PointMamba: A Simple State Space Model for Point Cloud Analysis

Transformers have become one of the foundational architectures in point cloud analysis tasks due to their excellent global modeling ability. However, the attention mechanism has quadratic complexity, making the design of a linear complexity method with global modeling appealing. In this paper, we propose PointMamba, transferring the success of Mamba, a recent representative state space model (SSM), from NLP to point cloud analysis tasks. Unlike traditional Transformers, PointMamba employs a linear complexity algorithm, presenting global modeling capacity while significantly reducing computational costs. Specifically, our method leverages space-filling curves for effective point tokenization and adopts an extremely simple, non-hierarchical Mamba encoder as the backbone. Comprehensive evaluations demonstrate that PointMamba achieves superior performance across multiple datasets while significantly reducing GPU memory usage and FLOPs. This work underscores the potential of SSMs in 3D vision-related tasks and presents a simple yet effective Mamba-based baseline for future research. The code will be made available at \url{https://github.com/LMD0311/PointMamba}.

Paper

Similar papers

Peer review

Reviewer MyQv5/10 · confidence 5/52024-06-16

Summary

In summary, PointMamba, an innovative state space model tailored for point cloud analysis, successfully harnesses the global modeling prowess of Mamba, a representative SSM from the NLP domain. By adopting a linear complexity algorithm, PointMamba addresses the computational challenges posed by traditional Transformer-based methods while maintaining their global modeling capabilities. Its key techniques include utilizing space-filling curves for effective point tokenization and employing a non-hierarchical Mamba encoder as the foundation. Extensive experiments across multiple datasets validate the superior performance of PointMamba while significantly reducing GPU memory usage and FLOPs. This work not only demonstrates the vast potential of SSMs in 3D vision tasks but also provides a simple yet effective baseline for future research in this domain.

Strengths

1. **Linear Complexity with Global Modeling**: PointMamba leverages state space models to achieve linear complexity while maintaining global modeling capabilities, overcoming the computational challenges of traditional Transformers. 2. **Efficient Point Cloud Representation**: The use of space-filling curves for point tokenization enables efficient representation of point clouds, capturing spatial structure while facilitating global feature extraction. 3. **Simple and Effective Mamba Encoder**: The non-hierarchical Mamba encoder provides a simple yet powerful backbone for PointMamba, enabling fast and accurate global feature modeling. 4. **Superior Performance**: Comprehensive evaluations show that PointMamba achieves state-of-the-art performance across multiple datasets, demonstrating its effectiveness for point cloud analysis tasks.

Weaknesses

1. Although PointMamba borrows structurally from Mamba, it may not take full advantage of the unique characteristics of point cloud data, such as spatial distribution, density variations, and local geometry. 2. PointMamba, while inheriting the strengths of Mamba, may have inherited some of the limitations of its design for natural language processing tasks. These limitations may not be applicable to point cloud analysis tasks, such as the lack of specific preprocessing steps, feature extraction methods, or post-processing techniques for point cloud data. The feature extraction and processing steps in the paper are largely similar to those of the previous PointMAE. 3. Although PointMamba achieves global modeling of linear complexity by introducing state-space models, there may be trade-offs between model complexity and performance in real-world applications. Because according to the experimental results of the thesis, the effect of partially using PointMamba becomes worse instead. 4. According to Table 5 of the ablation experiment, the serialization operations Hiber and Trans-Hiber have the most obvious impact on the experimental results. Meanwhile, the framework diagram at the core of the paper, i.e., Figure 4, shows that the only change is Hiber, from the directly cited literature [27], and there is almost no additional description or setting in the paper.

Questions

1. Based on weakness 4, is it possible that PointMamba actually only works well in Hiber conditions? As this very much affects the final results and conclusions of the paper. 2. Intuitively, PointMamba seems to be just a replacement of the Mamba block (Equation 4) for the previous Transformer block, and even if this results in good performance (Figure 1), I am not convinced. The paper is theoretically and experimentally fleshed out, but I still don't find an obvious innovative contribution. Therefore, I would like to confirm whether the authors have designed a specialized Mamba technique for 3D point clouds?

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

Content-wise, the paper has no obvious limitations, and its core technology builds on Mamba, which is popular in other fields. Formatting-wise, the paper shows the mamba icon several times; please confirm whether this representation is appropriate in an academic paper.

Reviewer vmtW5/10 · confidence 5/52024-06-16

Summary

This paper propose a simple but effective Mamba-based method named PointMamba for point cloud analysis. This paper is the first paper that studies the Mamba-based method for point clouds. The experiments are comprehensive and the paper is in very good shape.

