acceptodds
Withdrawn from ICLR 2027

Self-Supervised Point Cloud Dataset Distillation for Efficient Representation Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Dataset distillation (DD) aims to replace a large training dataset with a compact synthetic set that preserves its training utility, substantially reducing repeated storage and training costs. However, existing point-cloud DD methods are developed under supervised settings and rely on semantic annotations to construct distilled data. While self-supervised learning (SSL) makes large unlabeled point-cloud collections usable without annotation, repeatedly training on the full dataset remains computationally expensive. Moreover, a pretrained checkpoint merely provides model-specific rather than data-level reusability. This motivates a fundamental question: can the representation-learning utility of a massive unlabeled point-cloud dataset be distilled into a small, reusable synthetic set? We study self-supervised dataset distillation for point clouds and propose a distillation framework for constructing compact training sets from unlabeled 3D data. Rather than relying on semantic classes or discrete pseudo-labels, we formulate distillation around self-supervised representation learning and optimize synthetic point clouds according to their ability to train newly initialized encoders. Specifically, our proposed framework leverages masked predictive learning to preserve informative global and local structures under a highly constrained synthetic budget. We evaluate the distilled synthetic sets on multiple standard point-cloud benchmarks via frozen representation evaluation against equally sized real-subset controls. Extensive experiments demonstrate that the optimized synthetic data yields non-trivial representation-learning utility compared to equally sized real subsets across multiple compression regimes. These results demonstrate the potential of self-supervised point-cloud dataset distillation as a data-level mechanism for compressing and reusing large unlabeled 3D collections.

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

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_vY2nOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MPD-3D, which is a self-supervised dataset distillation method for 3D point clouds. The method uses a masked autoencoder for 3D point clouds to train a teacher network (e.g., PointNet) and extracts target features. The initial synthetic point clouds are sampled from the original point clouds and the points of synthetic ones are optimized so that a student network trained on them can predict the target feature. Experimental results on MN40, MVP100FULL, and SONN show promising gain over previous coreset selection methods.

Critical strengths
  1. The paper first formulates the problem of self-supervised dataset distillation for 3D point clouds.
  2. The paper reports extensive experimental results such as cross-architecture generalization results and the effect of coordinate optimization.
Critical weaknesses
  1. Unfair comparison. In table1, previous subset selection methods (Random, Herding, and K-center) are trained by supervised learning, while MPD-3D is trained by self-supervised learning. It is hard to say that the performance gains over previous subset selection methods are from the proposed method.

Q1. Please report the comparison with Random, Herding, and K-Center which is trained by self-supervised learning (masked prediction with the teacher targets).

  1. Unclear effectiveness of dataset distillation. In Table 18, the effectiveness of coordinate refinement is unclear. It implies that the performance gains in Table 1 come from the existing masked predictive learning, not from the dataset distillation itself.

Q2. Please justify the benefit of the coordinate refinement itself.

Q3. In the distillation, the student network is trained for only 3 iterations, while the consumer network is trained for far more iterations. Please show the learning curve during training.

Questions for authors

See the Q1-Q3 in Critial Weaknesses.

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

The paper studies label-free pointcloud dataset distillation through a compact predictive training resource. The proposed method, MPD-3D constructs a compact point-cloud training resource containing synthetic coordinates, fixed teacher-derived representation targets, and masking metadata. Coordinates are optimized by differentiating real-query prediction loss through short learner updates. Newly initialized consumers then learn from this resource without accessing the teacher or full source data.

Critical strengths
  1. The method provides a data-level interface for reusing unlabeled collections across newly initialized models
Critical weaknesses
  1. Unclear value of coordinate optimization: Table 18 shows the optimized coordinates improve LP@10% in only four of nine settings. And most of the improvements seem minor.

  2. Cross-architecture experiments lack matched unoptimized-support controls. Sensitivity to the producer architecture and inner-loop length is insufficiently characterized.

  3. Efficiency remains incompletely established: Storage reduction largely follows from retaining fewer objects. Quality-matched comparisons against predictive real subsets and complete production-cost accounting are needed to justify the additional distillation process.

  4. Main comparisons confound data and training objectives: Real-subset baselines use class-label supervision, whereas MPD-3D uses masked teacher-target regression. A matched full-data predictive baseline is also missing, preventing assessment of how much full-data utility is preserved.

Questions for authors
  1. How do compact resources compare with both real subsets and the full dataset under the same predictive objective and evaluation protocol?
  2. Do coordinate gains persist across consumer architectures, producer learners, and inner-loop lengths? Please include matched unoptimized-support controls.
  3. At equal representation quality, what storage and total computational savings remain after including targets, metadata, teacher training, and resource production?
  4. Given the mixed coordinate-refinement results, when does optimizing geometry provide a reliable advantage over exporting real supports with identical targets and masks?
Reviewer_pYUmOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces MPD-3D, a self-supervised method for 3D dataset distillation. It uses a pretrained teacher to generate feature targets and optimizes synthetic point clouds through a masked prediction objective. The distilled data can then train new models without needing the teacher or original dataset. Experiments across three datasets and different architectures show improvements over supervised real-subset baselines.

Critical strengths
  • Promising idea: The method provides a compact, reusable dataset for training new encoders without requiring teacher access.

  • Good evaluation and transparent reporting: The authors test different datasets, budgets, label settings, and architectures. particularly controlled comparison and the honest reporting of negative results, storage costs, and evaluation limitations.

Critical weaknesses
  • Unclear benefits of coordinate optimization: Table 1 compares different training pipelines, making it difficult to isolate the effect of distillation. The more controlled comparison in Table 18 shows improvements in only 4/9 settings for LP@10 and 3/9 for kNN@10, with worse LP@10 results across all three ScanObjectNN settings. This raises questions about the actual benefit of coordinate optimization and I think without it, the method largely becomes a real-data subset with cached teacher features.

  • Unfair baseline comparison: The proposed method benefits from a teacher trained on the full dataset, while the real-subset baselines use supervised training with very limited labeled data. More appropriate comparisons, such as self-supervised training or teacher distillation on the same real subsets, are missing. Some results also raise concerns about the evaluation: the teacher performs much worse than its students in linear probing, and the reported LP scores are unusually low compared with kNN results and prior work (e.g., OcCo). These issues make the reported improvements difficult to interpret.

  • Presentation: Appendix A is a template leftover ("You may include other additional sections here"). Also some figures in the appendix seem to be duplicated.

Questions for authors
  • Under the same evaluation protocol as Table 1, could the authors report LP and kNN results at 10% and 100% labels for an untrained PointNet, a fully supervised PointNet, and the frozen teacher? Also, why do students trained to match projected teacher features outperform the teacher's LP by over 20 points? These comparisons are important for understanding whether the gains come from distillation or simply from teacher supervision.

  • How much information is retained from full-data training? Could the authors compare their method with a consumer trained on the full dataset using the same objective, as well as an untrained encoder, under identical evaluation settings?

  • Is the 100-iteration L-BFGS probe converged? How sensitive are LP values to the regularization weight and to feature normalization, and why do LP and kNN disagree on the ranking with the full labeled bank?

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

Related papers

Discussion (0)

Sign in to comment.