acceptodds
Under review as a conference paper at ICLR 2027

R: Geometric Regressors Are Effective Point Cloud Registration Learners

Abstract

We propose R a point cloud registration framework that recasts registration as one-shot shared geometry regression. Instead of learning explicit correspondences or descriptors, R predicts a shared, pair-conditioned canonical space for both partial observations. Using a compact set of geometric anchors, R assigns canonical coordinates to sampled points from both views, including points outside the overlapping regions. Correspondences are then derived naturally from the predicted canonical geometry through mutual nearest-neighbor matching, and the rigid transformation is recovered in the original Euclidean space using a closed-form SVD-based Kabsch solver. To improve canonicalization under partial overlap, we introduce a cross-view geometric coupling loss that complements the coordinate regression loss with relational supervision. The resulting framework requires only pairwise pose supervision to construct canonical targets and does not rely on annotated point-wise correspondence labels. Experiments on 3DMatch and 3DLoMatch demonstrate that R achieves competitive registration accuracy with a simple RANSAC-free inference pipeline and favorable speed–accuracy trade-offs.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Point cloud registrtation, 3D reconstrction

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_F6R7OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes R²,a point cloud registration framework based on shared geometric regression.Via a single forward pass,the framework maps sampled anchor points from the two input point clouds into a shared canonical coordinate space.It then establishes point correspondences following the mutual nearest-neighbor rule and solves for the rigid transformation using SVD.For loss design,the method combines coordinate regression loss and cross-view geometric coupling loss to constrain the predicted coordinates and local geometric relations.

Critical strengths

(1)The idea of mapping two point clouds into a shared canonical space is inspiring, and it achieves a favorable trade-off between registration accuracy and computational overhead. (2)Ablation studies demonstrate that both cross-view geometric supervision and intermediate-layer supervision effectively boost registration performance, with particularly prominent gains under low-overlap scenarios.

Critical weaknesses

(1)While the ablation results validate the effectiveness of the proposed supervision designs,the intrinsic advantage brought by the shared canonical space itself remains unclear.Additional experiments are required to justify this claim. (2)Table 3 shows that replacing SVD with RANSAC provides only a small improvement in overall recall.However, the robustness of the default solver remains insufficiently characterized.Reporting correspondence inlier ratios and comparing the two solvers across overlap levels would help clarify when MNN+SVD is reliable and when robust estimation remains beneficial.

Questions for authors

(1)What is the correlation between the rigidity score adopted in TTA and registration success?Is it possible that predictions satisfying rigidity constraints individually yet yielding wrong relative alignment can still obtain low rigidity scores? (2)What criteria were used to set the triggering and stopping thresholds for TTA?Were these thresholds tuned solely on the validation set?How robust is the registration performance with respect to variations in these threshold values?

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

This paper presents a one‑shot point cloud registration framework that reformulates pairwise rigid registration as shared‑canonical‑space geometry regression, departing from conventional descriptor‑matching or iterative flow‑based paradigms.

Critical strengths

It provides a feed‑forward regression perspective for point cloud registration and adopts a simple deterministic pose solver without mandatory RANSAC to achieve a favorable speed‑accuracy trade‑off particularly for low‑overlap inputs, while its geometric coupling loss remedies the critical drawback of vanilla coordinate regression that low per‑point error cannot guarantee valid cross‑view relative geometry.