Strengths

1. This paper demonstrates excellent writing quality, with a clear and evident motivation. 2. Figure 1 is highly comprehensible, particularly in its comparisons with Point-MAE.

Weaknesses

1. (No need for experiments) The use of validation datasets is prevalent in the field, with consistently high metrics reported. I encourage the inclusion of indoor segmentation and detection tasks, as exemplified in Point-M2AE and MaskPoint, in future iterations of your work. Additionally, tackling the classification task on Objaverse-LVIS appears to be a more demanding and stimulating challenge. 2. To be honest, there are numerous papers that adhere to the evaluation paradigm established by Point-BERT. However, I am genuinely interested in exploring novel discoveries in self-supervised learning specifically applied to point clouds.

Questions

1. It has been experimentally determined that Point-MAE, Point-M2AE, and MaskPoint do not accurately replicate their reported performance, as observed in this study. It is important to investigate whether PointMamba exhibits consistent stability across all of these tasks. If not, providing mean and standard deviation values similar to those presented in Table 3 is crucial. 2. Additionally, it has been observed that conducting experiments with a larger number of points using Mamba results in an increase in training time. Have you encountered this phenomenon in your own experiment? If so, kindly elaborate on possible solutions to enhance its suitability for real-world applications, such as Auto-Driving. 3. Could you please elaborate on the reasons why incorporating Hilbert and Trans-Hilbert techniques in your current version has led to improved performance compared to the reordering strategy implemented in the initial version? Moreover, Hilbert and Trans-Hilbert techniques seem to borrow from Point Transformer. If so, the main contribution lies in integrating Mamba into Point-MAE. 4. (No need for experiments, just discuss.) There are several Mamba-based methods after this paper. Please just discuss the advantages and disadvantages compared with them.

Rating

5

Confidence

5

Soundness

3

Presentation

4

Contribution

3

Limitations

1. The authors have addressed the limitations and potential negative societal impact of their work.

Authorsrebuttal2024-08-05

Rebuttal by Authors (continued)

- **To question 4: “…several Mamba-based methods after this paper…”** **Reply:** Good suggestion! Indeed, there are several papers after this paper, we list the comprehensive comparisons as follows: **PCM [1]** combines Vision Mamba with PointMLP and incorporates consistent traverse serialization at each stage. To enhance Mamba’s capability in managing point sequences with varying orders, PCM introduces point prompts that convey the sequence’s arrangement rules. **Point Mamba [2]** uses an octree-based ordering scheme and combines Mamba with PCT and OctFormer as their baseline. Mamba blocks with bi-directional scanning extract hierarchical point features, and a Feature Pyramid Network (FPN) is utilized for classification or segmentation tasks. **Mamba3D [3]** introduces an enhanced Vision Mamba block, which includes both a token forward SSM and a backward SSM that operates on the feature channel. It proposes a Local Norm Pooling block to extract local geometric features. **PointTramba [4]** introduces a hybrid approach that integrates Transformers and Mamba. It segments point clouds into groups and utilizes Transformers to capture intra-group dependencies, while Mamba models inter-group relationships using a bi-directional, importance-aware ordering strategy. **Note that some methods (e.g., [4]) are clearly based on our code or baseline, which proves the value of our approach to the community. We will add these discussions in the revised version.** [1] Point could mamba: Point cloud learning via state space model [2] Point mamba: A novel point cloud backbone based on state space model with octree-based ordering strategy [3] Mamba3d: Enhancing local features for 3d point cloud analysis via state space model [4] PoinTramba: A Hybrid Transformer-Mamba Framework for Point Cloud Analysis

Reviewer QPV38/10 · confidence 5/52024-07-12

Summary

This paper introduces PointMamba, an interesting method for point cloud analysis that utilizes a linear complexity state space model (SSM) instead of traditional Transformer architectures. PointMamba employs space-filling curves for point tokenization and features a simple, non-hierarchical Mamba encoder. Additionally, the authors propose an effective order indicator and serialization-based mask modeling strategy. Comprehensive evaluations across multiple datasets demonstrate PointMamba's superior performance and significantly reduced computational costs compared to Transformer-based methods.

