CSI: Enhancing the Robustness of 3D Point Cloud Recognition against Corruption

We are glad that the reviewer found our topic crucial for real-world applications of point clouds and our method effective. We appreciate the opportunity to address the points you have raised.

The novelty of the proposed density-aware sampling may be limited as similar ideas have already been explored by previous methods [1,2], but the authors do not provide any comparisons with them.

We appreciate the reviewer’s concerns regarding the novelty of our proposed Density-Aware Adaptive Sampling (DAS) and the challenges in directly comparing it with methods presented in references [1] and [2]. The method in [1] fundamentally differs from ours in terms of output format and integration with subsequent processing layers. Specifically, our DAS outputs a format of (S, S, H) and is intricately linked with spherical voxel convolution layers, making it incompatible as a direct replacement for FPS or DAS in PCT architectures. Therefore, a direct comparison is not feasible due to these structural and functional differences.

Regarding the method in [2], we faced the challenge of no open-sourced code being available. Despite this, we endeavored to reproduce their method to the best of our ability. Due to time constraints, we implemented a version of Grid Density Sampling (GDS) with a grid size of 8, as larger grid sizes would have necessitated significantly longer training times. The results of this implementation are presented:

While it is true that the underlying motivation of our DAS may share similarities with these previous works, the implementation and the operational context significantly differ. A key distinction of our DAS is that it is not a grid-based approach, unlike GDS. We believe that grid-based operations can potentially disrupt the inherent structure of point clouds. Our ablation studies on ModelNet40-C and PointCloud-C demonstrate that our non-grid-based approach is more effective, supporting the unique contribution and novelty of our DAS methodology in preserving the integrity of point cloud structures while enhancing sampling efficiency.

The writing is hard to read and follow. For example, too many sentences that are too long, inconsistent usage of section references.

We greatly appreciate your feedback regarding the readability of our manuscript. We recognize that clear and accessible writing is crucial for effective communication of our research. To address your concerns, we will undertake a comprehensive review of the manuscript to simplify and shorten complex sentences, ensuring they are more reader-friendly. Additionally, we will systematically revise the usage of section references to maintain consistency throughout the text. Our goal for the final version is to enhance overall clarity and ease of reading, making our research more accessible to a broader audience. We are committed to improving these aspects to ensure our work is communicated as clearly and effectively as possible.

We are glad that the reviewer found our method has demonstrated improvements. We appreciate the opportunity to address the points you have raised.

The paper could improve the clarity of exposition. Some parts of the paper, particularly in the methodology section, are not explained in a clear and concise manner, which may impede the reader's understanding.

We appreciate the feedback on the clarity of our exposition, especially in the methodology section. We understand that clear and concise communication is crucial for the reader's comprehension. To address this, we will undertake a thorough revision of our manuscript, with a specific focus on enhancing the clarity and conciseness of the methodology section. We aim to ensure that our final version presents our methods and findings in a more accessible and understandable manner. This revision will include refining complex explanations, simplifying technical jargon where possible, and providing additional context or examples to facilitate better understanding.

The paper could benefit from more detailed evaluation and ablation studies. While the experimental results show the superiority of the CSI method, it would be valuable to have a more in-depth analysis of its performance and a comparison with widely-known baselines in the field.

Thank you for the constructive feedback. We acknowledge the importance of in-depth analysis and comprehensive comparisons in our research. Although our experimental section already evaluates numerous baselines within this domain, we understand the need for deeper insights into the performance of the CSI method. In light of your suggestions, we will expand our analysis to include more detailed evaluations and additional ablation studies. These enhancements will focus on a finer comparison with well-established baselines in the field, and on dissecting the individual contributions and impacts of the various components of the CSI method. Our aim is to provide a clearer, more comprehensive understanding of why and how CSI demonstrates superiority over these baselines, thereby enriching the overall value and contribution of our work to the field.

I think the method proposed in ModelNet40-C and PointCloud-C should also be compared and analyzed.

Thank you for emphasizing the importance of comparative analysis with ModelNet40-C and PointCloud-C. We have indeed conducted such comparisons and included them in our manuscript. Specifically, Table 1 in the main text and Tables 1 and 2 in the appendix present a comprehensive comparison of our method against others mentioned in the original papers of ModelNet40-C and PointCloud-C. These tables detail the performance metrics and demonstrate how our method stacks up against existing methods under similar conditions. We believe this thorough comparison effectively illustrates the strengths and limitations of our proposed approach in the context of these well-established datasets.

We are glad that the reviewer found our study valuable to this field and that our method was simple yet effective. We appreciate the opportunity to address the points you have raised.

Please explain the statement “local density of a point positively correlated with its significance”? Previous work has indicated that significant points are usually outward points in the point cloud, which are generally sparse[1]. Also, traditional sampling methods like FPS tend to find points with low density as representations.

The observation that significant points are predominantly located at the outer regions of the point cloud does not hold true in scenarios involving data corruption, as demonstrated in Figure 3. In such cases, Farthest Point Sampling (FPS) tends to retain an excessive number of noisy data points. This limitation of FPS is a key motivator for our proposal of the Density-Aware Sampling (DAS) method. DAS is specifically designed to mitigate the adverse effects caused by FPS, particularly its tendency to sample a higher number of noisy points in the presence of various forms of corruption, such as background noise and global noise.