Critical weaknesses

(1)Clarification of the baseline results in Table 1. In Table 1, the RRE and RTE results for GeoTransformer, BUFFER, PSReg, and RAP are marked as n/a, while their RR values are still reported. Please clarify the source and computation of these RR values, and explain why the corresponding RRE and RTE results are unavailable. It would be helpful to indicate which results were taken from the original papers and which were reproduced by the authors, and to specify the evaluation protocol used. This information is needed to assess the completeness and comparability of the results. (2)More thorough validation of the coupling loss. The ablation results for the coupling loss require more thorough validation. Table 3 shows that removing L_coup reduces the RR from 95.28% to 86.02% on 3DMatch and from 81.63% to 52.09% on 3DLoMatch. This substantial drop suggests that the loss is critical to the method, yet the current manuscript does not provide a hyperparameter sensitivity analysis. Please evaluate the effects of λ_reg, λ_coup, the KNN neighborhood size, the coupling radius, and the tolerance τ. This would establish whether the gains are stable and clarify how strongly they depend on these settings. (3)Direct evaluation of shared-geometry regression. The direct evidence for the shared-geometry regression claim remains limited. Although the paper states that the model predicts canonical coordinates for non-overlapping regions, it does not report the prediction error relative to the GT canonical targets. The cross-view nearest-neighbor residual in Figure 8 measures only the proximity between the two predicted point sets; it does not measure canonical-coordinate prediction accuracy, a limitation that the authors also acknowledge. Please report canonical-coordinate errors on the test set to provide a more direct evaluation of the paper’s central claim.

Questions for authors

Please see the weakness part.

Reviewer_hBXFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a point cloud registration framework that recasts registration as one-shot shared geometry regression. Experiments have been conducted to evaluate the performance of proposed technique.

Critical strengths

Table 1, 2 and 3 shows the evaluation of the proposed technique. Table 1 shows that the proposed technique achieves the best recall in both 3DMatch and 3DLoMatch.

Critical weaknesses

Table 1, 2 and 3 shows the evaluation of the proposed technique.

As shown in Table 1, REGTR has lower RRE and RTE in 3DMatch and lower RTE in 3DLoMatch. On the other hand, the time required for REGTR is significantly lower than proposed technique. shows that the proposed technique achieves the best recall in both 3DMatch and 3DLoMatch.

As shown in Table 2, most methods can achieve 99.8 TR. Some methods have lower RE and TE than proposed technique.

As shown in Table 3, MNN + RANSAC has lower RRE and RTE, and higher RR. The time required are similar.

Questions for authors

Nil

1 comment
AuthorsOpenReview ↗

Warm Thank You and Quick Summary of Tables 1, 2, and 3

Dear Reviewer hBXF,

Thank you again for your helpful review! Since our earlier reply above was quite detailed, we wanted to leave a short 3-point summary here for your easy reference, showing how our unified results highlight the strengths of R²:

  1. Table 1 (R² leads across Recall, RRE, and RTE at high speed): As you noted in Strengths, R² achieves the highest recall on both 3DMatch and 3DLoMatch. Once REGTR is evaluated under the exact same error rule (mean vs. mean, or median vs. median), R² not only beats REGTR on every metric (+16.71% recall on 3DLoMatch), but actually achieves the lowest RRE and lowest (or tied-lowest) RTE across all methods in Table 1, while running in one 59 ms pass (1.7x to 20x faster than other high-recall methods at 98-1,428 ms).
  2. Table 2 (554/555 outdoor ceiling in one step, zero refinement): On outdoor KITTI (benchmark threshold: 5° / 200 cm), 99.8% recall already means succeeding on 554 out of all 555 test pairs. While baselines with slightly smaller RE/TE (e.g., 0.23° / 4.9 cm vs. our 0.24° / 7.7 cm) rely on RANSAC or multi-step pose refinement, R² reaches this ceiling in one simple SVD step with zero extra refinement.
  3. Table 3 (working without RANSAC is R²'s core strength): Both MNN+SVD and MNN+RANSAC in Table 3 are two versions of our own R². In 3D registration, conventional methods fail badly without RANSAC because of wrong matches. The fact that our default R² (MNN+SVD) gets almost the exact same high accuracy (95.28% / 81.63%) as R² (MNN+RANSAC) (95.39% / 82.17%, random sampling) proves how clean and accurate R²'s one-shot predictions already are, while users can freely choose either version in practice.

Thank you very much for your effort and sharp observations, which helped make these strengths much clearer in our revised paper. If these clarifications address your concerns, we would be sincerely grateful if you could kindly consider updating your score, and we wish you all the best!

Warm regards,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.