Strengths

1. The paper reads very well. I appreciate the motivation of the paper and agree that we should make an effective method as simple as possible. As the first Mamba-based work for point cloud tasks, the proposed method is simple, elegant, and effective, establishing a solid Mamba-based baseline for point cloud analysis tasks. 2. The paper considers the limitations of Mamba's unidirectional modeling and proposes practical solutions like serialization-based mask modeling strategies and order indicators. 3. The theoretical analysis of Mamba used in point cloud is reasonable, and the paper is easy to reproduce. 4. The experiments are convincing and support the main idea of the paper. The authors provide thorough evaluations, including comparisons with SOTA methods, ablation studies, and analyses of each component.

Weaknesses

1. Some experiments are missing. For example, from Table 11, it would be beneficial to include more masking ratios, especially the performance when masking 90% point patches. Besides, from Table 12, could the authors provide a more detailed analysis of why average pooling performs better? Additionally, what about the performance of max pooling? 2. Have the authors considered the potential benefits of integrating PointMamba with Transformer architectures? For example, a transformer layer can be used as the final prediction head. Such a combination can take advantage of both architectures and will not introduce many computational costs. 3. The point clouds have complex structures, and this paper only considers two orders. In my view, introducing more orders (e.g., introducing three serialization methods and tripling the input length) can better capture the geometry information of the point clouds, which might be beneficial for learning. 4. During the pre-training strategy, do the authors use different order indicators or the same indicator? Figure 4 is somewhat ambiguous. Minor: Using space-filling curves to scan point clouds is an interesting attempt. Although it is common knowledge for some readers, it still would be helpful if the authors provided a more detailed introduction to space-filling curves in the preliminaries part. typos: PointMAE -> Point-MAE

Questions

See weakness.

Rating

8

Confidence

5

Soundness

4

Presentation

4

Contribution

4

Limitations

The authors have adequately discussed the limitations of their work.

Reviewer Daqu3/10 · confidence 5/52024-07-14

Summary

This work utilizes the Mamba architecture for point clouds. It employs Hilbert and Trans-Hilbert curves to order the point clouds, thus addressing the unidirectional modeling nature of Mamba. Additionally, it replaces transformer blocks with Mamba blocks. The proposed PointMamba demonstrates reasonable performance following pre-training.

Strengths

1. This work tries introducing a new network architecture to the point cloud domain, which is appreciated. 2. The presentation is clear, accompanied by well-crafted figures.

Weaknesses

1. While it is worthwhile to explore how the Mamba architecture performs when applied to point clouds, this work exhibits limited novelty or insights in architectural innovation for point clouds. The authors themselves acknowledge this by stating, “It should be noted that this paper does not claim algorithmic novelty but rather presents a simple and solid Mamba-based baseline.” In my opinion, works that directly adopt an existing network from other domains to point clouds without sufficient insights and innovation should not be accepted by top-tier conferences like NeurIPS. 2. The authors only demonstrate the performance of **pre-trained** PointMamba by comparing it with other widely used baselines like PointNext. This implies that PointMamba cannot surpass previous methods when trained from scratch. To effectively showcase the solid merits of PointMamba, it would be more reasonable to provide comparisons without pre-training. 3. In Table 4, an important baseline, PointNext, is omitted. According to the PointNext paper, PointNext exhibits significantly better performance than the proposed PointMamba (87% without pre-training vs. 84.4% with pre-training). This omission raises doubts about the effectiveness of PointMamba. To give a "solid Mamba-based baseline", comprehensive comparisons with previous methods should not be omitted. 4. In Figure 1, it would be beneficial to include comparisons with widely used methods like PointNext to convincingly demonstrate the necessity of introducing Mamba for point clouds. While PointMamba likely offers inference advantages, this contribution stems from the original Mamba paper, not this work. Additionally, does PointMamba have lower training efficiency? 5. The paper claims that the proposed Hilbert and Trans-Hilbert ordering is advantageous based on results from ScanObjectNN. However, I am not entirely convinced. I suggest the authors also present results on ShapeNet-Part and ModelNet40, for which only comparing against a random baseline is enough.

Questions

See the weakness section.

Rating

3

Confidence

5

Soundness

2

Presentation

3

Contribution

2

Limitations

See the weakness section.

Authorsrebuttal2024-08-12

We are open to any further discussion.

