Research map / Point clouds / Point cloud classification
Distilling Coarse-to-Fine Semantic Matching Knowledge for Weakly Supervised 3D Visual Grounding
Zehan Wang, Haifeng Huang, Yang Zhao, Linjun Li, Xize Cheng, Yichen Zhu, Aoxiong Yin, Zhou Zhao
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 |
|---|---|---|---|
| use coarse-to-fine semantic matching with distillation instead of MIL-NCE | weakly supervised 3DVG · ScanRefer (ScanNet detector) | R@3 IoU@0.25 Overall: 23.53 -> 34.12 | Improved |
| add class similarity (L_cls) to coarse candidate selection (ablation) | coarse-to-fine matching model · ScanRefer (ScanNet detector) | R@3 IoU@0.25: 12.28 -> 32.71 | Improved |
| add feature similarity (L_match) in addition to class similarity (ablation) | coarse-to-fine matching model · ScanRefer (ScanNet detector) | R@3 IoU@0.25: 32.71 -> 33.03 | Improved |
| add fine-grained masked-word reconstruction (L_recon) (ablation) | coarse-to-fine matching model · ScanRefer (ScanNet detector) | R@3 IoU@0.25: 33.03 -> 34.12 | Improved |
| select K=8 candidates rather than 4 or 32 (ablation) | coarse-grained selection · ScanRefer (ScanNet detector) | R@3 IoU@0.25: 33.68 (K=4), 31.29 (K=32) -> 34.12 (K=8) | Improved |
| add knowledge distillation into 3DJCG matching module (ablation) | coarse-to-fine matching model · ScanRefer (ScanNet detector) | R@3 IoU@0.25: 31.71 -> 34.12 | Improved |
| add knowledge distillation for inference efficiency (ablation) | matching phase · ScanRefer | matching time and params: 31.4 ms -> 8.22 ms; 5.85 M -> 1.93 M | Improved |
| distill into different matching architectures (SAT, 3DVG-Transformer, 3DJCG) (ablation) | distillation · ScanRefer (ScanNet detector) | R@3 IoU@0.25: 31.71 -> 33.70 (SAT), 34.61 (3DVG-Trans), 34.12 (3DJCG) | Improved |
| use detector pretrained on SUN RGB-D instead of ScanNet | Nr3D/Sr3D · Nr3D and Sr3D | R@3 IoU@0.25: 22.45 -> 10.36 (Nr3D) | Hurt |
| use class similarity alone vs adding feature similarity alone (ablation) | coarse selection · ScanRefer (SUN RGB-D detector) | R@3 IoU@0.25: 9.03 -> 10.53 (L_match), 9.03 -> 13.10 (L_cls) | Improved |
Closest papers by meaning
- G^3-LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding (CVPR 2024 · similarity 0.90)
- Cross-Modal Match for Language Conditioned 3D Object Grounding (AAAI 2024 · similarity 0.90)
- DenseGrounding: Improving Dense Language-Vision Semantics for Ego-centric 3D Visual Grounding (ICLR 2025 · similarity 0.89)
- From Objects to Anywhere: A Holistic Benchmark for Multi-level Visual Grounding in 3D Scenes (NeurIPS 2025 · similarity 0.89)
- AugRefer: Advancing 3D Visual Grounding via Cross-Modal Augmentation and Spatial Relation-based Referring (AAAI 2025 · similarity 0.89)
- Towards CLIP-driven Language-free 3D Visual Grounding via 2D-3D Relational Enhancement and Consistency (CVPR 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.
