Research map / 3D vision / 3D generation
DINO Eats CLIP: Adapting Beyond Knowns for Open-set 3D Object Retrieval
Xinwei He, Yansong Zheng, Qianru Han, Zhichuan Wang, Yuxuan Cai, Yang Zhou, Jingbo Xia, Yulong Wang, Jinhai Xiang, Xiang Bai
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 Chunking and Adapting Module (CAM) (ablation) | DINOv2 ViT-B/14 adapter · OS-MN40-core | mAP: 63.49 -> 66.43 | Improved |
| add Chunking and Adapting Module (CAM) (ablation) | DINOv2 ViT-L/14 adapter · OS-MN40-core | mAP: 65.89 -> 68.43 | Improved |
| add Virtual Feature Synthesis (VFS) (ablation) | DINOv2 ViT-B/14 + CAM · OS-MN40-core | mAP: 66.43 -> 67.62 | Improved |
| add Virtual Feature Synthesis (VFS) (ablation) | DINOv2 ViT-L/14 + CAM · OS-MN40-core | mAP: 68.43 -> 69.69 | Improved |
| use CAM instead of CLIP-Adaptor (ablation) | DINOv2 ViT-B/14 · OS-MN40-core | mAP: 64.35 -> 66.43 | Improved |
| use chunk size 3 instead of 1 (ablation) | CAM · OS-MN40-core | mAP: 63.62 -> 67.62 | Improved |
| use chunk size 3 instead of 5 (ablation) | CAM · OS-MN40-core | mAP: 64.91 -> 67.62 | Improved |
| use chunk size 3 instead of 7 (ablation) | CAM · OS-MN40-core | mAP: 65.72 -> 67.62 | Improved |
| use non-overlapping stride 3 instead of stride 1 (ablation) | CAM · OS-MN40-core | mAP: 64.64 -> 67.62 | Improved |
| use non-overlapping stride 3 instead of stride 2 (ablation) | CAM · OS-MN40-core | mAP: 65.52 -> 67.62 | Improved |
| naively fuse CLIP features with DINOv2 instead of using CLIP only for VFS (ablation) | DINOv2 baseline · OS-MN40-core | mAP: 62.77 -> 61.35 | Hurt |
| raise fusion weight lambda beyond about 0.4 (ablation) | CAM residual fusion · OS-MN40-core | mAP | Mixed |
Closest papers by meaning
- Describe, Adapt and Combine: Empowering CLIP Encoders for Open-set 3D Object Retrieval (ICCV 2025 · similarity 0.93)
- Duoduo CLIP: Efficient 3D Understanding with Multi-View Images (ICLR 2025 · similarity 0.88)
- A General Protocol to Probe Large Vision Models for 3D Physical Understanding (NeurIPS 2024 · similarity 0.87)
- OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views (ICLR 2024 · similarity 0.86)
- v-CLR: View-Consistent Learning for Open-World Instance Segmentation (CVPR 2025 · similarity 0.86)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (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.
