acceptodds
Withdrawn from ICLR 2027

EvoReg: Versatile and Robust Point Cloud Registration via Multi-Stage Alignment

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Point cloud registration methods remain specialized along two axes: most target rigid or non-rigid alignment under either supervised or self-supervised training. To address this fragmentation, we propose EvoReg, a unified multi-stage coarse-to-fine framework that covers all four (rigid/non-rigid) (supervised/self-supervised) settings via a single architecture through staged decoupling. Gradient-free CMA-ES pre-alignment over escapes local minima and provides a pose-agnostic initialization, iterative Sinkhorn soft correspondences with confidence-weighted Kabsch refinement tighten the rigid pose, a residual MLP head corrects remaining rigid error, and a conditional VAE predicts a residual non-rigid deformation field on the rigidly-aligned source; each stage inherits and solves a progressively tighter initialization. At inference, four optional training-free modules — concentrated gradient-free search, point-space diffusion denoising, and global and per-point Sinkhorn test-time optimization — combine selectively to trade compute for accuracy without retraining. Trained exclusively on ModelNet40 and evaluated under controlled and real-world protocols spanning ModelNet40, ShapeNet-13, 3DMatch, and FAUST, EvoReg achieves competitive correspondence accuracy and the best distributional alignment among tested baselines — significantly leading them on every Chamfer Distance and Earth Mover's Distance evaluation cell and realizing perfect or near-perfect Chamfer Distance Recall — across all four aforementioned settings. These results show that staged decoupling is an effective architectural principle for unifying rigid and non-rigid point-cloud registration under both pose-supervised and geometric self-supervised training, and position EvoReg as a dynamic and robust method for controlled and real-world registration. The code for this project can be found at: https://anonymous.4open.science/r/EvoReg-0507.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Geometric deep learning, Point cloud registration, Self-supervised learning, Non-rigid registration

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

EvoReg combines CMA-ES initialization, iterative Sinkhorn–Kabsch alignment, residual pose correction, and VAE-based deformation, with optional diffusion and test-time optimization. The authors evaluate rigid/non-rigid registration under supervised/self-supervised settings across four datasets.

Critical strengths

The paper addresses a relevant problem and provides broad evaluations, ablations, and runtime comparisons.

Critical weaknesses
  1. Limited methodological novelty.

EvoReg largely assembles established techniques rather than proposing a new registration mechanism. Its iterative Sinkhorn–SVD stage closely resembles RPM-Net [1], while rigid/deformable point cloud matching and unified registration have already been explored [2, 3]. The claimed staged decoupling lacks a clearly demonstrated contribution beyond combining existing modules. Also, the four settings use separately trained weights, not a single model that handles all four.

  1. Rigid-registration evaluation is potentially misleading.

Rigid EvoReg-Full uses per-point test-time optimization, which can reduce Chamfer Distance (CD) without preserving a single rigid transformation or improving pose accuracy. On native 3DMatch (Table E1), EvoReg-Sup-Full achieves 99.4% CD-Recall but has a 36.94° rotation error, compared with 5.72° for Predator and 6.24° for RoITr. Moreover, CD-Recall uses a custom CD threshold rather than standard pose-based registration success. The reported surface-alignment gains therefore do not establish superior rigid registration.

Questions for authors
  1. What is the specific algorithmic novelty beyond combining existing modules? Can the authors compare against a straightforward combination baseline under matched test-time compute?

  2. Can rigid results be evaluated using only one predicted SE(3) transformation, without per-point refinement, and include standard registration recall and pose errors?

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

In this paper, the authors propose a unified, multi-stage framework for both rigid and non-rigid point cloud registration. The framework integrates CMA-ES initialization, Sinkhorn–Kabsch refinement, residual correction, and deformation modeling via a Conditional Variational Autoencoder, with an option for test-time refinement.

Critical strengths

The evaluation covers multiple datasets and includes comprehensive ablation studies. The authors explicitly highlight the trade-off between distributional alignment and pose accuracy.

Critical weaknesses
  1. Limited methodological novelty. The framework primarily integrates established optimization and registration techniques. The paper does not convincingly establish a substantial methodological contribution beyond their combination.
  2. Insufficient comparisons with recent state-of-the-art methods. The evaluation omits several recent methods with published results on 3DMatch, including PARE-Net[1], CAST[2], BUFFER-X[3], and MHopReg [4]. These omissions are particularly concerning given the paper's broad claims of robustness and competitiveness. Without comparisons under consistent evaluation protocols, its performance relative to modern registration methods remains unestablished.
  3. Inconsistent hyperparameter selection. Appendix D.3.1 identifies a Chamfer-loss weight of 1.5 as the best configuration, yet Tables D13–D14 report severe performance degradation at this value. This discrepancy undermines confidence in the reported hyperparameter analysis and requires clarification.
  4. Insufficient evidence for outdoor generalization. Despite claiming versatility and robustness, the paper provides no outdoor LiDAR registration experiments. Its generalization claims are therefore not sufficiently supported across diverse real-world environments.

[1] Yao et al. "PARE-Net: Position-Aware Rotation-Equivariant Networks for Robust Point Cloud Registration." ECCV, 2024.
[2] Huang et al. "A Consistency-Aware Spot-Guided Transformer for Versatile and Hierarchical Point Cloud Registration." NeurIPS, 2024.
[3] Seo et al. "BUFFER-X: Towards Zero-Shot Point Cloud Registration in Diverse Scenes." ICCV, 2025.
[4] Wu et al. "MHopReg: Efficient Hierarchical Multi-Hop Graph Search for Point Cloud Registration." CVPR, 2026.

Questions for authors
  1. How can the rigid registration performance be fairly assessed when per-point test-time optimization can modify the point cloud non-rigidly?
  2. Why is the Chamfer-loss weight of 1.5 described as optimal despite the poor results reported in Tables D13–D14?
Reviewer_9bRyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a unified multi-stage coarse-to-fine framework that covers all four settings (rigid / non-rigid x supervised / self-supervised) for point cloud registration. Experiments have shown that the proposed technique is promising.

Critical strengths

There are two axes in point cloud. The technique proposed in this paper covers all four settings (rigid / non-rigid x supervised / self-supervised). Experiments have shown that the proposed technique have good performance in all four settings (rank #1 or #2 in Table 1, 2, 3, and 4).

Critical weaknesses

Nil

Questions for authors

Nil

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

Related papers

Discussion (0)

Sign in to comment.