acceptodds
Withdrawn from ICLR 2027

Density Tells Where Matching Is Ambiguous: A Linear-Time Mamba-Transformer for Unordered 3D Points

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Point cloud registration is a fundamental prerequisite in 3D vision. While existing Transformer-based methods excel at feature matching, their quadratic computational complexity limits scalability. Conversely, linear-time Mamba models inherently struggle to capture non-sequential 3D spatial symmetries and global context. Furthermore, current approaches largely overlook local point density, an intrinsic geometric prior crucial for resolving matching ambiguities in sparse or repetitive regions. To bridge these gaps, we propose the Density-Aware Mamba-Transformer (DAMT). DAMT couples a lightweight attention mechanism, which provides permutation-invariant global initialization via a novel Global Norm Pooling (GNP) module, with a bidirectional Mamba block for linear-scaling deep feature aggregation. To fully exploit density cues without overfitting to sensor noise, we explicitly introduce a Density-Guided Attention Weighting (DGAW) mechanism and dedicated loss constraints, enforcing strict neighborhood density consistency. Extensive experiments across synthetic CAD (ModelNet/ModelLoNet), indoor RGB-D (3DMatch/3DLoMatch), and outdoor LiDAR (KITTI) benchmarks demonstrate that DAMT achieves a highly competitive performance, offering a superior trade-off between registration accuracy and computational efficiency, paving the way for scalable and robust 3D registration.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Point Cloud Alignment

Reviews

Among:
Shape:
1234this paper, Bottom 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.25 over 4 reviews, above 49% of the others.
Reviewer_t7DWOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Density-Aware Mamba-Transformer (DAMT), a hybrid Transformer-Mamba architecture for coarse-to-fine point cloud registration, aiming to address two limitations of existing methods:

  1. Transformer-based coarse matching suffers from quadratic computational complexity, while linear-complexity Mamba models struggle to capture the non-sequential geometric structures of unordered 3D point clouds.
  2. Local point density, as an intrinsic geometric cue, remains underexplored in existing registration frameworks despite its potential to alleviate matching ambiguities in sparse and repetitive regions.

To address these challenges, the authors introduce two main designs. First, DAMT employs a lightweight self-attention stage to establish an initial permutation-invariant global receptive field, together with a Global Norm Pooling (GNP) module inspired by Mamba3D. A bidirectional Mamba module is then used for efficient deep feature aggregation, reducing the reliance on computationally expensive attention layers.

Second, the authors propose a Density-Guided Attention Weighting (DGAW) mechanism that incorporates local point density as both a structural embedding and an auxiliary loss constraint, aiming to learn more discriminative features for point cloud matching.

Critical strengths

1. Favorable trade-off between registration accuracy and computational efficiency.

DAMT demonstrates strong registration performance across multiple benchmarks, including ModelNet/ModelLoNet, 3DMatch/3DLoMatch, and KITTI. Compared with the evaluated baselines, the proposed approach achieves competitive or superior registration accuracy while maintaining relatively low inference latency. These results suggest that the hybrid Transformer-Mamba architecture offers a promising trade-off between matching accuracy and computational efficiency.

2. A potentially useful geometric prior for feature matching.

Explicitly incorporating local point density into geometric feature learning is an interesting direction. The proposed DGAW mechanism complements conventional distance and angular embeddings with density-related information. The reported ablation results indicate that density-aware modeling contributes to the overall registration performance, particularly in challenging matching scenarios.

Critical weaknesses

1. Insufficient experimental evidence supporting the theoretical rationale for point density embedding.

In Appendix F.1, the authors provide two theoretical motivations for incorporating local point density: Breaking Geometric Ambiguity in Repetitive Structures and Implicit Delimitation of Structural Boundaries. While these arguments are intuitively appealing, their effectiveness is not sufficiently demonstrated by the current experiments.

