Detect What You Need: Chain-of-Causal Reasoning for 3D Intent Grounding
Organizations: School of Artificial Intelligence, College of Intelligence and Computing, Tianjin University, China · School of Computer Science and Information Engineering, Hefei University of Technology, China
Abstract
Accurately matching human intentions in 3D space is an important goal of artificial intelligence. Recently, 3D Intention Grounding (3D-IG) has emerged, aiming to localize target 3D objects given a natural-language intent. Unlike conventional visual grounding with descriptive referring expressions, 3D-IG intents are abstract and non-descriptive, making object localization substantially more challenging. This requires inferring latent functional requirements from non-descriptive intents and aligning them with object-level 3D representations. However, existing methods largely rely on implicit intent-object matching, leading to logical gaps and limited interpretability, robustness, and generalization. To address these challenges, we propose Chain-of-Causal Reasoning (CoCR), a causality-inspired functional dependency reasoning framework that explicitly bridges abstract intentions and candidate objects through intermediate functional requirements. Specifically, CoCR progressively decomposes complex intents into ordered functional requirements, thereby forming an explicit intent-function-object reasoning chain that links abstract intentions to object suitability. Building on this chain, we construct a causality-inspired functional dependency graph to model requirement-attribute relationships and introduce a causal-visual alignment module that aligns function-aware representations with the geometric-semantic features of 3D point clouds, enabling bidirectional verification between structured functional reasoning and visual evidence. Extensive experiments on 3D Intention Grounding and 3D Visual Grounding demonstrate that our method enhances intent-aware object localization. Our code is available at https://2490o.github.io/CoCR.
Figures & tables
| Method | Detector | Val Set | Test Set | ||||||
|---|---|---|---|---|---|---|---|---|---|
| [email protected] | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] | [email protected] | ||
| BUTD-DETR [ 19 ] | GroupFree [ 20 ] | 47.12 | 24.56 | 31.05 | 13.05 | 47.86 | 25.74 | 31.41 | 13.46 |
| EDA [ 21 ] | GroupFree [ 20 ] | 43.11 | 18.91 | 14.02 | 5.00 | 44.00 | 19.62 | 14.56 | 5.18 |
| 3D-ViSTA [ 22 ] | Mask3D [ 23 ] | 42.76 | 30.37 | 36.10 | 19.93 | 43.88 | 31.44 | 37.29 | 22.00 |
| Chat-3D-v2 0 [ 24 ] | PointGroup [ 25 ] | 5.86 | 5.24 | 0.15 | 0.13 | 5.63 | 4.93 | 0.14 | 0.11 |
| Chat-3D-v2 [ 24 ] | PointGroup [ 25 ] | 36.71 | 32.78 | 3.23 | 2.58 | 33.46 | 29.32 | 2.67 | 2.10 |
| Methods | Nr3D | Sr3D | ||
|---|---|---|---|---|
| Overall | Hard | Overall | Hard | |
| ReferIt3D [ 12 ] | 35.6 | 27.9 | 40.8 | 31.5 |
| TGNN [ 26 ] | 37.3 | 30.6 | 45.0 | 36.9 |
