Research map / Point clouds / Point cloud classification

SIEVE: Overcoming Topological Obstruction in Equivariant Self-Supervised Learning

NeurIPS 2026 · Point clouds › Point cloud classification · approach: Masked point autoencoders

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes SIEVE, an SE(3)-equivariant masked-autoencoding framework for point clouds that uses a spherical projection mask (instead of constant/noise masks) to avoid a claimed topological collapse, and yields frozen invariant scalar embeddings that improve downstream point cloud tasks.

Problem

Standard masking (constant or random tokens) in masked autoencoding is argued to be incompatible with SE(3)-equivariant networks on point clouds, causing representational collapse; the paper seeks an equivariant SSL scheme producing transferable invariant embeddings.

Approach

A spherical projection mask that projects each point onto a unit sphere centered at the point cloud mean, preserving global pose while obscuring local semantics, combined with an equivariant GATr encoder that bottlenecks to invariant scalar embeddings; justified via a vector-bundle/pullback theory argument.

Key results

• On 3D-EPN unseen categories, PCN+SIEVE reduces average Chamfer distance from 11.98 to 10.49 (x10^4), about 12.5% reduction vs 1.6% on seen categories.
• RNA segmentation: SIEVE improves PointNet++ accuracy from 71.55% (xyz) to 86.28%, DeltaNet 87.21% to 89.58%, DiffusionNet 81.20% to 89.30%.
• On the synthetic cuboid/toroid collapse experiment, spherical mask reaches validation L1 below 0.01 while constant and ellipsoid noise masks stagnate around 0.5 loss.

Surprising findings

• Constant and ellipsoid noise masks fail to reduce even training loss on the simple synthetic topology task, which the authors attribute to representational collapse.

Limitations

• Theoretical guarantees rely on vector bundle approximation arguments that do not cover all non-trivial bundles (Appendix A.5).
• Spherical projection mask requires Gaussian noise injection to avoid trivial radius shortcut; proof steps are informal and some arguments are heuristic.
• Downstream gains are shown on a small set of tasks with generalist backbones and no comparison against SOTA task-specific models.

Experimental setup

Datasets: OpenShape ShapeNet synthetic cuboid/toroid dataset RNA molecular segmentation dataset 3D-EPN ModelNet40

Benchmarks: RNA molecular segmentation 3D-EPN shape completion ModelNet40 registration

Metrics: test accuracy Chamfer distance MSE RMSE MAE L1 reconstruction loss

Baselines: PCN DCP PointNet++ DeltaNet DiffusionNet SE(3)-Transformer constant vector mask ellipsoid noise mask

Builds on: Masked Autoencoders GATr Point-BERT DCP

Closest papers by meaning

• RI-MAE: Rotation-Invariant Masked AutoEncoders for Self-Supervised Point Cloud Representation Learning (AAAI 2025 · similarity 0.88)
• Adaptive Siamese Masked Autoencoder with Global Optimization for Unsupervised Point Cloud Shape Correspondence (AAAI 2025 · similarity 0.87)
• 3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining (ICLR 2024 · similarity 0.87)
• Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.86)
• Neural Isometries: Taming Transformations for Equivariant ML (NeurIPS 2024 · similarity 0.86)
• Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 2024 · similarity 0.86)