Research map / Point clouds / Point cloud classification
Towards Compact 3D Representations via Point Feature Enhancement Masked 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 Point-FEMAE, a single-modal point cloud MAE with a shared-parameter Transformer encoder over global-random and local-block masked branches plus a Local Enhancement Module (local patch convolution), pre-trained on ShapeNet.
Problem
Single-modal point cloud MAE pre-training with global random masking learns limited local detail representations, while cross-modal MAE methods are computationally heavy and need paired image/text data.
Approach
Combine a global random-mask branch and a local block-mask branch with a shared-parameter Transformer encoder, and add a Local Enhancement Module that applies KNN-based edge convolution over neighboring patch coordinates in the local branch, using only point clouds.
Key results
- Point-FEMAE reaches 95.18%, 93.29%, and 90.22% on ScanObjectNN OBJ-BG, OBJ-ONLY, and PB-T50-RS, improving on Point-MAE by 5.16, 5.00, and 5.04 points.
- On ModelNet40 Point-FEMAE scores 94.0% (w/o vote) and 94.5% (w/ vote), versus 93.2% and 93.8% for Point-MAE.
- Pre-training takes 21 hours with 41.5M parameters, about 1.6x Point-MAE's time, versus 34 hours for the cross-modal Recon method (140.9M parameters).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use Point-FEMAE instead of Point-MAE | ScanObjectNN · ScanObjectNN OBJ-BG | accuracy: +5.16 | Helped |
| use Point-FEMAE instead of Point-MAE | ScanObjectNN · ScanObjectNN OBJ-ONLY | accuracy: +5.00 | Helped |
| use Point-FEMAE instead of Point-MAE | ScanObjectNN · ScanObjectNN PB-T50-RS | accuracy: +5.04 | Helped |
| use Point-FEMAE instead of Point-MAE | ModelNet40 · ModelNet40 w/o Vote | accuracy: +0.8 | Helped |
| use Point-FEMAE instead of Point-MAE | ModelNet40 · ModelNet40 w/ Vote | accuracy: +0.7 | Helped |
| add hybrid global and local masking to Point-MAE (ablation) | Point-MAE · ScanObjectNN OBJ-BG | accuracy: 90.02 -> 89.67 | Hurt |
| add hybrid global and local masking to Point-MAE (ablation) | Point-MAE · ScanObjectNN OBJ-ONLY | accuracy: 88.29 -> 88.30 | No clear effect |
| add LEM to Point-MAE (ablation) | Point-MAE with global random mask · ScanObjectNN OBJ-ONLY | accuracy: 88.29 -> 89.33 | Helped |
| add LEM to Point-MAE (ablation) | Point-MAE with global random mask · ScanObjectNN OBJ-BG | accuracy: 90.02 -> 90.36 | Helped |
| use simple rotate augmentation instead of the baseline augmentation (ablation) | ScanObjectNN models · ScanObjectNN OBJ-BG | accuracy: 90.02 -> 92.94 | Helped |
| reduce LEM MLP scale from 1 to 1/16 (ablation) | Point-FEMAE · ScanObjectNN OBJ-BG | accuracy: 95.18 -> 94.84 | Hurt |
| reduce LEM MLP scale from 1 to 1/16 (ablation) | Point-FEMAE · ScanObjectNN OBJ-ONLY | accuracy: 93.29 -> 93.12 | Hurt |
Surprising findings
- Adding a parameter-matched MLP or self-attention module instead of LEM gives much smaller gains, suggesting the gain is not from added parameters alone.
- Fully shared encoder across global and local branches performs worse than the proposed partial sharing, attributed to conflicting reconstruction targets.
Limitations
- Reported ModelNet40 gains are small (0.7-0.8 points), and the method is roughly on par with Recon on ScanObjectNN and ModelNet40.
- Parameter count (27.4M) and pre-training time are higher than Point-MAE, though the authors call the increase negligible.
- Pre-training only on ShapeNet; no evaluation on scene or LiDAR data.
Experimental setup
Datasets: ShapeNet ScanObjectNN ModelNet40 ShapeNetPart
Benchmarks: ScanObjectNN ModelNet40 ShapeNetPart
Metrics: overall accuracy few-shot accuracy mIoU Chamfer distance
Baselines: Point-MAE Point-M2AE ACT I2P-MAE Joint-MAE Recon Point-BERT MaskPoint
Builds on: Point-MAE PointNet EdgeConv Chamfer distance reconstruction
Closest papers by meaning
- Point-M2AE: Multi-scale Masked Autoencoders for Hierarchical Point Cloud Pre-training (NeurIPS 2022 · similarity 0.93)
- Learning 3D Representations From 2D Pre-Trained Models via Image-to-Point Masked Autoencoders (CVPR 2023 · similarity 0.92)
- 3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining (ICLR 2024 · similarity 0.91)
- PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024 · similarity 0.89)
- RI-MAE: Rotation-Invariant Masked AutoEncoders for Self-Supervised Point Cloud Representation Learning (AAAI 2025 · similarity 0.88)
- Multi-Scale Neighborhood Occupancy Masked Autoencoder for Self-Supervised Learning in LiDAR Point Clouds (CVPR 2025 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.