  • Breaking Geometric Ambiguity in Repetitive Structures: The authors argue that local point density can distinguish geometrically similar regions. However, consider two repetitive planar patches in 3DMatch with similar local sampling densities. According to the proposed density formulation, these patches would receive nearly identical density values, leaving their geometric ambiguity unresolved. Conversely, two geometrically corresponding patches may have substantially different observed densities due to occlusion, missing regions, or incomplete sampling. In this case, density inconsistency could potentially suppress correct correspondences.

  • Implicit Delimitation of Structural Boundaries: I generally agree that local density variations may provide useful cues for identifying structural boundaries.

However, neither of these claims has been directly validated through experiments. I suggest that the authors conduct additional controlled experiments:

Breaking Geometric Ambiguity: Investigate whether the proposed method can produce a greater number of correct correspondences and a higher proportion of correct matches for local patch pairs with high geometric ambiguity.

Implicit Delimitation of Structural Boundaries: Investigate whether the proposed method can produce a greater number of correct correspondences and a higher proportion of correct matches for local patch pairs with high curvature.

Although the existing ablation studies demonstrate the overall effectiveness of density-related components, they do not directly validate these two specific theoretical claims. More targeted experiments would substantially strengthen the motivation and reliability of the proposed density prior.

2. Limited validation of the effectiveness of point density normalization.

The authors introduce range-normalized point density to compensate for variations in sampling density caused by sensor distance. However, the experimental evidence for its effectiveness under varying sensing distances remains limited.

The KITTI benchmark already exhibits relatively high registration recall for several existing methods without explicitly incorporating point density priors. Consequently, performance improvements on the standard KITTI setting alone may not be sufficient to demonstrate the advantages of the proposed normalization strategy, particularly under more challenging long-range conditions.

I suggest conducting additional experiments under larger spatial separations, such as KITTI-30m and KITTI-50m [1], to evaluate the effectiveness and robustness of density normalization. Comparisons among models without density guidance, with unnormalized density, and with normalized density would provide stronger evidence for the necessity of the proposed normalization strategy.

Summary

The proposed DAMT achieves advantages in both registration performance and computational efficiency compared with the evaluated methods. However, its methodological novelty is somewhat limited, and the claimed effectiveness of the point density prior lacks direct experimental validation. In addition, there are inconsistencies between the textual descriptions and figures. For rate improving, I believe the paper would be substantially strengthened by solving the weakness .

[1] Xiong, K., Xiang, H., Xu, Q., Wen, C., Shen, S., Li, J., & Wang, C. (2024). Mining and transferring feature-geometry coherence for unsupervised point cloud registration. Advances in Neural Information Processing Systems, 37, 35468-35491.

Questions for authors

1. Inconsistency between the description and illustration of the Global Norm Pooling module.

In Section 3.2, the textual description and mathematical formulation of the GNP module do not appear to be fully consistent with its illustration in Figure 2.

In particular, the G-Norm operation is described as normalizing point features relative to their assigned anchor centers, followed by a learnable affine transformation. However, the operations presented in Figure 2 do not clearly correspond to this formulation.

I suggest that the authors carefully revise the GNP illustration and its corresponding description to ensure consistency among the textual explanations, and architectural diagram.

2. Important evidence supporting the main contributions should be emphasized in the main paper.

The primary contributions of DAMT lie in its computational efficiency and the incorporation of point density as a geometric prior. However, some of the most relevant supporting analyses are either placed in the appendix or insufficiently explored.

In particular, Table 8 (Scalability Test) provides valuable evidence for the computational advantages of the hybrid architecture as point cloud size increases. I recommend moving this experiment to the main paper, since scalability is one of the central motivations of the proposed method.

Similarly, the targeted density-related analyses suggested in Weakness 1 should be included in the main paper in my opinion. These experiments would directly support the claimed geometric benefits of point density embedding and make the methodological contributions more convincing.

