acceptodds
Withdrawn from ICLR 2027

RAD: Reconstruction-Anchored Diffusion Pretraining for Corruption-Robust Point Cloud Recognition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Corruption-robust point-cloud recognition has become an essential task for the realm of 3D vision. Current globally conditioned diffusion pretraining typically optimizes the encoder for whole-shape denoising through a pooled global condition. Although effective in transferring features for clean recognition, this approach leaves masked-patch geometry without an explicit prediction objective. In this work, we propose a framework, named as RAD, for corruption-robust point-cloud recognition, via complementing whole-shape generative pretraining with an explicit geometric prediction task. Specifically, RAD consists of two key modules, the Local Geometry Reconstructor and the Conditional Flow Denoiser. The former module, the Local Geometry Reconstructor, is designed to reconstruct masked patches from visible encoder tokens under Chamfer distance, thus supplying the missing geometric prediction objective. The Conditional Flow Denoiser retains whole-shape denoising as an auxiliary pretraining objective, thereby keeping the generative branch. Through the integration of these two modules, RAD jointly pretrains a single encoder that is transferred alone, effectively leaving the downstream classifier and fine-tuning protocol unchanged. Extensive experiments across diverse corruption benchmarks demonstrate that adding geometric reconstruction reduces mean Corruption Error relative to the diffusion baseline.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3D vision, Point Cloud, corruption

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_Agw6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces an anchored diffusion model for point cloud recognition and unlike the previous works that complete missing parts from the point cloud, it uses pretrained generative point cloud backbone. Also, this method did not add any extra parameters. This method has two major modules one for the local geometry reconstructor and another one for the conditional flow denoiser. This work jointly trains two modules to get the classifier. This work mainly uses three datasets including ShapeNet, ModelNet-C and ScanObjectNN-C for the classification task. And this paper uses the mean corruption error to validate the corruption robustness.

Critical strengths
  • I like how RAD only add a small thing. It is a clean method without adding more parameters.
  • The gains are good and even compared to other baselines. (But see the below…)
Critical weaknesses
  • Point-MaDi, which is the closest method, this work is cited but not benchmarked. Please specify the reason why this is discarded.
  • Table 3, using only LGR already matches the performance in Scan-C with Diffusion+LGR. This weakens the contribution of the paper of “Diffusion + Reconstruction”. And it makes me think this method is very similar to Point-MaDi, which is cited and not benchmarked.
  • I am not convinced about the “effective”. I wonder this method is better in pretraining wall-clock or FLOPs compared to PointDif.
  • Figure 3, the error difference between PointDif and Rad is very small.
  • The contribution of this paper is not backed up by the results. Per the paper’s contribution, the Drop-Local should be the largest gain in the table, but it is not the largest gain. For example, this is only a partial gain compared to the Jitter gain.
Questions for authors

I would like to see the response for #1 Weakness, missing related work and the baseline, and the #5 Weakness about the Drop-Local v.s. others.

Reviewer_Sn6gOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents RAD, a self-supervised pretraining framework for corruption-robust point cloud recognition. It jointly trains a shared encoder through masked-patch coordinate reconstruction and whole-shape noise prediction, with a linear warmup for the reconstruction loss weight. Only the encoder is transferred downstream, leaving the classifier architecture and fine-tuning protocol unchanged.

Critical strengths

RAD achieves mCE scores of 65.2 and 64.1 on ModelNet-C and ScanObjectNN-C, respectively, improving over PointDif by 8.6 and 7.5 points and outperforming the other methods in the main tables. This demonstrates the model’s practical value for point cloud recognition under corruptions. The core design motivation is clear and partially supported by the ablation studies. The paper provides representation diagnostics beyond classification metrics.4.

Critical weaknesses

Figure 1(b) states “Illustrative corruptions, not model outputs,” referring to AI-generated outputs. This is highly unrigorous for a paper submitted for publication. Figure 2 explains the Conditional Flow Denoiser, but no corresponding component can be found in the diagram. This mismatch raises concerns that the paper was rushed and was not carefully proofread. The methods used for comparison are somewhat dated. The layouts of Figures 1 and 2 are cluttered and difficult to follow. There are numerous inconsistencies in terminology. For example, Figure 2 uses “Global Condition Aggregator (CANet),” whereas the main text and appendix use “Condition Aggregation Network (CANet),” assigning two different English names to the same acronym. Similarly, the section heading and main text use “Shared Mask Encoder” or “Mask Encoder,” while Figure 2 uses “Masked Encoder fθf_\theta,” among other inconsistencies.

