Research map / 3D vision / 3D generation
Describe, Adapt and Combine: Empowering CLIP Encoders for Open-set 3D Object Retrieval
Zhichuan Wang, Yang Zhou, Zhe Liu, Rui Yu, Song Bai, Yulong Wang, Xinwei He, 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 InternVL descriptions at inference (ablation) | DAC (CLIP ViT-B/32) · OS-ABO-core | mAP: +9.65% mAP | Improved |
| add InternVL descriptions at inference (ablation) | DAC (CLIP ViT-B/32) · OS-ESB-core | mAP: 57.45 -> 58.70 | Improved |
| add AB-LoRA adaptation (ablation) | DAC · OS-MN40-core | mAP: +7.01% mAP (ViT-B/32), +9.21% (ViT-L/14) | Improved |
| add additive bias to LoRA (ablation) | DAC (CLIP ViT-B/32) · OS-MN40-core | mAP: 59.85 -> 62.40 | Improved |
| add LoRA adaptation to CLIP (ablation) | DAC (CLIP ViT-B/32) · OS-MN40-core | mAP: 55.39 -> 59.85 | Improved |
| fuse by addition instead of concatenation (ablation) | DAC (CLIP ViT-B/32) · OS-MN40-core | mAP: 55.30 -> 62.40 | Improved |
| fuse by addition instead of concatenation (ablation) | DAC (CLIP ViT-L/14) · OS-MN40-core | mAP: 56.74 -> 68.98 | Improved |
| raise LoRA rank from 8 to 12 (ablation) | DAC · OS-MN40-core | mAP: 62.40 -> 62.17 | No clear effect |
| add Tanh normalization after fusion (ablation) | DAC (CLIP ViT-B/32) · OS-ABO-core | mAP: 64.63 -> 66.10 | Improved |
| use ReLU after fusion (ablation) | DAC (CLIP ViT-B/32) · OS-ABO-core | mAP: 64.63 -> 63.13 | Hurt |
| train with InternVL category descriptions instead of hand-crafted template (text fusion at retrieval) (ablation) | DAC (CLIP ViT-B/32) · OS-ABO-core | mAP gain from text: +9.06 -> +9.65 | Improved |
| use category-level instead of individual-level descriptions (ablation) | DAC (CLIP ViT-B/32) · OS-ESB-core | mAP: 48.78 -> 58.70 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Uni3D — Uni-3D: A Universal Model for Panoptic 3D Scene Reconstruction (ICCV 2023) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Duoduo CLIP: Efficient 3D Understanding with Multi-View Images (ICLR 2025 · similarity 0.88)
- ImageNet3D: Towards General-Purpose Object-Level 3D Understanding (NeurIPS 2024 · similarity 0.87)
- Cross-Modal 3D Representation with Multi-View Images and Point Clouds (CVPR 2025 · similarity 0.86)
- Tango3D: Towards Alignment for Global and Local 2D-3D Correspondence (NeurIPS 2026 · similarity 0.86)
- 3DCoMPaT200: Language Grounded Large-Scale 3D Vision Dataset for Compositional Recognition (NeurIPS 2024 · similarity 0.86)
- OpenNeRF: Open Set 3D Neural Scene Segmentation with Pixel-Wise Features and Rendered Novel Views (ICLR 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.