2 comments
Reviewer_t7DWOpenReview ↗

Final Justification: improve the rating from 2 to 3

I thank the authors for their thoughtful response. Although the density prior is applied only in self-attention, this design alone does not establish on theoretical grounds that the prior can help the network recover true correspondences between patches sampled at different densities in the two views. In light of the experiments reported in the rebuttal, one possible explanation is that the density prior guides the network during self-attention to learn distinct features for geometrically similar superpoint patches with different local densities, thereby reducing ambiguity among the features from the two input views during cross-attention. Overall, the authors have fully addressed my concerns, and their results show that combining the density prior with the hybrid Transformer–Mamba architecture improves the trade-off between registration accuracy and computational efficiency. I encourage the authors to consider whether this explanation could clarify the role of the density prior in the revised paper. I have raised my rating to 3.

AuthorsOpenReview ↗

Thanks for the suggestion and rating improvement

Thanks for raising the rating and the clear framing. We agree that the density prior gives similar patches with different densities distinct descriptors during self-attention, reducing cross-view ambiguity at the cross-attention stage. We'll add this explanation to the revision.

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

This paper proposes DAMT for point cloud registration. It builds on the GeoTransformer framework and adds three parts:

  • a global pooling module (GNP)
  • a bidirectional Mamba block after the attention layers
  • a density module (DGAW) that adds local point density to the attention weights, with a density loss

The method is evaluated on KITTI, ModelNet/ModelLoNet, and 3DMatch/3DLoMatch. The paper reports accuracy comparable to or better than recent methods with lower runtime. Ablations are done on ModelNet.

Critical strengths
  1. The method is evaluated on standard benchmarks across domains, and it shows competitive accuracy with lower reported runtime than recent methods.
  2. Using local density as an extra prior is a simple idea with very low cost. It can be easily added to existing registration frameworks.
Critical weaknesses
  1. Reported results are not consistent. The ModelNet RRE of the default full model is 1.127 in Tables 4, 5, 11, and 12, but 1.194 in Table 9. The text of App. D.4 also says 1.194, while Table 11, which it describes, shows 1.127. In addition, the formula in App. G.2 seems different from Eq. (7). The model without the density module has an RRE of 1.192 (Sec. 4.3). So whether the density module helps depends on which value is correct.

  2. Main claims are not well supported.

  • All ablations are done on ModelNet only (Sec. 4.3). A key part of the density design, range normalization, is mainly meaningful for LiDAR, but its effect is not tested on KITTI or 3DLoMatch. App. F.2 says this is validated by ablation studies, but I could not find such results. Also, the number of runs or standard deviations is not reported. Since the gains are small (e.g., 62.7% → 63.3% RR when adding the density module), it is hard to tell whether the improvement is meaningful.
  • The title says linear-time, but attention seems to remain in every block (Eq. (10), Fig. 2). It is not clear to me why FLOPs and parameters are lower than GeoTransformer (Table 7).
  • The paper claims generalization across modalities, but each model is trained per dataset and the density normalization is defined per data type.
  1. Limited novelty. The method mainly combines existing parts: the GeoTransformer framework, a global version of Mamba3D’s pooling, and a bidirectional SSM. The main new part is the density prior, but I think it is the least validated part, since its ablations are limited to ModelNet. The contribution seems incremental for ICLR.
Questions for authors
  1. Tables 4 and 9 seem to report the same default model, but the RRE is different. Which value is the final result?
  2. What is the effect of the density module on KITTI and 3DLoMatch? Since the gains are small, could the results be reported over multiple runs?
  3. For indoor RGB-D data, density is only normalized by a constant, so the same surface can have different densities in two fragments. Could the density module suppress correct matches in this case?
  4. The paper claims generalization across modalities, but each model is trained separately per dataset. In what sense does this show generalization?
  5. The paper does not describe the model dimensions or backbone settings. Why does DAMT have fewer FLOPs and parameters than GeoTransformer, even though it adds modules to a similar framework?
  6. How does DAMT scale compared to existing methods such as GeoTransformer, not only Model B?
3 comments
AuthorsOpenReview ↗

Follow-up on soundness and the density-artifact concern

Dear Reviewer ACuu,

Thank you again for your review. As the discussion ends on Nov 18, we would like to draw your attention to the two points most relevant to the soundness concerns underlying your rating:

  1. Numerical reconciliation: we confirmed that the discrepancies you identified (1.127 vs. 1.192, and the Table 14 "None" row) stemmed from a single outdated checkpoint. All affected numbers are now reconciled in our response, and the corrected configuration has been re-run with consistent results.

