Research map / Vision-language and multimodal / Human and embodied modeling
G^3-LQ: Marrying Hyperbolic Alignment with Explicit Semantic-Geometric Modeling for 3D Visual Grounding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes G3-LQ, a DETR-style 3D visual grounding model with geometric point-proxy features (PAGE), language-scene-graph-guided fine-grained query selection (Flan-QS), and a Poincaré-ball contrastive alignment loss (PSA), reporting SOTA on ScanRefer, Nr3D and Sr3D.
Problem
3D visual grounding of free-form referring expressions in point clouds suffers from weak modeling of geometric/spatial structure, coarse text guidance for object query generation, and Euclidean alignment that fails to capture hierarchical semantic-geometric relations.
Approach
Combines explicit local geometric proxies with adaptive refinement, selects object queries using a dependency-parsed language scene graph, and aligns vision and text features on a Poincaré ball with a hyperbolic contrastive loss.
Key results
- On ScanRefer, G3-LQ (one-stage) reaches 55.95% overall at IoU 0.25 and 44.72% at IoU 0.5, versus EDA's 53.83% and 41.70%.
- G3-LQ (two-stage) reaches 56.90% / 45.58% overall on ScanRefer, outperforming EDA by 2.31 and 2.34 points.
- G3-LQ achieves 58.4% on Nr3D and 73.1% on Sr3D overall, above prior methods such as 3D-VisTA (57.5% and 69.6%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add PAGE geometric feature modeling (ablation) | G3-LQ on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 85.67 -> 86.89 | Helped |
| add Flan-QS fine-grained language-guided query selection (ablation) | G3-LQ on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 85.67 -> 86.75 | Helped |
| add PSA hyperbolic semantic alignment (ablation) | G3-LQ on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 85.67 -> 86.32 | Helped |
| combine PAGE and Flan-QS (ablation) | G3-LQ on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 85.67 -> 87.66 | Helped |
| add PSA to PAGE and Flan-QS (ablation) | G3-LQ on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 87.66 -> 88.09 | Helped |
| add relative-distance prior to point-position prior (ablation) | PAGE on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 85.67 -> 85.83 | Helped |
| use max pooling for geometric feature aggregation (ablation) | PAGE on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 85.67 -> 85.98 | Helped |
| use softmax geometric feature aggregation (ablation) | PAGE on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 85.67 -> 86.04 | Helped |
| use affine geometric refinement (ablation) | PAGE on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 85.67 -> 86.57 | Helped |
| replace parametric queries with language scene graph query selection (ablation) | G3-LQ on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 81.40 -> 87.66 | Helped |
| replace Top-K selection with language scene graph query selection (ablation) | G3-LQ on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 85.67 -> 87.66 | Helped |
| use language scene graph rather than sentence-level language guidance (ablation) | G3-LQ on ScanRefer · ScanRefer Unique, IoU 0.25 | accuracy: 87.10 -> 87.66 | Helped |
Limitations
- Gains over EDA on ScanRefer are modest (about 2.3 points overall), and the ablation is limited to ScanRefer.
- Relies on an off-the-shelf scene graph parser for language graph construction, adding a dependency on external parsing quality.
- Paper does not report compute, training cost, or inference efficiency.
Experimental setup
Datasets: ScanRefer Nr3D Sr3D
Benchmarks: ScanRefer Nr3D Sr3D
Metrics: accuracy@0.25 IoU accuracy@0.5 IoU
Baselines: ScanRefer TGNN InstanceRefer SAT FFL-3DOG 3DVG 3D-SPS BUTD-DETR ViL3DRel EDA 3DJCG 3D-VLP
Builds on: EDA GroundingDINO PointNet++ RoBERTa MDETR
Closest papers by meaning
- Grounded 3D-Aware Spatial Vision-Language Modeling (CVPR 2026 · similarity 0.92)
- DenseGrounding: Improving Dense Language-Vision Semantics for Ego-centric 3D Visual Grounding (ICLR 2025 · similarity 0.91)
- Towards CLIP-driven Language-free 3D Visual Grounding via 2D-3D Relational Enhancement and Consistency (CVPR 2024 · similarity 0.91)
- Let Language Constrain Geometry: Vision–Language Models as Semantic and Spatial Critics for 3D Generation (ICML 2026 · similarity 0.90)
- From Objects to Anywhere: A Holistic Benchmark for Multi-level Visual Grounding in 3D Scenes (NeurIPS 2025 · similarity 0.90)
- G$^2$VLM: Geometry Grounded Vision Language Model with Unified 3D Reconstruction and Spatial Reasoning (CVPR 2026 · similarity 0.90)
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.