acceptodds
Under review as a conference paper at ICLR 2027

ShapeDINO: Scaling Self-Supervised Representation Learning for 3D Shapes

Abstract

Large-scale self-supervised learning has enabled general-purpose representations in language and vision. We investigate how far this paradigm can advance reusable representations of 3D shapes. We introduce ShapeDINO, a family of self-supervised point cloud encoders built on a standard vision Transformer backbone, ranging from Tiny to a billion-parameter Giant. ShapeDINO combines global and patch-level self-supervised learning to jointly learn object-level semantics and dense local features. To support pretraining at scale, we curate a dataset of 1.5 million 3D assets through quality filtering, geometric deduplication, and semantic rebalancing. We further distill the largest model into smaller variants to accommodate different computational budgets. Under frozen-feature evaluation, ShapeDINO achieves state-of-the-art performance on multiple classification and part-segmentation benchmarks, matching or surpassing the evaluated weakly supervised baselines. Additionally, we demonstrate that ShapeDINO can be seamlessly integrated into downstream tasks such as representation alignment, point segmentation and completion, yielding consistent performance improvements. These results demonstrate that large-scale self-supervised learning can yield reusable 3D representations that benefit diverse downstream tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
3d representation learning, 3d foundation model

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_ZZ4QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ShapeDINO, a family of large-scale self-supervised encoders for object-centric 3D shapes. The method uses a PointNet-style tokenizer with FPS/KNN grouping followed by a standard Vision Transformer with 3D RoPE, and adopts a DINO-style teacher–student framework that combines global DINO supervision with patch-level iBOT learning to capture both object-level semantics and dense local structure. To scale training efficiently, the authors introduce a 3D multi-crop strategy that reuses a shared spatial grouping across crops, and construct a curated 1.5M-object pretraining corpus from Objaverse-XL through geometric filtering, deduplication, and feature-space rebalancing. The model family ranges from 6.9M to 1.14B parameters, with smaller variants distilled from the Giant model. Experiments show strong frozen-feature performance on classification and part-segmentation benchmarks, as well as transfer gains in 3D generation, promptable segmentation, and point-cloud completion. Overall, the paper argues that a carefully engineered DINO-style self-supervised recipe can be scaled effectively to produce strong and reusable object-level 3D representations

Critical strengths
  1. Strong engineering execution and comprehensive empirical validation. ShapeDINO presents a complete large-scale 3D self-supervised pretraining pipeline, including efficient point-cloud multi-crop construction, data filtering/deduplication/rebalancing, model scaling up to 1B parameters, and distillation into smaller variants. The evaluation is also broad: beyond frozen kNN/linear probing on classification and part segmentation, the learned representations are transferred to 3D generation, promptable segmentation, and point-cloud completion. This breadth makes the empirical contribution substantial even though the individual methodological components are largely based on prior work.

  2. The learned representations are strong at both global and dense local levels. ShapeDINO achieves competitive frozen-feature performance on both object classification and part segmentation, with particularly strong dense-feature results under simple point-wise linear probing. The downstream experiments further show that the pretrained features are not only benchmark-specific but can improve different 3D tasks, including Point-SAM segmentation and point-cloud completion. This provides convincing evidence that the representation itself is reusable and practically useful.

Critical weaknesses
  1. Limited methodological novelty and insufficient comparison with closely related prior work.
    Most of ShapeDINO’s core ingredients have clear precedents, including the point tokenizer + ViT architecture, DINO/iBOT/KoLeo objectives, EMA teacher–student training, 3D RoPE, knowledge distillation, and feature-space-based data rebalancing. Closely related works such as PointCAM, AsymDSD already explore DINO-style/self-distillation-based representation learning for object-centric point clouds, including multi-crop training and large-scale pretraining. The current paper therefore appears primarily to scale and integrate an established recipe not introduce a fundamentally new representation-learning principle. A direct comparison with these closely related methods is important for establishing the paper’s novelty.

  2. The central scaling claim is confounded by distillation and data curation.
    The model-scaling results do not provide a clean from-scratch scaling curve: the Giant model is pretrained from scratch, whereas the smaller variants are largely distilled from the Giant teacher. Since the paper itself shows that distillation provides non-negligible gains, improvements across model sizes cannot be attributed solely to parameter scaling. Similarly, the data ablation suggests that curation/rebalancing may be at least as important as raw scale: on ScanObjectNN, balanced 100K already outperforms raw 1.5M, while increasing balanced data from 100K to 1.5M provides only a modest gain (table 3). Consequently, the experiments more clearly establish the effectiveness of the overall training recipe than model/data scaling itself.

  3. The data contribution is primarily curation of an existing dataset not a new dataset contribution.

Questions for authors
  1. Could the authors clarify the methodological distinction between ShapeDINO and closely related prior work such as PointCAM (Self-supervised Adversarial Masking for 3D Point Cloud Representation Learning) and AsymDSD (Asymmetric Dual Self-Distillation for 3D Self-Supervised Representation Learning), and ideally provide a direct comparison under a matched evaluation protocol? From my current understanding, the main differences of ShapeDINO appear to be a more DINOv2-like recipe, larger-scale training, and several engineering improvements. A clearer comparison would help establish what is genuinely new beyond scaling and integration.

  2. Could the authors provide a cleaner from-scratch scaling study that disentangles model size from distillation, and data scale from curation? The current model-scaling curve mixes scratch-trained Giant with smaller models distilled from the Giant teacher, while the data ablation suggests that balancing/curation contributes substantially relative to raw data scale. Since “scaling” is central to the paper’s framing, a matched from-scratch comparison across several model sizes, or a more explicit decomposition of these effects, would substantially strengthen the main claim.

