Research map / 3D vision / 3D generation
ConDense: Consistent 2D-3D Pre-training for Dense and Sparse Features from Multi-View Images
Xiaoshuai Zhang, Zhicheng Wang, Howard Zhou, Soham Ghosh, Danushen L Gnanapragasam, Varun Jampani, Hao Su, Leonidas Guibas
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 |
|---|---|---|---|
| unfreeze the 2D encoder during joint training (ablation) | ConDense · ScanObjectNN / ImageNet-1k | accuracy: 91.3 -> 94.1 (3D); 86.5 -> 89.6 (2D) | Improved |
| add 2D-3D key point loss L_p (ablation) | ConDense · ScanObjectNN / ImageNet-1k | accuracy: 90.5 -> 94.1 (3D); 87.1 -> 89.6 (2D) | Improved |
| add 2D fidelity loss to DINOv2 features (ablation) | ConDense · ScanObjectNN / ImageNet-1k | accuracy: 89.7 -> 94.1 (3D); 79.9 -> 89.6 (2D) | Improved |
| use full pre-training data instead of 10% (ablation) | ConDense · ScanObjectNN / ImageNet-1k | accuracy: 88.7 -> 94.1 (3D); 79.1 -> 89.6 (2D) | Improved |
| 2D-3D consensus pre-training of DINOv2 backbone | DINOv2 ViT-g/14 · ImageNet-1k, Places205, ADE20k, PascalVOC | accuracy / mIoU: ImageNet 86.5 -> 89.6; ADE20k 49.0 -> 53.6 | Improved |
| use higher-quality NeRFs (4k steps instead of 2k) for pre-training (ablation) | ConDense (lighter version) · ImageNet / ScanObjectNN | accuracy: 86.2 / 88.9 -> 87.7 / 93.2 | Improved |
| use MinkowskiNet instead of PointBERT as 3D backbone (ablation) | ConDense · ScanObjectNN | accuracy: 87.2 -> 93.2 | Improved |
| use key point matching instead of global features for scene queries (ablation) | ConDense · ScanNet scene retrieval | accuracy: 70.1 -> 78.4 | Improved |
| apply 2D-3D consensus to DINOv2 features for depth estimation | ConDense · NYUd depth (lin. 4) | RMSE: 0.298 -> 0.322 | Hurt |
| pre-train with ConDense when 3D labeled data is scarce | 3D segmentation · ScanNet 1% | mIoU: Swin3D 54.8 -> ConDense 67.3 | Improved |
Closest papers by meaning
- Multi-View Representation is What You Need for Point-Cloud Pre-Training (ICLR 2024 · similarity 0.86)
- From Thousands to Billions: 3D Visual Language Grounding via Render-Supervised Distillation from 2D VLMs (ICML 2025 · similarity 0.86)
- ULIP-2: Towards Scalable Multimodal Pre-training for 3D Understanding (CVPR 2024 · similarity 0.86)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.86)
- UniGS: Unified Language-Image-3D Pretraining with Gaussian Splatting (ICLR 2025 · similarity 0.86)
- Uni3D: Exploring Unified 3D Representation at Scale (ICLR 2024 · similarity 0.85)
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.