  2. Density as sensor artifact: the KITTI-30m ablation is a direct test of this concern. Unnormalized density performs WORSE than no density at long range (RTE 19.1 vs. 18.4 cm), while range-normalized density is clearly better (15.6 cm; RR 96.4 vs. 94.2). A prior that merely encoded sensor statistics could not produce this pattern — under long-range attenuation it would degrade the same way the unnormalized variant does. The result indicates the formulation captures geometric structure rather than sensor-specific biases.

May we ask whether these clarifications address the soundness concerns underlying your rating, or whether any remaining issue should be addressed before the discussion closes?

Best regards,

The Authors

Thank you for the detailed responses and the additional experiments. The clarifications have helped me better understand the motivation and design choices.

However, while the additional results shared in the rebuttal are helpful, I feel they are not yet sufficient, and the core experiments are not well organized. The ablations in the manuscript are conducted on ModelNet, where the improvement from the density module appears relatively small, while the additional results suggest larger gains on real scans, such as 3DLoMatch and long-range KITTI. Since the method is built on point density, I think its main benefit would be better demonstrated on real scans with non-uniform density than on uniformly sampled CAD data. I would recommend including these ablations in the main paper, with results over multiple seeds, and making sure that the numbers are consistent across all tables. It is also somewhat unfortunate that the paper does not include qualitative comparisons with other methods or a visualization of where the density cue actually helps resolve matching ambiguity.

Overall, I appreciate the authors' efforts. I lean toward keeping my current rating, but I am open to the discussion with the other reviewers and the AC.

AuthorsOpenReview ↗

Follow-up on ablation placement and the qualitative comparison

Thanks for the constructive feedback. We agree that real scans with non-uniform density better demonstrate the density module; the revision will move the real-scan ablations (3DLoMatch, long-range KITTI) into the main paper with multi-seed results and consistent numbers across all tables.

On the qualitative comparison, we did run PTT and END on ModelNet/ModelLoNet and CAST and GeoTransformer on 3DMatch/3DLoMatch, and inspected their correspondence visualizations against ours. Since these baselines already reach near-saturated metrics on these benchmarks, we could not find visually distinguishable differences --- their correspondence maps look essentially identical to ours. Including such figures would add length without conveying additional information, so we opted not to include them.

Reviewer_2HM1OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposed a Transformer-Mamba hybrid architecture for point cloud registration, and designed a global norm pooling module that effectively extracts pertumation-invariant global context to capture non-sequential 3D spatial symmetries. The authors also employ point cloud density to boost robustness. The provided quantitative results show that the presented DAMT outperforms other methods. The main concerns are the limited qualitative and efficiency experiments, unjustified parameter and blocks choice, missing values in some quantitative results.

Critical strengths

-This paper proposed a hybrid archiecture coupling Transformer and Mamba for point cloud registration, which achieves SOTA performance in several datasets.

-The derivation part in METHOD is solid.

Critical weaknesses

-The experiments focus on quantitative evaluations. In the qualitative experiments, the authors do not provide registration results from other compared methods, which makes it hard to evaluate the effectiveness of the proposed DAMT.

-The quantitative results of some methods are missing. For example, the inference time of several Rethinking, MART, etc in Table 1, as well as some values in Table 2 and 3.

-The authors do not provide sufficient efficiency comparison as only GeoTrans and Model B are compared.

-In Table 9, 4-layer blocks seems have better performance as four indicators are best and another one indicator (BRE) is the second best. It is inferior to 3-layer configuration only in RR, while all other indicators are better.