| 3D-SPS [ 27 ] | 51.5 | 45.1 | 62.6 | 65.4 |
| EDA [ 21 ] | 52.1 | 46.1 | 68.1 | 62.9 |
| MVT [ 28 ] | 55.4 | 49.1 | 64.5 | 58.8 |
| Methods | Metrics | |||||
|---|---|---|---|---|---|---|
| ICP | ICG | CVFA | [email protected] | [email protected] | [email protected] | [email protected] |
| 55.32 | 38.70 | 40.72 | 24.40 | |||
| 57.82 | 40.31 | 42.85 | 25.70 | |||
| 58.68 | 41.07 | 43.76 | 25.37 | |||
| 59.28 | 42.46 | 43.71 | 27.01 | |||
| 60.99 | 43.67 | 44.25 | 28.38 | |||
| Method | Training Time | Inference | Memory | Params |
|---|---|---|---|---|
| (hours) | (ms) | (GB) | (M) | |
| IntentNet [ 5 ] | 32 | 27.1 | 7.1 | 148.6 |
| Baseline | 26 | 25.9 | 6.8 | 147.3 |
| Baseline + ICP | 34 | 28.4 | 7.3 | 207.8 |
| Baseline + ICP + ICG | 38 | 29.1 | 7.6 | 208.6 |
| Full Model (Ours) | 40 | 32.7 | 7.6 | 209.5 |
| Perturbation | [email protected] | [email protected] | [email protected] | [email protected] |
|---|---|---|---|---|
| Full Model | 61.37 | 45.91 | 45.73 | 30.28 |
| Top-contribution removal | 52.51 | 35.85 | 34.13 | 21.45 |
| Random removal | 58.45 | 42.39 | 42.27 | 27.57 |
| Edge shuffling | 54.56 | 38.28 | 39.61 | 24.32 |
| Pruning Threshold ( ) | 0.1 | 0.2 | 0.3 | 0.4 | 0.5 |
|---|---|---|---|---|---|
| [email protected] | 25.41 | 27.33 | 25.82 | 22.35 | 21.92 |
| [email protected] | 41.75 | 43.24 | 42.63 | 40.89 | 38.41 |
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Method | Venue | Unique ( 19%) | Multiple ( 81%) | Overall | |||
|---|---|---|---|---|---|---|---|
| 0.25 | 0.5 | 0.25 | 0.5 | 0.25 | 0.5 | ||
| ScanRefer [ 11 ] | ECCV20 | 67.6 | 46.2 | 32.1 | 21.3 | 40.0 | 26.1 |
| EDA [ 21 ] | CVPR23 | 85.8 | 68.6 | 49.1 | 37.6 | 54.6 | 42.3 |
| VPP-Net [ 3 ] | CVPR24 | 86.1 | 67.1 | 50.3 | 39.0 | 55.7 | 43.3 |
| M3DRef-CLIP [ 37 ] | ICCV23 | - | 77.2 | - | 36.8 | - | 44.7 |
| G3-LQ [ 29 ] | CVPR24 | 88.6 | 73.3 | 50.2 | 39.7 | 56.0 | 44.7 |
| Methods | Nr3D | Sr3D | ||
|---|---|---|---|---|
| Overall | Hard | Overall | Hard | |
| Baseline | 64.7 | 55.1 | 75.3 | 69.0 |
| + ICP | 65.1 | 55.4 | 75.3 | 69.2 |
| + ICG | 66.5 | 58.3 | 81.1 | 71.4 |
| + ICP + ICG | 66.4 | 58.9 | 80.9 | 72.4 |
| + ICP + ICG + CVFA | 67.1 | 59.8 | 81.6 | 73.7 |
| Method | Detector | [email protected] | [email protected] | [email protected] | [email protected] |
|---|---|---|---|---|---|
| Gemini [ 44 ] | GroupFree | 39.21 | 29.34 | 12.17 | 9.36 |
| GPT-4 [ 42 ] | GroupFree | 41.40 | 28.40 | 15.10 | 7.76 |
| GPT-5 [ 43 ] | GroupFree | 42.50 | 29.90 | 14.70 | 8.50 |
| IntentNet [ 5 ] | GroupFree | 58.34 | 40.83 | 41.90 | 25.36 |
| Ours | GroupFree | 60.71 | 43.24 | 42.65 | 27.33 |
| Method | Top-1 Acc (%) | Top-5 Acc (%) |
|---|---|---|
| GPT-5 [ 43 ] | 70.6 | 87.4 |
| IntentNet [ 5 ] | 76.8 | 88.9 |
| Ours | 83.2 | 91.3 |
3D Visual Grounding
Three-dimensional visual grounding links language or visual queries to specific objects, parts, regions, and spatial relations in 3D scenes.
See all 149 papers