acceptodds
Under review as a conference paper at ICLR 2027

UPCDD: Unsupervised Point Cloud Dataset Distillation

Abstract

Dataset distillation (DD) aims to condense a large real dataset into a compact synthetic set that preserves its training utility while reducing storage and repeated training costs. While recent approaches have extended DD to 3D point clouds, they predominantly rely on ground-truth semantic annotations, which are difficult and expensive to acquire in the 3D domain. This dependence prevents unlabeled point-cloud data from being directly distilled and is particularly restrictive for tasks requiring costly fine-grained annotations, such as part segmentation. To address this limitation, we propose UPCDD (Unsupervised Point Cloud Dataset Distillation). To the best of our knowledge, UPCDD is the first framework to enable label-free dataset distillation for point clouds. UPCDD leverages self-supervised representation learning and adaptive feature-space clustering to discover latent semantic partitions, effectively substituting manual labels with proxy supervision. UPCDD can be instantiated by adapting existing label-dependent 3D distillation methods, while a novel augmented objective is designed to stabilize label-free distillation, facilitating effective adaptation and transfer. Experiments on 3D shape classification and part segmentation demonstrate that UPCDD retains useful semantic information under severe compression and transfers across distillation methods, architectures, and datasets. Moreover, the distilled dataset can be reused for downstream training without repeatedly accessing the full unlabeled collection, providing a compact and architecture-flexible alternative for exploiting large unlabeled point-cloud datasets.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Dataset Distillation, Unsupervised Learning, Self-Supervised Learning, Point Clouds, Representation Learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_np1JOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces UPCDD, a label-free dataset distillation for object-wise point clouds. First an encoder is trained via masked-reconstruction, to later on supply features for the adaptive clustering. The adaptive clustering is done via k-means and is used to partition the dataset into subsets, to provide similar partition to how labeled dataset can be split into object categories. The resulting partition indices and confidence weights replace the semantic labels in DD3D- and SADM-based distillation, so reusing these approaches becomes trivial. Feature alignment and dispersion regularization augment the matching objective used during the distillation process. The paper evaluates classification, part segmentation, and transfer across datasets and architectures.

Critical strengths
  • The problem setting is well motivated: producing reusable synthetic training data without requiring shape or point annotations is an extension of point-cloud distillation. The formulation makes the role of proxy supervision explicit and is instantiated with both gradient and distribution matching. Extending the construction to pseudo-parts is also relevant to the annotation-cost motivation.

  • There is evidence that the distilled sets carry does carry useful training information. Table 2 shows improvements over a frozen randomly initialized backbone. On ScanObjectNN at m=10, UPCDD-SADM reaches 42.99 percent versus 39.90 percent for Best PL-Core with the same reported K=5. The supervised experiments reported in table 4 also provide evidence that the proposed augmented objective can help existing distillation pipelines. Evaluation with fresh downstream networks, frozen linear probes, and label mappings fitted on the training split provides useful controls that strengthen these results.

Critical weaknesses
  • Incomplete reporting makes the compression and comparative results difficult to assess. Table 1 omits the selected K*, although the total distilled-set size is directly determined by it, preventing readers from determining the compression achieved by each configuration. Furthermore, Table 2 compares unequal total budgets, eg: on MVPNet100 at m=10, UPCDD-SADM uses 1,050 clouds versus 510 for the core-set. This does not establish an advantage at fixed storage. Network roles, dataset construction, and evaluation details also require piecing together the main text and appendix. No qualitative examples of the final synthetic clouds are provided, limiting inspection and intuition of their structure and diversity.

  • The practical motivation and contribution beyond existing pseudo-labeling approaches remain insufficiently established. The evaluated setting assumes already isolated, object-centric point clouds. This object-level scope should be clearer in the opening motivation. A relevant practical alternative is to obtain pseudo-labels from an external pretrained model and then apply an existing distillation method. A comparison at equal synthetic-data budgets, or an explanation of why such models are unsuitable for the intended setting could provide a stronger motivation. Moreover, replacing class labels with feature clusters for dataset distillation has precedent in 2D. Although the domain is different, the paper should identify which challenges specific to point clouds require new solutions and how its components address them. As presented, the methodological advance beyond adapting and combining existing ideas remains unclear.

