Research map / Vision-language and multimodal / Cross-modal learning
Is Contrastive Distillation Enough for Learning Comprehensive 3D Representations?
Yifan Zhang, Junhui Hou
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 image reconstruction and occupancy estimation heads (ablation) | pixel-point contrastive pretraining · nuScenes pretraining ablation (linear probing) | mIoU: 38.5 -> 43.4 | Improved |
| add multi-modal unified codebook (ablation) | PP + Rec. · nuScenes pretraining ablation (linear probing) | mIoU: 43.4 -> 45.6 | Improved |
| add geometry-enhanced masked image modeling (ablation) | PP + Rec. + Codebook · nuScenes pretraining ablation (linear probing) | mIoU: 45.6 -> 46.3 | Improved |
| add L_kl loss (ablation) | PP + Rec. + Codebook + Geo. · nuScenes pretraining ablation (linear probing) | mIoU: 46.3 -> 51.2 | Improved |
| increase codebook size from 128 to 512 (ablation) | CMCR · nuScenes codebook size (linear probing) | mIoU: 45.3 -> 51.2 | Improved |
| increase codebook size from 512 to 1024 (ablation) | CMCR · nuScenes codebook size (linear probing) | mIoU: 51.2 -> 49.5 | Hurt |
| increase codebook size from 512 to 2048 (ablation) | CMCR · nuScenes codebook size (linear probing) | mIoU: 51.2 -> 47.6 | Hurt |
| train the 2D backbone jointly with MIM instead of freezing (ablation) | CMCR · nuScenes 2D backbone (linear probing) | mIoU: 48.5 -> 51.2 | Improved |
| add image features to 3D occupancy estimation (ablation) | occupancy head · nuScenes occupancy estimation | mIoU: 56.56 -> 57.34 | Improved |
| add orthogonal loss between shared and specific features (ablation) | CMCR · feature disentanglement visualization | separation | Improved |
| add self-supervised pretraining to WaffleIron | WaffleIron WI-768 · nuScenes semantic segmentation (1% labels) | mIoU: 35.41 -> 53.64 | Improved |
| add self-supervised pretraining to WaffleIron | WaffleIron WI-768 · nuScenes semantic segmentation (100% labels) | mIoU: 78.06 -> 78.93 | Improved |
Closest papers by meaning
- Hybrid Distillation: Connecting Masked Autoencoders with Contrastive Learners (ICLR 2024 · similarity 0.89)
- Fine-grained Image-to-LiDAR Contrastive Distillation with Visual Foundation Models (NeurIPS 2024 · similarity 0.88)
- Connect, Collapse, Corrupt: Learning Cross-Modal Tasks with Uni-Modal Data (ICLR 2024 · similarity 0.87)
- Beyond Modality Collapse: Representation Blending for Multimodal Dataset Distillation (NeurIPS 2025 · similarity 0.87)
- MM-Mixing: Multi-Modal Mixing Alignment for 3D Understanding (AAAI 2025 · similarity 0.86)
- Towards Compact 3D Representations via Point Feature Enhancement Masked Autoencoders (AAAI 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.
