acceptodds
Under review as a conference paper at ICLR 2027

G2M-Reg: Graph-to-Manifold Verification for Robust Point Cloud Registration

Abstract

Robust point cloud registration remains challenging under low overlap, where sparse true matches can be overwhelmed by locally consistent outliers that form strong compatibility structures. Consequently, an incorrect pose may receive substantial correspondence support while failing to explain the underlying common surface. This reveals a fundamental limitation of correspondence-space reasoning: evidence that is effective for generating a pose is not necessarily sufficient for verifying its geometric validity. We propose G2M-Reg, a training-free Graph-to-Manifold Verification framework that separates pose proposal from pose verification. A correspondence graph first generates a set of plausible pose hypotheses. Instead of ranking them solely by correspondence-level evidence, G2M-Reg evaluates the cross-cloud surface relation induced by each candidate pose. Bidirectional partial-overlap consensus measures whether the two fragments mutually explain a plausible common region, while structural reliability evaluates whether the supporting geometry is coherent and sufficiently informative to constrain the pose. Graph, surface, and structural evidence is subsequently integrated in rank space to select a surface-valid hypothesis, followed by constrained surface-guided refinement. Experiments across indoor and outdoor benchmarks and diverse hand-crafted and learned correspondence sources demonstrate that G2M-Reg consistently improves robust pose selection, particularly in challenging low-overlap scenarios.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
point cloud registration, robust registration, pose verification, low-overlap registration, geometric matching, 3D vision

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

This work presents the G2M-Reg framework for point cloud registration,targeting low-overlap scenarios where erroneous matches can receive high graph-level support and induce misalignment.The author separates candidate pose generation from geometric validation:candidate poses are first derived from the correspondence graph,and their credibility is subsequently assessed based on bidirectional partial-overlap surface consensus,alongside topological preservation,overlap consistency,and pose observability metrics.Pose candidates are then selected by aggregating rankings across these multiple cues,prior to constrained surface refinement.

Critical strengths

(1)This work decouples candidate pose generation from surface geometric verification to address the issue where false matches can receive strong graph support under low overlap,with a well-articulated design. (2)In addition to surface alignment quality,overlap coherence and pose observability are incorporated,enabling the detection of candidates that achieve plausible local fitting but suffer from weak constraints.

Critical weaknesses

(1)The method mainly integrates geometric residuals,bidirectional consensus,structural scores and rank fusion.Its substantial innovation over existing geometric verification and candidate reranking approaches has not been adequately elaborated. (2)Comparisons against baselines such as naive bidirectional distance and truncated point-to-plane scoring on the same set of candidate poses are not sufficiently presented,making it difficult to justify the necessity of this complex verification pipeline. (3)The full configuration retained for each ablation group is ambiguously described.The three sub-components for structural reliability are not ablated individually,which limits the assessment of the independent contribution of each design. (4)Sensitivity analysis of key parameters,stability evaluation of results,and runtime breakdown for each module are missing.Therefore,the reliability of marginal performance gains and the reasonableness of the extra computational overhead cannot be fully evaluated.

Questions for authors

(1)When repetitive structures yield high surface consistency for multiple erroneous poses,how does the method distinguish true alignment from geometrically plausible false alignment?Are there inherent ambiguities that cannot be resolved? (2)The bidirectional consensus leverages the same set of candidate overlap ratios,yet the two point clouds may differ in size,density,and visible field of view. How does this design cope with discrepancies in the actual overlap ratio between the two directions? (3)The pose observability score penalizes degenerate geometries.For samples where the real overlapping region mainly consists of planes or corridors,could this downgrade the ranking of the correct pose?

Reviewer_4UekOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The authors introduce a training-free approach for point cloud matching and registration (for 3D scenes specifically - indoor and outdoor). Their core method finds empirical evidence that information relevant for generating a pose is not necessarily sufficient for verifying its geometric validity. Thus they propose to separate pose proposal from pose verification thus leaning on older more traditional approaches in improving a common problem.

