluations, while ablation studies assess the proposed reconstruction model.

Critical strengths
  1. The proposed framework provides a systematic approach to evaluating the topology of 3D objects reconstructed using different methods, complementing conventional geometric metrics.
  2. The extensive experimental evaluation highlights discrepancies between geometric and topological measurements. The findings are supported by comparisons across multiple reconstruction methods and datasets, together with instance-level analysis. The controlled ablation studies provide additional insight into the proposed reconstruction model, while the authors acknowledge differences in input supervision.
Critical weaknesses
  1. The experimental evaluation is limited to cardiac datasets, leaving it unclear how well the framework generalizes to more complex, high-genus 3D shapes. In addition, the paper does not sufficiently investigate whether converting different 3D representations into a common SDF introduces topological artifacts, which could affect the reliability of the evaluation.
  2. The proposed reconstruction model uses substantially more supervision than the competing methods, limiting conclusions about their relative effectiveness. Moreover, the capacity-matched MLP achieves better mean results in the controlled ablation studies, raising questions about the additional value of the proposed architecture.
Questions for authors

Q1. How reliably can the proposed framework detect topological errors involving multiple handles, small holes, and narrow tunnels in high-genus shapes, such as those in Thingi10K, GSO, or Objaverse?

Q2. How sensitive are the reported topology measurements to the resolution and numerical choices used during SDF conversion, particularly for small holes and narrow tunnels?

Q3. How does the choice of SDF offsets affect the reported topology metrics, particularly when the number of reference-stable offsets varies across instances?

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

The paper proposes a framework for evaluating the quality of reconstruction methods that disentangles topological from surface-based accuracy. For measuring topological accuracy, the proposed framework converts each method's output to a common SDF and evaluates Betti numbers/errors across a predefined set of thresholds. The paper proposes a reconstruction method with which they probe their claim of a mismatch between surface distance and topological accuracy.

Critical strengths

The problem itself is relevant. The results seem plausible, and the claims are not overstated.

Critical weaknesses

The paper should be rejected because (1) it is an intermediate version of some unfinished and incomplete project, (2) its writing is flawed, and (3) the contribution itself seems minor or even wrong. I will go into more detail below.

First, I suspect the paper is an intermediate version of a project to develop a topologically accurate reconstruction method. There are several indications. The authors consistently refer to the M&M version as "repair-free" without clearly explaining what that means. In line 267, they speak about "The historical ACDC/MM-WHS correction stage", which indicates some kind of postprocessing from a previous version that led to test-set leakage. In Table 1, they mention "GT-assisted successes." The paper is clearly not ready for review yet.

Second, the paper is very hard to follow, and it exhibits typical LLM-generated flaws. Concepts are being named without introducing them, adjectives are being used in uncommon contexts and without clearly defining what they mean, and lots of unnecessary information is included.

Third, the core contribution is an evaluation framework intended to disentangle surface- and topology-based accuracy. The observation that low surface distance does not imply topological accuracy is not new. In my understanding, the proposed framework has two characteristics that go beyond that observation. First, they transfer the output to a common space, and second, they measure topology not only at the zero-level set but also at other predefined distant thresholds. Both seem potentially problematic to me. The transfer to the common SDF can hide or introduce topological artifacts, making the evaluator itself biased. This bias is not measured or systematically described in the paper. Furthermore, measuring the topology at different distance thresholds conflates surface-based distance with topology, which, ironically, is the problem the paper set out to solve.

Questions for authors

Given the state of the paper, I don't have specific questions that could potentially change my rating.

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

This paper presents TOPOAUDIT3D, a framework with two components: Betti-stable output audit and TopoAudit3D-SDF, a controlled localized-residual SDF model. The first part is intended to evaluate topological errors in the reconstruction rather than just surface accuracy. The TopoAudit3D-SDF model aims to link bounded residuals to topology through a newly provided theorem. Experiments on cardiac data show that common surface distance metrics are not correlated with topological errors and that providing GT SDF improves these topological errors compared to using only a sparse set of surface points.

Critical strengths
  • Introducing topological metrics is important for better inspection of surface quality.

  • Theoretically bounding topological change based on the residual bounds is interesting and important.

These are consequential to my overall assessment because they address real gaps in how surface reconstruction is evaluated.

Critical weaknesses
  • Evaluating only on cardiac cases is not broad enough. As topological accuracy is important in general, I think more object classes should be explored.

  • The contribution of TopoAudit3D-SDF is not clear, and its superiority in topological accuracy probably comes from training with GT SDF rather than any benefit that follows from the theorem. As shown in the ablation study, a flat MLP performs better when using the same inputs.

  • The paper is really hard to read. The flow is very non-standard, and although acknowledged, using mixed names in the figures (“HSDF”) is not acceptable in my opinion.

These points are consequential to my overall assessment because they leave both the scope and the method’s actual contribution unclear.

Questions for authors

Please address the weaknesses above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.