-The figures are not sufficiently clear as some visual elements are not adequately explianed, such as the yellow/orange and blue rectangles in Figure 2.

-The choices of certain parameters are not sufficiently justified, such as the setting of loss weight which lacks rationale or ablation experiments.

Questions for authors

-Could the author provide visual results of other compared methods under fair setting?

-Could the author explain why they choose 3-layer blocks instead of 4-layer ones.

-Could the author explain why the fields of some quantitative results are empty?

1 comment
AuthorsOpenReview ↗

Follow-up on the depth choice and Eq. (11)

Dear Reviewer 2HM1,

Thank you again for the careful review. With the discussion ending on Nov 18, we would like to highlight the two points most central to your concerns:

  1. Depth choice: Table 9 of the paper already contains the direct comparison — a 4-block variant yields RRE 1.193 (vs. 1.194), PIR 0.957 (vs. 0.955), but RR 62.6 (vs. 63.3). The 3-block configuration is therefore a deliberate accuracy-efficiency trade-off supported by the reported numbers rather than a cost shortcut.

  2. Eq. (11): we confirmed that the coefficient omega is missing from the equation as printed and acknowledged the erratum in our response; the accompanying text and the reported weight (0.5) are unaffected.

All six of your concerns are addressed in our response above using numbers already present in the paper or its supplementary material; no unverified figures were introduced. Could you let us know whether any specific point remains unresolved for you before the discussion closes? We are happy to clarify further.

Best regards,

The Authors

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

DAMT is a registration network that stacks a shallow attention stage and a bidirectional Mamba block on superpoints. Global Norm Pooling (GNP) picks 32 anchor centers by farthest point sampling, normalizes features relative to the nearest center, and pools them with a channel-wise softmax. Density enters through range-normalized neighbor counts. An MLP turns pairs of these counts into an additive bias on the self-attention logits (DGAW), and an MSE loss supervises a predicted density feature. Superpoints are serialized in Z-order before the Mamba block. Results cover KITTI, ModelNet and ModelLoNet, and 3DMatch and 3DLoMatch.

Critical strengths

•Evaluation spans outdoor LiDAR, indoor RGB-D, and synthetic CAD data, and the supplement includes code. Appendix ablations cover GNP against max and average pooling (Table 11), serialization order (Table 6), the number of GNP centers (Table 10), the density radius (Table 13), and alternative density estimators (Table 14).

•On 3DLoMatch the improvement is steady across sample counts: registration recall is 76.3, 76.0, 75.8, and 75.4 at 2500 down to 250 samples, against 74.8, 74.5, 73.9, and 73.2 for CAST, and inlier ratio at 250 samples is 60.9 against 57.7 for GeoTransformer.

•The scaling test (Table 8) shows an attention-only variant running out of memory at 100k points while DAMT uses 9.2 GB.

•Eq. 4 is readable once the extraction artifacts are ignored: the density term is an additive logit bias, in the same position as the geometric embedding.

Critical weaknesses

•The "linear-time" claim in the title is overstated. Fig. 2 shows self-attention and cross-attention repeated three times, each followed by the bidirectional SSM, so quadratic attention remains in every block. On KITTI the gain over a pure-attention model is modest (0.22 s to 0.15 s, 24.1 to 16.5 GFLOPs, Table 7), and the pure-attention model is as accurate or better on some metrics (PIR 0.957 against 0.955, Table 5). N in Table 8 is not defined, so it is unclear whether it counts points or superpoints.

•Density does not act where the motivation says it does. The text says matches across different sampling densities are penalized (Eq. 7), but Φ is added to the self-attention logits (Eq. 4) and is computed between superpoints of the same cloud. The cross-attention in Eq. 8, where matching between clouds happens, has no density term. The text calls Φ a "gate" while Eq. 4 shows an additive bias, and Appendix G.2 describes Eq. 7 as (r_i + r_j)W^De, which differs from the MLP in the main text.