Critical strengths

The work analyzes well what key issues can be observed in state-of-the-art and clearly finds a novel issue in the verification process. The work focuses on reliability which is key in many real-world applications.

Critical weaknesses
  • Would it be possible to clarify the difference between the proposed "Bidirectional Partial-Overlap Consensus" and a classical Chamfer Distance? For me this seems to be missing in the comparison, since both essentially aim to address the same problem. The other metrics all seem to novel in this application area but seem to be extension from statistical analysis.
  • In ablation study, Structural Reliability, Grouped Evidence Fusion, and Surface-Guided Refinement seem to all have the exact same impact on the performance, is this correct? If yes, how can this and what's the intuition behind this?
  • The main claim of the paper is: "...where sparse true matches can be overwhelmed by locally consistent outliers that form strong compatibility structures. Consequently, an incorrect pose may receive substantial correspondence support while failing to explain the underlying common surface." , but I lack general theoretical or empirical analysis to validate this claim conditions under which surface verification provably rejects false hypotheses. Would it be possible to extend this observation to an actual analysis (with best case some kind of error bound). If not, please explain why and end-to-end accuracy is already sufficient to explain this observation.
  • Are the current used verification metrics related to existing 3D distance / structural metrics (similar to Chamfer Distance with partial overlap). I think this would require and additional related work subsection.

I am very open to improving my initial rating, I am currently mainly confused about the actual novelty since the metrics mainly seem to be already known. I would require a clear definition what are the old metrics / what the new ones and what's the difference to improve my score.

I am not familiar with the larger literature in this are especially considering the training-free approach this seems to very required as it could potentially contain already existing idea from old-school traditional method. Due to my lack of knowledge in this I am assigning a low-confidence score for myself.

Questions for authors

How can you explain the difference between KITTI and 3DMatch/3DLoMatch performance? Rest see in weaknesses

Reviewer_UXo1OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors propose a training-free point cloud registration framework named G2M-Reg, which combines graph-based pose generation, surface-based verification, and constrained refinement. Comparative experiments demonstrate its effectiveness on indoor and outdoor benchmarks.

Critical strengths

The method is well organized. Experimental comparisons and component ablations support the method’s effectiveness.

Critical weaknesses
  1. The method appears to have limited novelty. In Section 3.2, SC²-PCR is employed to generate hypotheses, while Equation (1) combines standard metrics such as point-to-point distance, point-to-plane distance and normal vector alignment. The paper fails to provide a compelling justification for how this specific combination offers a substantive contribution beyond simple geometric verification. Furthermore, the proposed method is not compared against state-of-the-art (SOTA approaches such as SC²-PCR++ [1] and Regor [2].
  2. Several definitions lack clarity. For instance, the gig_i in Section 3.3, the Ur\mathcal{U}_r in Equation (11), and the wi(h)w_i^{(h)} in Equation (13). Additionally, Table 2 does not explicitly detail the configurations for the Graph and Manifold.
  3. The group-based fusion mechanism requires further justification. While converting scores into rankings, it does not resolve conflicts in evidence: a single hypothesis might be ranked highly in one group but poorly in another. The paper lacks an analysis of the conditions under which weighted aggregation improves selection outcomes, nor does it explore how unreliable groups might negatively impact the final results.

[1] Z. Chen, K. Sun, F. Yang, L. Guo, W. Tao. SC²-PCR++: Rethinking the Generation and Selection for Efficient and Robust Point Cloud Registration. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2023, 45(10): 12358–12376.

[2] G. Zhao, S. Ao, Y. Zhang, K. Xu, Y. Guo. Progressive Correspondence Regenerator for Robust 3D Registration. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2025: 1210–1219.

Questions for authors

While the paper presents promising results, its methodological breakthroughs and the reliability of its integration mechanisms have not yet been sufficiently established. To substantiate its core arguments, the paper requires further comparative analysis and a more explicit description of its experimental protocols.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.