Research map / Point clouds / Point cloud classification

Asymmetric Dual Self-Distillation for 3D Self-Supervised Representation Learning

NeurIPS 2025 · Point clouds › Point cloud classification · approach: Masked point autoencoders · 0 citations (OpenAlex)

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 AsymDSD, an asymmetric dual self-distillation framework that combines latent masked point modeling with cross-view invariance learning for 3D point cloud self-supervised pretraining, reaching 90.53% on ScanObjectNN PB_T50_RS (93.72% with 930k-shape pretraining).

Problem

Masked point modeling for 3D point clouds relies on reconstruction objectives that emphasize low-level detail and can fail to capture high-level semantics; combining reconstruction with invariance objectives often interferes, and existing fusion methods are costly.

Approach

Replace reconstruction targets with latent predictions from an EMA teacher, using a joint-embedding student with an encoder-predictor design, attention disabled between mask queries to prevent shape leakage, inverse block-wise masking, multi-mask sampling, and point-cloud multi-crop, all trained with centering/sharpening on discrete latent distributions.

Key results

• AsymDSD-S achieves 90.53% on ScanObjectNN PB_T50_RS under full fine-tuning, a 5.35% gain over Point-MAE's baseline as stated in the paper's introduction.
• Pretraining on the 930k-shape Mixture dataset with ViT-B (AsymDSD-B*) reaches 93.72% on ScanObjectNN PB_T50_RS, exceeding PointGPT-L by 2.6%.
• Ablations show removing cropping drops ScanObjectNN full fine-tune accuracy from 88.72% to 82.13%, and the predictor is necessary (removing it drops FFt to 69.74%).

Measured effects reported in the paper

Surprising findings

• Multi-mask gives virtually no accuracy loss while cutting memory and throughput cost by nearly 70% relative to a single-mask batch-1024 setup.
• Without sharpening and centering the posterior collapses to uniform, while sharpening alone collapses the CLS marginal entropy to zero.

Limitations

• Pretraining and evaluation are primarily on object-centered datasets.
• The flat (non-hierarchical) encoder does not scale well to larger scene-level point clouds.

Experimental setup

Datasets: ShapeNetCore ModelNet40 ScanObjectNN ShapeNet-Part Objaverse 3D-FUTURE S3DIS SUN RGB-D OmniObject3D Toys4K

Benchmarks: ModelNet40 ScanObjectNN ModelNet40 few-shot ShapeNet-Part Objaverse-LVIS few-shot

Metrics: overall accuracy mIoU few-shot accuracy

Baselines: Point-MAE Point-BERT PointGPT ReCon Point-RAE Point-FEMAE MaskPoint Point-M2AE

Builds on: DINO Point-MAE data2vec I-JEPA multi-crop MSN

Closest papers by meaning

• DOS: Distilling Observable Softmaps of Zipfian Prototypes for Self-Supervised Point Representation (AAAI 2026 · similarity 0.87)
• Masked Scene Modeling: Narrowing the Gap Between Supervised and Self-Supervised Learning in 3D Scene Understanding (CVPR 2025 · similarity 0.86)
• SIEVE: Overcoming Topological Obstruction in Equivariant Self-Supervised Learning (NeurIPS 2026 · similarity 0.86)
• Sonata: Self-Supervised Learning of Reliable Point Representations (CVPR 2025 · similarity 0.86)
• Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.86)
• Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 2024 · similarity 0.85)