Research map / Vision-language and multimodal / Multimodal LLMs
3D CoCa v2: Contrastive Learners with Test-Time Search for Generalizable Spatial Intelligence
H. W. Tang, Ting Huang, Zeyu Zhang
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 test-time search (best-of-8, LLM judge) to 3D CoCa | 3D CoCa · ScanRefer (w/o 2D, IoU 0.25) | CIDEr: 85.42 -> 86.95 | Improved |
| add test-time search to 3D CoCa | 3D CoCa · ScanRefer (w/o 2D, IoU 0.50) | CIDEr: 77.13 -> 78.63 | Improved |
| add test-time search to 3D CoCa | 3D CoCa · Nr3D (IoU 0.5) | CIDEr: 52.84 -> 54.45 | Improved |
| add test-time search to 3D CoCa | 3D CoCa · TOD3Cap (zero-shot OOD, IoU 0.5) | CIDEr: 47.2 -> 50.8 | Improved |
| add test-time search to 3D CoCa | 3D CoCa · TOD3Cap (zero-shot OOD, IoU 0.25) | CIDEr: 55.8 -> 59.6 | Improved |
| add test-time search to 3D CoCa | 3D CoCa · ScanRefer (w/ 2D, IoU 0.50) | CIDEr: 74.52 -> 75.60 | Improved |
| add contrastive loss (lambda 0 -> 1.0) (ablation) | 3D CoCa backbone · ScanRefer (IoU 0.25, no TTS) | CIDEr: 74.12 -> 85.42 | Improved |
| raise contrastive weight from 1.0 to 2.0 (ablation) | 3D CoCa backbone · ScanRefer (IoU 0.25, no TTS) | CIDEr: 85.42 -> 76.89 | Hurt |
| use CoCa-style multimodal decoder instead of GPT-2 captioner (ablation) | 3D CoCa backbone · ScanRefer (IoU 0.25, no TTS) | CIDEr: 76.20 -> 85.42 | Improved |
| use proposed point-tokenizer + frozen CLIP encoder instead of PointNet++ (ablation) | 3D CoCa backbone · ScanRefer (IoU 0.25, no TTS) | CIDEr: 72.48 -> 85.42 | Improved |
| use GPT-5 judge instead of Qwen3-VL Pro (ablation) | 3D CoCa v2 TTS · ScanRefer (IoU 0.5) | CIDEr: 78.00 -> 78.63 | Improved |
| use GPT-5 judge instead of Qwen3-VL Pro (ablation) | 3D CoCa v2 TTS · TOD3Cap (zero-shot OOD, IoU 0.5) | CIDEr: 49.20 -> 50.80 | Improved |
Closest papers by meaning
- SpatialCLIP: Learning 3D-aware Image Representations from Spatially Discriminative Language (CVPR 2025 · similarity 0.90)
- From Flatland to Space: Teaching Vision-Language Models to Perceive and Reason in 3D (NeurIPS 2025 · similarity 0.90)
- RegionPLC: Regional Point-Language Contrastive Learning for Open-World 3D Scene Understanding (CVPR 2024 · similarity 0.90)
- GPT4Scene: Understand 3D Scenes from Videos with Vision-Language Models (ICLR 2026 · similarity 0.90)
- Grounded 3D-Aware Spatial Vision-Language Modeling (CVPR 2026 · similarity 0.89)
- MMScan: A Multi-Modal 3D Scene Dataset with Hierarchical Grounded Language Annotations (NeurIPS 2024 · similarity 0.89)
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.
