ConDaFormer: Disassembled Transformer with Local Structure Enhancement for 3D Point Cloud Understanding

Transformers have been recently explored for 3D point cloud understanding with impressive progress achieved. A large number of points, over 0.1 million, make the global self-attention infeasible for point cloud data. Thus, most methods propose to apply the transformer in a local region, e.g., spherical or cubic window. However, it still contains a large number of Query-Key pairs, which requires high computational costs. In addition, previous methods usually learn the query, key, and value using a linear projection without modeling the local 3D geometric structure. In this paper, we attempt to reduce the costs and model the local geometry prior by developing a new transformer block, named ConDaFormer. Technically, ConDaFormer disassembles the cubic window into three orthogonal 2D planes, leading to fewer points when modeling the attention in a similar range. The disassembling operation is beneficial to enlarging the range of attention without increasing the computational complexity, but ignores some contexts. To provide a remedy, we develop a local structure enhancement strategy that introduces a depth-wise convolution before and after the attention. This scheme can also capture the local geometric information. Taking advantage of these designs, ConDaFormer captures both long-range contextual information and local priors. The effectiveness is demonstrated by experimental results on several 3D point cloud understanding benchmarks. Code is available at https://github.com/LHDuan/ConDaFormer .

Paper

Similar papers

Peer review

Reviewer 32ZL6/10 · confidence 4/52023-06-28

Summary

This paper proposed a network for point cloud understanding. The basic idea is to process point clouds in 3 orthogonal 2D planes (triplanes). This can largely reduce the computation cost compared to processing point clouds in 3D space. Experiments on several tasks (segmentations and detection) are conducted to show the effectiveness of this method.

Strengths

This paper showed an insight of point cloud processing: when we are dealing with data of large dimensionality, it is better that we can first project the data into a smaller space. The figures nicely illustrated the core idea. Even though the triplane-style network has been used in many other works, I still believe the paper did a great job in designing the backbone network. Some experiments on segmentation and detection can also show the potential usage of the proposed network. Overall, I believe the paper delivered a great idea in designing 3D networks and some good results. My major concern is about the experiement part (see the weakness section).

Weaknesses

1. As a general backbone network for point cloud understanding. Some experiments are missing. For example, object-level classification (ShapeNet, ModelNet, ScanObjectNN), object-level segmentation (ShapeNetPart, PartNet). 2. As mentioned above, the triplane-style network has been used in some prior works (e.g., [1]). 3. A minor issue, it is difficult to understand the notations of attention matrix for 3 planes (-, ~ and ^). I know because of the superscripts and subscripts, the authors had to choose (-, ~ and ^) to denote different planes. But maybe we can find something better. [1] Efficient Geometry-aware 3D Generative Adversarial Networks

Questions

See the weakness section.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

N/A

Reviewer Kwdg5/10 · confidence 5/52023-07-06

Summary

This paper proposes a new window partitioning method, which can save a lot of computation costs by sacrificing a small amount of precision. At the same time, it proposes a kind of depth wise sparse convolution, which can better capture local structure by using before and after self attention. The experimental results fully prove the effectiveness of the method

Strengths

1. This paper innovatively puts forward a new window partitioning method, which divides 3D cubic into 3 2D planes and then divides the Windows. This method can save a lot of calculation costsThis paper presents a new regularization method to optimize networks by predicting relative position differences 2. By using depth wise sparse convolution to capture local structures, the experimental results fully demonstrate the effectiveness of this design 3. The content is sufficient and the experimental results are abundant

Weaknesses

1. The interaction between different planes is only additive operation, which may lead to the loss of 3D structure information, 2. More relevant work on window partitioning methods is basically not discussed and compared

Questions

1. I think a better interaction between the planes might improve task performance even more

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

limitations have benn discussed in the paper

Reviewer teZK5/10 · confidence 5/52023-07-06

Summary

This paper studies point cloud understanding. They propose ConDaFormer, a novel Transformer architecture for 3D point cloud. Specifically, ConDaFormer disassembles the cubic window into three orthogonal 2D planes, leading to fewer points when modeling the attention in a similar range. Together with local sparse convolutions, ConDaFormer can capture both long-range contextual information and local priors. They evaluate their method on point cloud detection and segmentation datasets and achieve good performance.

Strengths

- Their method has substantial improvements over the existing point cloud Transformers. - The proposed architecture is simple and effective. The method has sufficient novelty. - They evaluate their approach on widely-used point cloud datasets and achieve satisfactory results. - Paper writing is good and easy to follow.

Weaknesses

- The proposed model has some similarities with FlatFormer [1], which groups the patches by axis sorting. The axis sorting is similar to disassembled window attention proposed in this paper. The authors need to clarify the difference with the existing works. - The proposed model failed to outperform the SoTA detectors on SUN-RGBD. [1] Liu et al. FlatFormer: Flattened Window Attention for Efficient Point Cloud Transformer.

Questions

Please refer to the weaknesses.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

Please refer to the weaknesses.

Reviewer JCcK5/10 · confidence 5/52023-07-07

Summary

In this paper, CondaFormer, an innovative 3D transformer architecture, is presented. It ingeniously dissects the cubic window into three orthogonal 2D planes and incorporates a local structure enhancement strategy that uses depth-wise convolutions to capture local geometric information. Through rigorous experiments on point cloud semantic segmentation and 3D detection, CondaFormer's efficacy is demonstrably validated.