•The density loss may be close to trivial. L_den (Eq. 12) regresses a predicted density feature onto D(p̂), which comes from the same neighbor counts that produce the input feature r^De (Eq. 5). The network is asked to reproduce something it already sees, and no ablation isolates this loss. In Table 4, the density modules add 0.6 points of recall and 0.065° of rotation error over GNP + Mamba, from one run on ModelNet.

•Some numbers do not line up. Table 14's "None" baseline gives RR 60.1 and RRE 1.41, while Table 4's GNP + Mamba without density gives 62.7 and 1.192. Sec. 3.5 says registration uses LGR, but Sec. 4 says 50k RANSAC iterations are run and left out of the reported time. Baselines that use LGR, such as GeoTransformer, are therefore not timed on the same footing.

•Margins on saturated benchmarks are tiny. On KITTI, recall is 100.0 for both DAMT and CAST, translation error is 2.3 against 2.5 cm, and rotation error 0.22 against 0.23° (Table 1). On ModelNet, rotation error is 1.127 against 1.138 for IAPReg, and on ModelLoNet 1.613 against 1.618 for MTAR (Table 2). On 3DMatch, recall is 95.4 against 95.2 for CAST. Without seeds, these differences cannot be separated from noise.

•The robustness argument is asserted more than tested. Appendix F says robustness is "guaranteed by design" and that heatmaps are unnecessary. No experiment uses a mis-specified density profile or a different sensor, and the radius study (Table 13) is on ModelNet, where the 1/d² profile does not apply. The claim that α is absorbed by the MLP covers a scale error, not a wrong profile shape.

Questions for authors

1.Why is Φ used only in self-attention? What happens if density is added to the cross-attention logits as well?

2.Which estimator, LGR or 50k-iteration RANSAC, produced each reported number, and what are end-to-end times including it for all methods?

3.What does N mean in Table 8, and how many superpoints enter attention on each dataset?

4.Can you ablate L_den separately, and compare injection designs (concatenation, additive bias, multiplicative gate) with the same density feature?

5.Which configuration is "None" in Table 14, and how does it relate to the rows of Table 4?

6.Can you provide mean and standard deviation over several seeds for the main tables and ablations?

7.How does the method behave when the expected density profile is wrong, or on a sensor with a different beam count?

1 comment
AuthorsOpenReview ↗

Follow-up on the timing protocol and Table 14

Dear Reviewer ocTB,

Thank you again for the detailed and technically engaged review. As the author-reviewer discussion ends on Nov 18, we would like to briefly consolidate the points that most directly address your main concerns (full details in our four comments above):

  1. Timing protocol: the reported comparison is now single and explicit — accuracy and latency are both obtained under 50k-iteration RANSAC (0.180 s vs. 0.260 s when appending LGR, +31%; 0.150 s vs. 0.192 s at matched accuracy, +22%), so all comparisons are on the same footing.

  2. Table 14: the "None" row is reconciled to 62.7% RR / 1.192° RRE under the same configuration as all four rows, and the resulting ordering — KNN 61.7 < none 62.7 < KDE 62.9 < range-normalized 63.3 — indicates that the gain comes from the range normalization rather than from adding raw density.

  3. Density injection: the cross-attention control (RRE 1.185 / RR 62.4 vs. 1.127 / 63.3) and the L_den ablation (1.192 / 1.154 / 1.127) support the intra-cloud design choice.

May we ask whether these clarifications resolve your concerns, or whether any specific point remains that we should address before the discussion closes?

Best regards,

The Authors

Comments on the paper

1 comment
Area_Chair_nNuPOpenReview ↗

Discussion reminder

Dear reviewer:

At first, thank you for being a reviewer for ICLR 2027, and thanks for your contributions so far.

We are half way through the discussion period. Please check if the authors' response have addressed your concerns, any remaining/extended concerns, and update your final rating with justification as early as possible.

Thank you for your contributions again

AC

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

Related papers

Discussion (0)

Sign in to comment.