Questions for authors
  • What K* was selected for each Table 1 configuration? would be great to know the total number of synthetic clouds and its ratio to the original training-set size.

  • Are there any representative examples from the final synthetic datasets, alongside real point clouds from the corresponding partitions, to illustrate the geometry and diversity retained through distillation?

  • How does UPCDD compare with using an external pretrained model to obtain pseudo-labels, followed by the existing distillation method? Are there any technical challenges specific to object-centric point clouds that required changes beyond the clustering-based strategy done in the 2D domain?

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

This paper presents UPCDD (Unsupervised Point Cloud Dataset Distillation) to enable label-free dataset distillation for 3D point clouds. It uses pseudo-partitions derived from self-supervised representations as proxy supervision for distillation. Evaluations have been performed on two 3D point-cloud tasks, namely shape classification and part segmentation. Shape classification was evaluated on ScanObjectNN , ModelNet40, and MVPNet100, and part segmentation on ShapeNet-Part.

Critical strengths
  • The paper is well-written and easy to follow.
  • Validation is comprehensive although the tables provide a mixed picture.
Critical weaknesses
  • It is mentioned in the paper that fine-grained annotations are costly for part segmentation. Yet, the results presented in Table 5 for part segmentation are not very strong. No standard deviation is included, and the gain with fine-tuning over training from scratch is very small. In addition, linear and fine-tune results barely change from 1 to 10 samples per partition.

  • In Table 2, K is different for different columns/methods, with UPCDD-SADM having the most samples. So, it is not clear if UPCDD-SADM is the best or is this result is due to storing more.

  • Since no upper bound is listed, one cannot figure out how much is lost by distilling. What would be the performance of an encoder pretrained with self-supervised learning on the full unlabeled dataset?

Questions for authors

Please see the points listed under Weaknesses.

1 comment

W3: Full-Data and self-supervised References

We actually have conducted these experiments and the results are already shown in Appendix A.5 (Table 13 ) and full-data supervised upper bounds in (Appendix A.4, Table 12). We now list the two references here: (1) full-data supervised PointNet trained from scratch using all labeled training samples (original in Table 12), and (2) the frozen Phase-1 encoder pretrained on the full unlabeled training set (original Table 13).

DatasetFull-data supervised OA (DD3D / SADM)Full-data SSL LP@50Compression ratio (DD3D / SADM)
ScanObjectNN65.22 / 63.2744.18±1.603.03% / 2.17%
ModelNet4087.76 / 84.7079.57±0.942.24% / 3.15%
MVPNet10071.29 / 67.8942.00±0.142.05% / 3.59%

The compression ratios further quantify the storage reduction achieved by UPCDD under CPP=10.

Please let us know if any concerns remain unaddressed. We are happy to discuss them.

Reviewer_kW8rOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper first proposed an unsupervised dataset distillation method for 3D point clouds. The proposed method obtains pseudo latent partitions by self-supervised representation learning for proxy supervision. Also, it designs an augmentation objective to stabilize label-free distillation and enforce semantic consistency. Experiments are performed on 3D shape classification and part segmentation tasks to demonstrate the usability of the proposed method.

Critical strengths

Unsupervised setting of dataset distillation for 3D point clouds is a novel research topic.

Critical weaknesses
  • Limited novelty: The proposed method seems to be a straightforward extension of the existing frameworks of 2D image dataset distillation into 3D point cloud domain by simply combining self-supervised representation learning and pseudo-label-based distillation.

  • Insufficient validation: The specific modules in the proposed method were not designed dedicated to 3D point clouds. It seems to be applicable to 2D image domain also. However, the paper does not provide experimental results demonstrating the benefits of the proposed method when applied to 2D image dataset distillation.

Questions for authors
  • What is the meaning of “We use the aligned setting” in the experimental setup?

  • The authors show the sensitivity to latent partition granularity in Table 7. How do we determine optimal numbers of partitions (K) for different datasets?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.