Questions for authors

How were the illustrative corruption examples in Figure 1(b) produced? Were they generated using AI tools, manually constructed, or obtained by applying corruptions to actual point clouds? Which components in Figure 2 constitute the Conditional Flow Denoiser? Could you clarify their correspondence with the module described in the text? How were the comparison methods selected? Are there results for more recent, closely related point cloud pretraining methods under the same evaluation protocol? Do “Global Condition Aggregator” and “Condition Aggregation Network” refer to the same module? Likewise, are “Mask Encoder” and “Masked Encoder” interchangeable, or do these names indicate architectural differences?

Reviewer_FGJ6OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper proposes a new reconstruction-anchored diffusion framework for corruption-robust point cloud recognition, named as RAD. The RAD consists of two components, which includes a Local Geometry Reconstructor and a Conditional Flow Denoiser on a shared Mask Encoder. Specifically, Local Geometry Reconstructor was designed to reconstruct masked local patches from visible encoder tokens using Chamfer Distance, and Conditional Flow Denoiser was proposed to preserve the whole shape generative denoising objective inherited from PointDif. Experiments conducted on ModelNet-C and ScannObjectNN-C demonstrate improved corruption robustness compared with previous methods.

Critical strengths
  1. The paper proposes a new framework for corruption robust point cloud recognition.
  2. The propose framework is straightforward and easy to understand.
  3. The experiments conducted on two benchmarks demonstrate the effectiveness of proposed method.
Critical weaknesses
  1. The main contribution of RAD seems like the integration of two established pretraining objectives. The Local Geometry Reconstructor closely follows previous masked point modeling approaches. Besides, the Condition Flow Denoiser is largely inherited from PointDif. Although the experimental results improved, the overall architecture looks like incremental.
  2. The paper argues that previous methods lack explicit masked patch geometric supervision and introducing such supervision improves robustness. However, the absence of local reconstruction objective does not necessarily imply that diffusion-pretrained representations fail to capture local geometric information. In fact, the original PointDif paper explicitly aims to learn both local and global geometric priors through conditional point cloud generation. Furthermore, PreDifPoint[1] (ICCV 2025) introduces mechanisms for integrating local and global features within diffusion-based point cloud pretraining.
  3. Although authors discussed Point-MAE in the related work, its performance is not directly reported in the main experimental tables. This is an important omission because the proposed LGR closely resembles the masked coordinate reconstruction paradigm.

Reference:
[1] Qiu, Chang, Feipeng Da, and Zilei Zhang. "Feature Extraction and Representation of Pre-Training Point Cloud Based on Diffusion Models." 2025 IEEE/CVF International Conference on Computer Vision (ICCV). IEEE, 2025.

Questions for authors
  1. What is the key method contribution of RAD beyond integrating PointDif-style diffusion pretraining with Point-MAE-style masked geometric reconstruction?
  2. How does the approach differ substantively from existing diffusion and masked autoencoding combinations, particularly Point-MaDi[2]?
  3. How does the proposed LGR mitigate the potential positional leakage associated with providing ground-truth masked-patch centers, as discussed in PCP-MAE[1] and Point-MaDi[2]? Can the authors provide an ablation using predicted or withheld masked centers?
  4. What is the performance comparison between RAD compare and standard Point-MAE and a naive PointDif + Point-MAE joint-pretraining baseline under identical encoder architectures and downstream protocols?
  5. PointDif aims to capture both local and global geometric priors. As discussed in the paper, what evidence supports the claim that its lack of direct masked-patch reconstruction constitutes a significant limitation in learned local geometry?

References:
[1] Zhang, Xiangdong, Shaofeng Zhang, and Junchi Yan. "Pcp-mae: Learning to predict centers for point masked autoencoders." Advances in Neural Information Processing Systems 37 (2024): 80303-80327.
[2] Xiao, Xiaoyang, et al. "Point-MaDi: Masked autoencoding with diffusion for point cloud pre-training." Advances in Neural Information Processing Systems 38 (2026): 50302-50330.

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

Related papers

Discussion (0)

Sign in to comment.