Even though I challenge the paper’s framing somewhat reframing of established ideas and the scaling claims overstated, I still consider the engineering effort, empirical coverage, and resulting representations strong, and I hope authors can address my concerns during rebuttal.

Reviewer_tTB9OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces ShapeDINO, a family of self-supervised 3D shape encoders that scale from 6.9M to 1.1B parameters. Building on DINOv2-style self-distillation, the method combines global DINO supervision with patch-level iBOT supervision, and introduces an efficient multi-crop construction that reuses spatial grouping across crops to reduce the repeated cost of point-cloud patch construction. The authors also curate a 1.5M-object pretraining corpus from Objaverse-XL through geometric filtering, deduplication, and feature-space rebalancing, and distill the largest model into smaller variants. ShapeDINO is evaluated through frozen-feature classification and part segmentation, model and data scaling studies, and downstream transfer to 3D generation, promptable segmentation, and point cloud completion, where it shows strong performance across the evaluated settings.

Critical strengths
  1. The paper presents a comprehensive and well-executed scaling study of self-supervised representation learning for 3D shapes. ShapeDINO scales from 6.9M to 1.14B parameters and demonstrates consistent improvements in both global and dense local representations, providing useful evidence that DINO-style self-supervised learning can benefit from model scaling in the 3D domain.
  2. The paper presents a comprehensive and well-executed scaling study of self-supervised representation learning for 3D shapes. ShapeDINO scales from 6.9M to 1.14B parameters and demonstrates consistent improvements in both global and dense local representations, providing useful evidence that DINO-style self-supervised learning can benefit from model scaling in the 3D domain.
  3. The experimental evaluation is extensive and covers multiple aspects of representation quality and transferability. In addition to frozen-feature classification and part segmentation, the paper studies model scaling, data curation, and distillation, and further evaluates transfer to 3D generation, promptable segmentation, and point-cloud completion. The consistently strong performance across these settings provides broad empirical support for the usefulness of the learned representations.
Critical weaknesses
  1. The interpretation of the scaling and model-size comparisons is somewhat confounded by knowledge distillation. ShapeDINO-Giant is pretrained from scratch, whereas the smaller variants are distilled from the 1.14B Giant teacher. As a result, both the Tiny-to-Giant scaling trend and comparisons with similarly sized baselines reflect not only model capacity, but also access to a much larger pretrained teacher and different pretraining resources. Controlled from-scratch experiments at multiple model scales would help isolate the effect of model capacity.
  2. The efficiency benefit of the proposed shared-grouping multi-crop strategy is not directly quantified. The method is motivated by avoiding repeated FPS/KNN operations for independently constructed crops, but the paper does not provide a direct runtime, memory, or computational-cost comparison against a naive multi-crop implementation with independent grouping. Since this is one of the main 3D-specific technical contributions, such a comparison would better substantiate its practical advantage.
Questions for authors
  1. Could the authors provide controlled from-scratch results for several ShapeDINO model scales to better disentangle the effect of model capacity from knowledge distillation?
  2. Could the authors quantify the efficiency gain of the shared-grouping multi-crop strategy compared with independently constructing point-cloud crops?
  3. Can the authors provide further analysis of why ShapeDINO-based REPA improves over the baseline while Utonia-based REPA remains below it in the 3D generation experiment?
Reviewer_SxJmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

ShapeDINO is a family of self-supervised encoders for object-centric point clouds, ranging from 6.9M to 1.14B parameters. It combines global DINO and local iBOT objectives and constructs multi-scale crops through shared geometric grouping. The Giant model is pretrained on 1.5 million 3D assets and distilled into smaller variants. Experiments cover frozen-feature classification, part segmentation, and several downstream transfer tasks. Giant achieves 84.1% linear-probing accuracy on ScanObjectNN.

Critical strengths

DINO and iBOT supervise object-level and local representations, respectively, supporting both classification and dense prediction in a general-purpose shape encoder.

Frozen-feature performance is strong, and smaller models retain practical utility. Giant performs well on ScanObjectNN and both part-segmentation benchmarks.

Critical weaknesses

Giant is pretrained from scratch, while the other main variants are distilled from Giant. The scaling curve in Table 2 therefore combines different training regimes and does not isolate the effect of model capacity.

The contributions should be better summarized. The statement "We systematically evaluate model scaling and distillation, and demonstrate transferability across frozen-feature benchmarks and diverse downstream applications." cannot be considered as a contribution. Conducting experiments is a means to verify the effectiveness of your technical innovations, not an end in itself.

Cross-deduplication between pretraining assets and downstream test sets is not described. Appendix B details deduplication within the training corpus but does not report overlap checks with ModelNet40, ShapeNetPart, PartNet, PCN, or the completion test assets.

Shared geometric grouping avoids repeated FPS and neighborhood searches, but the paper does not report throughput, memory usage, preprocessing time, or end-to-end speedups relative to constructing each crop independently.

Table 3 compares Raw 1.5M and Balance 1.5M without clarifying whether they are drawn from the same quality-filtered and deduplicated pool. It is therefore difficult to separate the contributions of filtering, deduplication, and rebalancing.

Table 1 reports PartNetE results of 38.4 and 39.3 for Point-MAE and Point-M2AE, respectively, while Table 10 reports 44.1 and 44.9. This is a discrepancy of approximately 5–6 points.

Questions for authors

Please explain the approximately 5–6-point discrepancy between Tables 1 and 10 in the PartNetE results for Point-MAE and Point-M2AE.

Was cross-deduplication performed between the pretraining corpus and ModelNet40, ShapeNetPart, PartNet, PCN, and the completion test sets?

Could you clarify whether the scaling curve in Table 2 combines the Giant model pretrained from scratch with smaller models obtained through distillation?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.