Research map / Point clouds / Point cloud classification
Multi-modal Relation Distillation for Unified 3D Representation Learning
Huiqun Wang, Yiping Bao, Panwang Pan, Zeming Li, Xiao Liu, Ruijie Yang, Di Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (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 relation distillation with normalized similarity (intra-modal) (ablation) | contrastive 3D-image-text alignment · ModelNet40 zero-shot | Top-1 accuracy (%): 72.2 -> 73.1 | Improved |
| use partial-order relation representation (intra-modal) instead of normalized similarity (ablation) | relation distillation · Objaverse-LVIS zero-shot | Top-1 accuracy (%): 11.0 -> 10.0 | Hurt |
| add both intra- and cross-modal distillation without dynamic weights (ablation) | MRD · ScanObjectNN zero-shot | Top-1 accuracy (%): 54.6 -> 53.5 (vs intra only) | Mixed |
| add dynamic distillation weights (learnable softmax weights) (ablation) | MRD with intra and cross relations · ModelNet40 zero-shot | Top-1 accuracy (%): 73.0 -> 74.2 | Improved |
| set loss weight lambda to 3 versus 5 (ablation) | MRD · ModelNet40 zero-shot | Top-1 accuracy (%): 72.4 -> 74.2 | Improved |
| scale PointBERT backbone from 5M to 307M with MRD (ablation) | MRD · Objaverse-LVIS zero-shot | Top-1 accuracy (%): 47.6 -> 53.6 | Improved |
| scale MRD-B to MRD-L on ModelNet40 (ablation) | MRD · ModelNet40 zero-shot | Top-1 accuracy (%): 88.8 -> 87.3 | Hurt |
| pretrain with MRD instead of ULIP before fine-tuning | PointBERT-22M · ScanObjectNN fine-tuning | Instance accuracy (%): 86.4 -> 89.6 | Improved |
| use MRD instead of OpenShape/Uni3D for text-to-3D retrieval | Cap3D to Objaverse retrieval · instance-wise retrieval | Top-1 accuracy (%): 18.9 -> 20.5 | Improved |
| use ULIP2 image-text data instead of default data for MRD training (ablation) | MRD-T and MRD-S on ShapeNet · Objaverse zero-shot | accuracy | Improved |
Closest papers by meaning
- TAMM: TriAdapter Multi-Modal Learning for 3D Shape Understanding (CVPR 2024 · similarity 0.87)
- Beyond Modality Collapse: Representation Blending for Multimodal Dataset Distillation (NeurIPS 2025 · similarity 0.87)
- L4D-Track: Language-to-4D Modeling Towards 6-DoF Tracking and Shape Reconstruction in 3D Point Cloud Stream (CVPR 2024 · similarity 0.87)
- MM-Mixing: Multi-Modal Mixing Alignment for 3D Understanding (AAAI 2025 · similarity 0.86)
- ULIP-2: Towards Scalable Multimodal Pre-training for 3D Understanding (CVPR 2024 · similarity 0.86)
- MMScan: A Multi-Modal 3D Scene Dataset with Hierarchical Grounded Language Annotations (NeurIPS 2024 · similarity 0.86)
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.