The authors should compare with other SOTA train-time methods under the same settings in the main experiment, such as data augmentation or modules addition to the model.

We acknowledge the suggestion to benchmark our approach against other state-of-the-art (SOTA) train-time methods under equivalent experimental conditions, including factors like data augmentation and module additions. In response, we direct attention to Table 2 in the main manuscript, which provides a comprehensive comparison where all evaluated methods, including ours, are integrated with the data augmentation technique PointCutMix-R. Additionally, for a more extensive comparison, Table 2 in the supplementary material, ranging from the entry 'DGCNN+OcCo' to 'PCT+CSI+PointCutMix-R', offers a detailed analysis of various other train-time methods. This inclusion ensures a thorough and fair evaluation of our method against current SOTA techniques under consistent experimental settings.

The authors present the SEM method for Global Feature in Table 5. Given that the authors are making a critical point selection, what is the purpose of this?

The incorporation of the SEM method for the global feature, as detailed in Table 5, is intended as an additional ablation study. Similar to the critical point selection process, the application of SEM is designed to enhance the distinctiveness of global features. The experiments presented in Table 5 investigate the individual and combined effects of critical point selection and SEM application. Our findings indicate that while applying SEM to both local and global features yields performance improvements, the point-level selection demonstrates a more significant benefit. This comparison underscores the unique contribution of each technique to the overall model performance.

The experiments should consider including datasets not based on ModelNet40.

In addressing the recommendation to include datasets beyond ModelNet40 in our experiments, we emphasize our focus on point cloud recognition within this work. ModelNet40, along with its two corrupted variants, ModelNet40-C and PointCloud-C, are extensively recognized and employed in related research. These datasets are particularly relevant and suitable for the objectives of our study. Our decision to utilize these datasets is grounded in their widespread acceptance in the field, ensuring that our findings are comparable and relevant to current standards in point cloud recognition research.

Tables 1 and 2 in Supplementary show that CSI cannot effectively handle point dropping (e.g., occlusion) and transformations (e.g., rotation), and may even be harmful.

We appreciate the reviewer's observation regarding the performance of CSI under conditions of point dropping and transformations, as shown in Tables 1 and 2 of the supplementary material. We acknowledge that these results may highlight a limitation in the generalizability of the CSI approach, particularly in handling occlusions and rotations. This limitation is indeed an area that warrants further investigation and improvement in future iterations of our work. Regarding the use of ModelNet40-C, we understand the concern that the results derived from this dataset might present a skewed perspective. We will consider this point critically in our analysis and discussion, ensuring that we clearly communicate the potential limitations of our findings in the context of different datasets and conditions. However, we argue that it is hard for a method to cure all corruption types. This feedback is invaluable for guiding our ongoing research efforts to refine and enhance the robustness and applicability of our approach.

Thank you to the authors for providing new experimental results and additional interpretations. However, some concerns remain unaddressed.

These datasets are particularly relevant and suitable for the objectives of our study.

ModelNet40-C and Point Cloud-C might be insufficient due to significant overlaps in these datasets, such as the "add global" in Point Cloud-C and the "background" in ModelNet40-C. The authors might consider incorporating a wider variety of noise datasets to strengthen the argument.

Table 2 in the supplementary material, ranging from the entry 'DGCNN+OcCo' to 'PCT+CSI+PointCutMix-R', offers a detailed analysis of various other train-time methods...

The additional data provided by the authors are greatly appreciated. However, these data suggest that the pure CSI method does not outperform methods with data augmentation (PCT+CSI v.s. PCT+WOLFMix), and even the enhanced version (PCT+ CSI +PointCutMix-R) fails to achieve the best results.

Finally, I understand and agree with the authors' statement that it is challenging for a method to address all types of corruption. The suggestion is for the authors to focus on specific types of corruption to further explore and demonstrate the unique capabilities of CSI.

We are glad that the reviewer found our method effective. We appreciate the opportunity to address the points you have raised.

SEM is mostly based on previous knowledge that entropy minimization helps classification robustness, which slightly undermines the significance of the proposed techniques. Nonetheless, the paper provides a detailed discussion of how entropy minimization should be applied to transformer-based point classifiers in both attention layers and the classification head, accompanied by sufficient ablation studies.

Although previous work has investigated how entropy minimization helps classification robustness, in their works, they merely investigated how to minimize entropy during inference. For example, tent [1] minimizes Shannon Entropy on classification logits through several iterations. The biggest difference is that we give a thorough analysis of how to apply entropy minimization during training and which features contribute most by conducting ablation studies not only on transformer self-attention maps but also on intermediate features.

[1] Wang, Dequan, et al. "Tent: Fully test-time adaptation by entropy minimization." arXiv preprint arXiv:2006.10726 (2020).

It is not clear how general DAS is and how it affects the classifier's robustness to more types of corruptions other than global noise addition shown in Figure 3. Table 1 and Table 2 in the supplementary material ablate CSI as a whole so they can not show the effect of DAS. It would be better if DAS could be individually studied on different types of corruption.

More visualization results are available in the appendix of our revised manuscript. One can observe the effectiveness of DAS compared with FPS (Farthest Point Sampling) and RS (Random Sampling). Besides, our quantitative experimental results have also validated that DAS has outperformed other baseline methods in improving corruption robustness.