Dear Reviewer Daqu, We sincerely appreciate your time and effort in reviewing our paper. We hope our explanations have addressed your concerns. As we are in the discussion phase, we welcome any additional comments or questions regarding our response or the main paper. If further clarification is needed, please do not hesitate to mention it, and we will promptly address your inquiries. We look forward to receiving your feedback. Best regard, Paper 940 Authors

Authorsrebuttal2024-08-13

Dear Reviewer Daqu, Thank you for your time and valuable feedback. As the discussion phase is nearing its end, we remain open to addressing any remaining questions or concerns. We would greatly appreciate it if you could consider improving the evaluation after reviewing our responses. Thank you very much for your consideration. Sincerely, Paper 940 Authors

Reviewer Daqu2024-08-13

Dear Authors, Thank you very much for providing additional experimental results. I have no further questions at the moment and would like to decide on my final score after discussing it with AC and the other reviewers. Have a good day! Best, Reviewer

Authorsrebuttal2024-08-14

Dear reviewer Daqu, We sincerely thank you for the time and feedback. We hope our existing rebuttal has addressed your previous concerns well. If you have any further questions during the next discussion period, please let us know, and we would be happy to answer them. Thank you once again! Sincerely, Paper 940 Authors

Reviewer MyQv2024-08-08

Reviewer's Response

Thanks to the authors for the reply. My concerns have been partially addressed, mainly because I am not completely sold on PointMamba's results. As a result, after the authors' rebuttals, I still believe it has merit, so I'm willing to change my rating to positive.

Authorsrebuttal2024-08-08

We appreciate your thought-provoking reviews and are pleased to see your positive decision. To substantiate our results, we will release the code. Thank you once again for your positive rating.

Reviewer vmtW2024-08-08

Reviewer's Response

Thanks for the authors' responses, most of my questions have been solved. Therefore, I will keep my rating as positive. I look forward to your future work exploring PointMamba on more challenging datasets.

Reviewer QPV32024-08-11

Thank you for the rebuttal! Most of my concerns have been addressed. I have one more question regarding the Hilbert space-filling curve: How is the starting point of the curve determined?

Authorsrebuttal2024-08-11

We appreciate the reviewer acknowledging that our rebuttal addressed the concerns. In response to the new question, specifically, we voxelize the key points of the point cloud and shift the minimum coordinate to the origin, with the voxel at (0,0,0) serving as the starting point of the Hilbert curve. We will add this explanation in the revised version.

Reviewer QPV32024-08-13

Thank you for the further response. Now, all of my concerns have been addressed well. After considering other reviewers' comments and the authors' rebuttal, I am now more convinced about the value of the paper. I believe this work has great potential to contribute to the NeurIPS community. Therefore, I tend to accept this paper and encourage the authors to incorporate the above discussions.

Authorsrebuttal2024-08-13

We sincerely thank the reviewer for the constructive feedback and support. Your comments are valuable for us to improve the quality of this work. We will incorporate your suggestions and clarify key points in the revision. Many thanks again!

Reviewer Daqu2024-08-12

Hi authors, Thanks for your time. For the performance of "Simply replace Transformer with Mamba in Point-MAE", did you pre-train this model or train it from scratch?

Authorsrebuttal2024-08-12

Dear reviewer, Thanks for your kind reply! To ensure a fair comparison, the setting of "**Simply replace Transformer with Mamba in Point-MAE**” refers to swap the Transformer with Mamba in the architecture of Point-MAE, followed by applying the same pre-training strategy as used in the default Point-MAE. Please note that the data augmentations used are also kept the same. To further compare the results under the from-scratch, we are currently conducting experiments and will update you within the next few hours. We appreciate your patience and will provide the results as soon as possible. Thank you once again for your kind reply!

Authorsrebuttal2024-08-12

Dear Reviewer Daqu, Thank you for your patience. We have now provided a comprehensive comparison, including Point-MAE, simply replacing Transformer with Mamba in Point-MAE, and our proposed PointMamba. As shown in the Table, the results clearly demonstrate that **simply replacing the Transformer with Mamba does not perform well, whether trained from scratch or with pre-training**. In contrast, our proposed PointMamba achieves promising results, benefiting from the key designs we proposed. We will be happy to include the discussions of pre-trained and scratch in the revised version, and we hope this will address the raised concerns. | Method | OBJ-BG (scratch/pre-training) | OBJ-ONLY (scratch/pre-training) | PB-T50-RS (scratch/pre-training) | ModelNet40 (scratch/pre-training) | | --- | --- | --- | --- | --- | | Point-MAE | 91.05/92.77 | 90.02/91.22 | 86.05/89.04 | 92.3/93.2 | | Simply replace Transformer with Mamba in Point-MAE | 90.36/92.25 | 89.50/90.69 | 85.58/87.11 | 91.8/92.4 | | PointMamba (ours) | 91.74/94.32 | 90.19/92.60 | 87.27/89.31 | 92.4/93.6 | Given that the discussion phase is quickly passing, we look forward to your reply and thank you once again for your time. Best regard, Paper 940 Authors

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC