Research map / Point clouds / Point cloud classification
Learning 3D Representations From 2D Pre-Trained Models via Image-to-Point Masked Autoencoders
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add I2P-MAE image-to-point pre-training rather than use Point-M2AE | 3D point-cloud representation learning · ModelNet40, linear SVM | accuracy: 92.9 -> 93.4 | Helped |
| Add I2P-MAE image-to-point pre-training rather than use Point-M2AE | 3D point-cloud representation learning · ScanObjectNN PB-T50-RS | accuracy: +3.68% | Helped |
| Add 2D-guided masking that preserves important tokens instead of random masking (ablation) | I2P-MAE · ModelNet40 and ScanObjectNN OBJ-BG, linear SVM | accuracy: +0.4% and +0.9% | Helped |
| Mask important tokens and preserve unimportant tokens (ablation) | I2P-MAE · ModelNet40 and ScanObjectNN OBJ-BG, linear SVM | accuracy: -0.9% and -3.3% | Hurt |
| Change masking ratio from 0.8 to 0.7 (ablation) | I2P-MAE · ModelNet40, linear SVM | accuracy: 93.4 -> 92.5 | Hurt |
| Change masking ratio from 0.8 to 0.9 (ablation) | I2P-MAE · ModelNet40, linear SVM | accuracy: 93.4 -> 92.8 | Hurt |
| Add 2D-semantic reconstruction from visible tokens (ablation) | I2P-MAE · ModelNet40 and ScanObjectNN OBJ-BG, linear SVM | accuracy: +0.5% and +2.4% | Helped |
| Remove masked-token 3D-coordinate reconstruction (ablation) | I2P-MAE · ModelNet40 and ScanObjectNN OBJ-BG, linear SVM | accuracy: 93.4 -> 91.9; 87.1 -> 80.2 | Hurt |
| Remove 2D-semantic targets and reconstruct only masked-token 3D coordinates (ablation) | I2P-MAE · ModelNet40 and ScanObjectNN OBJ-BG, linear SVM | accuracy: 93.4 -> 91.2; 87.1 -> 77.8 | Hurt |
| Reconstruct both 3D and 2D targets from masked tokens instead of separating visible and masked targets (ablation) | I2P-MAE · ModelNet40 and ScanObjectNN OBJ-BG, linear SVM | accuracy: 93.4 -> 92.6; 87.1 -> 84.9 | Hurt |
| Remove 2D guidance during pre-training (ablation) | I2P-MAE · ScanObjectNN PB-T50-RS, fine-tuning | accuracy: 87.52 -> 90.11 | Hurt |
| Use fewer projected views than the default three (ablation) | I2P-MAE · ModelNet40, linear SVM | accuracy: 93.4 -> 93.0; 93.4 -> 92.9 | Hurt |
Closest papers by meaning
- Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 2024 · similarity 0.92)
- 3D Feature Prediction for Masked-AutoEncoder-Based Point Cloud Pretraining (ICLR 2024 · similarity 0.90)
- Masked Scene Modeling: Narrowing the Gap Between Supervised and Self-Supervised Learning in 3D Scene Understanding (CVPR 2025 · similarity 0.88)
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.87)
- RI-MAE: Rotation-Invariant Masked AutoEncoders for Self-Supervised Point Cloud Representation Learning (AAAI 2025 · similarity 0.87)
- PCP-MAE: Learning to Predict Centers for Point Masked Autoencoders (NeurIPS 2024 · similarity 0.87)
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.