acceptodds
Under review as a conference paper at ICLR 2027

LiMamba: Rethinking Point Cloud Serialization in Mamba-Based Diffusion Models for LiDAR Scene Completion

Abstract

Diffusion-based methods now dominate LiDAR point cloud completion, yet every existing point-based scene-completion denoiser relies on sparse 3D convolutions whose receptive fields are inherently local, leaving long-range structural correlations largely uncaptured. Explicit global self-attention is uneconomical at outdoor scale, while state space models such as Mamba offer linear-complexity modeling but require flattening 3D voxels into a 1D sequence. Our central observation is that, while multi-directional scanning is a minor detail for object-level point clouds, at LiDAR scene scale the serialization order becomes a central design choice: the ordering governs how far the state space can propagate structure, and any single ordering is lossy. We propose LiMamba, a hybrid MinkowskiEngine–Mamba denoiser built around Multi-Scan Mamba (MSM) blocks. Each MSM block serializes voxel features along orderings derived from spinning-LiDAR acquisition physics (a radial range-density ordering and an azimuthal scan-line ordering, alongside Z-order), processes them through parallel Mamba streams, and fuses the outputs via learned gating. A Scan Dropout strategy prevents the streams from collapsing onto one during training, so that multi-scan improves accuracy rather than merely adding computation. Experiments on SemanticKITTI show that LiMamba improves over purely convolutional diffusion baselines on Chamfer Distance, JSD, and coarse-resolution occupancy IoU, while retaining linear-complexity, memory-feasible global modeling at a scale where explicit self-attention is not feasible, and ablations confirm the complementary value of multi-scan fusion over any single serialization.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
LiDAR scene completion, point cloud, diffusion models, state space models, Mamba, serialization, 3D vision, autonomous driving

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_H2YCOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents LiMamba, a diffusion model for LiDAR scene completion that combines a sparse convolutional backbone with Mamba blocks to capture broader spatial context. Its main idea is to process point features in three parallel orders—Z-order, radial, and azimuth—so that each branch captures different spatial relationships. Learned gates fuse the branches, while scan dropout encourages the model to use multiple views. Experiments show improved completion quality on SemanticKITTI and zero-shot transfer to KITTI-360.

Critical strengths

1.The diffusion-only gains are convincing: 0.1 m IoU rises from LiDPM’s 6.3 to 8.5. The range analysis also shows larger gains far from the sensor, where completion is harder. 2. The scan, dropout, fusion, and placement ablations are helpful. Zero-shot transfer and separate training/inference cost reports make the evaluation more useful.

Critical weaknesses

1.The benefit of physically motivated scans is not fully isolated. The claimed capacity matching lacks per-variant details, and there is no repeated-order, three-branch control. The gains could partly come from extra branches rather than the chosen geometry. 2. The overall advantage is mixed. After refinement, 0.1 m IoU trails LiDPM (11.3 vs. 14.9), while sampling is slower (29.2 vs. 26.4 s). Windowed attention has no accuracy comparison, so its timing alone cannot establish a better accuracy–cost trade-off. 3. The claim that scan dropout prevents gate collapse lacks direct evidence. Table 6 shows better completion scores, but no gate-weight statistics or training curves. This supports dropout’s usefulness, not the proposed explanation. 4. The ablations use only 100 validation scans, with no reported variation across training seeds. Several differences are small, so it is hard to tell whether the preferred dropout rate and fusion design are consistently better.

Questions for authors

1.How were parameters and FLOPs matched within 2% across the ablations? Please provide the actual configurations and counts. 2. Was windowed attention used at all four deep stages or only enc3/enc4, as Table 10 states? Is completion accuracy available under the same protocol?

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

This paper proposes LiMamba for diffusion-based LiDAR scene completion. It combines sparse convolutions with Mamba branches using Z-order, radial, and azimuth serialization, followed by gated fusion and Scan Dropout. Experiments on SemanticKITTI and zero-shot KITTI-360 evaluate completion quality, efficiency, and individual components.

Critical strengths

1.The diffusion-only results demonstrate consistent improvements over a competitive baseline, including an increase in 0.1 m IoU from 6.3 to 8.5. 2.The paper provides extensive component ablations and zero-shot evaluation, supporting the empirical usefulness of the overall design.

Critical weaknesses

1.Limited methodological novelty. The method primarily combines established sparse-convolution, Mamba, and multi-order serialization designs within a diffusion framework. The specific technical advance introduced by the sensor-aware orderings is not sufficiently demonstrated beyond adapting existing ideas to scene-scale completion. Consequently, the overall contribution appears incremental.

2.Insufficient coverage of recent methods. The quantitative comparisons do not adequately cover recent advances in LiDAR scene completion. A more up-to-date evaluation under consistent protocols is needed to establish whether the reported improvements remain competitive in both completion quality and computational efficiency.

Questions for authors

1.What new technical insight does the proposed sensor-aware serialization provide beyond established multi-order sequence modeling?

2.Can the authors provide a more up-to-date comparison with representative approaches under consistent evaluation settings? Flag For Ethics Review

Reviewer_3c7kOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes LiMamba, a diffusion denoiser for scene-scale LiDAR point cloud completion. It augments a sparse-convolution U-Net with Mamba blocks at deeper stages. Each block processes voxel features in Z-order, radial, and azimuthal order, then combines the three streams through learned gating; Scan Dropout encourages the streams to remain useful. Experiments on SemanticKITTI and zero-shot KITTI-360 report improvements on several geometric and coarse-resolution occupancy metrics, with ablations supporting the multi-scan design.

Critical strengths
  1. The paper identifies a concrete challenge in applying sequence models to scene-scale point clouds: the serialization order affects which geometric relationships the model can capture. The component ablation shows a consistent gain from the proposed combination: on the reduced SemanticKITTI subset, 0.1 m IoU increases from 7.4 with the best single scan to 8.5 with all three scans.
  2. The evaluation includes diffusion-only and refined results, cross-dataset testing, and runtime measurements. This makes both the benefits and the costs of the proposed backbone relatively clear.
Critical weaknesses
  1. The comparison does not establish an advantage over a practical diffusion Transformer. Explicit global attention runs out of memory, but this only rules out that particular implementation at the tested resolution. The feasible windowed-attention baseline reports memory and speed without completion quality, and its reported module placement differs from the default LiMamba placement. A Transformer operating on pooled or lower-resolution features would be a more informative comparison. Without a matched quality–latency–memory evaluation, the architectural motivation for choosing Mamba remains insufficiently tested.
  2. The quality gains come with costs and are not consistent across metrics. LiMamba’s reported 50-step sampling time is 29.2 s, versus 26.4 s for LiDPM. In the refined SemanticKITTI results, LiMamba improves 0.5 m IoU from 36.6 to 38.0, but trails LiDPM at 0.1 m IoU (11.3 versus 14.9). These results make the overall benefit less clear for applications that require fine-resolution completion.
Questions for authors
  1. Can the authors report completion quality, sampling time, and peak memory for a memory-feasible Transformer denoiser, for example one using pooled or lower-resolution features? Please clarify how its output resolution and computational budget compare with LiMamba.
  2. What is the completion quality of the windowed-attention baseline in Appendix Table 10? Can the comparison use the same module placement as LiMamba?
  3. What explains the lower fine-resolution IoU after refinement, and how should readers assess this trade-off when selecting the final pipeline?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.