Strengths

1. The rationale behind dissecting the cubic window into tri-planes is well-articulated and has a solid foundation. This approach considerably reduces computational cost by limiting query-key pairs. 2. CondaFormer showcases remarkable improvements in performance in the context of semantic segmentation, as evidenced by the results. 3. A series of comprehensive ablation studies are conducted, providing evidence for the effectiveness of the window disassembly design, local structure enhancement, and the impact of hyper-parameter choices. 4. The paper is well-structured and clearly presented, making it accessible and easy to follow.

Weaknesses

1. The overall concept, though practical, does not break new ground in terms of novelty. The disassembly of 3D windows into 2D planes can be seen as a straightforward adaptation of the Axial Transformer[1], which similarly disassembles 2D windows into 1D axis attention. ([1] Ho J, Kalchbrenner N, Weissenborn D, et al. Axial attention in multidimensional transformers[J]. arXiv preprint arXiv:1912.12180, 2019.) 2. The performance in 3D object detection leaves room for improvement, as CondaFormer does not exhibit a substantial edge over the baseline FCAF3D. 3. It is recommended that the authors further validate the model's performance and robustness by conducting additional experiments on outdoor perception tasks such as 3D object detection or segmentation on datasets like KITTI or Waymo.

Questions

see the strengths and weaknesses.

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

2 fair

Limitations

NA

Reviewer BFnK6/10 · confidence 4/52023-07-08

Summary

Recent advancements in 3D point cloud understanding have explored the use of Transformers, resulting in notable progress. However, the computational demands of applying global self-attention to large point cloud datasets, which contain over 0.1 million points, present a significant challenge. To mitigate this issue, researchers have proposed using Transformers within local regions, such as spherical or cubic windows. Nevertheless, these approaches still involve a considerable number of Query-Key pairs, leading to high computational costs. Moreover, previous methods often neglect the local 3D geometric structure by employing linear projections to learn the query, key, and value. In this paper, a new transformer block named ConDaFormer is introduced to address these challenges while also considering the local geometry prior. ConDaFormer decomposes the cubic window into three orthogonal 2D planes, reducing the number of points involved in attention modeling within a similar range. Although this disassembling operation sacrifices some contextual information, a local structure enhancement strategy is implemented using depth-wise convolutions before and after the attention step. This strategy effectively captures local geometric information. By leveraging these innovative designs, ConDaFormer is capable of capturing both long-range contextual information and local priors. Experimental results on various benchmarks for 3D point cloud understanding demonstrate the effectiveness of ConDaFormer.

Strengths

(1) The authors propose a novel disassembled window attention module for 3D point cloud semantic segmentation by disassembling the 3D window into three orthogonal planes for self-attention. This strategy effectively reduces computational overhead with negligible performance decrease. (2) To enhance the modeling of local features, the authors introduce depth-wise sparse convolution within the disassembled window attention module. This combination of self-attention and convolution provides a comprehensive solution for capturing both long-range contextual information and local priors in 3D point cloud data. (3) Experiments show that our method achieves state-of-the-art performance on widely adopted large scale 3D semantic segmentation benchmarks and comparable performance in 3D object detection task. Extensive ablation studies also verify the effectiveness of the proposed components.

Weaknesses

(1) In Table 1, why is there no results on test set for ConDaFormer? (2) In Table 5, you said "in comparison with FCAF3D, our method achieves comparable performance but performs more steadily.". What do you mean by more steadily? Do you have any experimental results to support that? (3) In Table 7, why don't try the window size smaller than 0.16m?

Questions

I'm positive about this paper. I really like the idea to disassemble 3D cubic window into three orthogonal planes for self-attention. It can reduce the computational cost. However, I still have some questions about the experimental results. Please see the Weaknesses and respond to those questions. Thank you.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

This paper still have some limitations when using larger attention window size. If the authors enlarge the window size from 0.32m to 0.48m, the training loss drops from around 0.52 to around 0.47 while the mIoU does not increase on the S3DIS dataset.

Area Chair GFRd2023-08-12

Author-reviewer discussion starts

Dear reviewers, Thanks for serving as reviewers. The authors have submitted a rebuttal. Please review through the rebuttal and reviews from other authors. If you have any questions, please feel free to let the authors know. You are more than welcome to post comments for further explanation or clarification before 1pm EDT on 8.21. Best, AC

Area Chair GFRd2023-08-18

Kind reminders for reviewers

Dear reviewers, If you have not responded to the authors' feedback, please take some time to read through their responses and reviews from other reviewers. We would be very pleased to hear your thoughts. Thanks, AC

Reviewer JCcK2023-08-18

The authors have conducted additional experiments that provide an effective rebuttal to my previous concerns. The new results demonstrate that CondaFormer outperforms the previous state-of-the-art CAGroup3D method on the SUN RGB-D dataset and improves the outdoor perception methods. This addresses my main criticism about the strength of the empirical results. I am willing to increase my score to be borderline accept.

Reviewer Kwdg2023-08-20

Thanks for the reply

Thanks to the author's thoughtful response, I feel that my questions have been mostly resolved, and I will maintain my rather positive rating

Reviewer 32ZL2023-08-21

Thanks for the clarification and the new results. I believe this paper proposed an interesting method and showed some good results.

Reviewer teZK2023-08-21

Thanks for the reply. The reviewers have addressed my concerns so I would keep my rating and recommend this paper for acceptance.

Reviewer BFnK2023-08-21

Keep my current rating

Thanks for the author's rebuttal. It resolved all of my concerns and I'll keep my rating for weak accept.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC