Overall grounding accuracy on ScanRefer, Nr3D and Sr3D among all methods compared in Tables 1–2.
01 · Motivation
Ground locally,
not globally.
Most 3D grounding systems reason over every proposal in a scene. But referring expressions usually depend on a small neighborhood: the target, one or two anchors, and their spatial relations.
Clutter creates ambiguity
Same-class distractors and imperfect segmentation enlarge the candidate set, increase computation, and propagate errors.
Language narrows the search
A frozen VLM keeps only the relevant spatial region, then adds visual context to make the query easier to ground.
02 · Interactive Explorer
Same room.
Language changes the target.
When several chairs look alike, their relationships to the TV, doorway and table distinguish the target. Compare two descriptions in one room to see why PruneGround uses spatial context to narrow the search.
Same room. Different descriptions.
Several chairs look alike. Choose a description to see which one its spatial relationship identifies.
Mesh statistics & data sources
Scene data and chair A: official examples ↗. Chair B’s displayed region is manually specified from its reference bounds; MCDR was rerun on two fresh views of this crop. Crop details ↗ · MCDR run ↗ · Run records ↗.
Object boxes use ScanNet instance annotations, not downstream model predictions: chairs ↗ · sofa by door ↗ · sofa near window ↗. The sofa instances were matched to the displayed examples; these are not released query-ground-truth pairs. The displayed A/B descriptions and sofa-chair coffee-table relation use user-supplied wording; original model outputs remain unchanged in the source files. The sofa-by-door example retains its released MCDR text.
03 · Method
Three stages.
One focused search.
The framework progressively reduces spatial noise and improves language structure before predicting the target 3D bounding box.
Language-Guided
Spatial Pruning
LGSP renders one top view plus four oblique RGB-depth views. Conditioned on the query, a frozen VLM predicts a relevant 2D region that is projected back into the 3D point cloud.
- Removes language-irrelevant points
- Reduces same-class distractors
- Retains full height along the z-axis
Multi-View-Conditioned Description Reformulation
MCDR uses local side views to split free-form language into compact target–anchor statements, then augments missing spatial cues visible in the scene.
- Simplifies complex expressions
- Adds visually grounded anchors
- Encodes explicit spatial relations
LLM-based
Grounder
A detection-pretrained SpatialLM is repurposed for language-conditioned grounding. Point features and text tokens are jointly decoded into one target box.
- Sonata point-cloud encoder
- Qwen2.5-0.5B language decoder
- Single target-selective prediction
04 · Results
Grounding accuracy across
three 3DVG benchmarks.
Complete comparisons from Tables 1 and 2 on ScanRefer, Nr3D, and Sr3D. PruneGround achieves the highest overall scores on all three benchmarks among the methods reported.
ScanRefer overall accuracy at IoU ≥ 0.25, up 2.8 percentage points over PanoGrounder.
Nr3D overall accuracy at IoU ≥ 0.5 on human-written descriptions with same-class distractors.
Sr3D overall accuracy at IoU ≥ 0.5 on descriptions that distinguish objects through spatial relations.
| Method | Venue | Unique | Multiple | Overall | |||
|---|---|---|---|---|---|---|---|
| Acc@0.25 | Acc@0.5 | Acc@0.25 | Acc@0.5 | Acc@0.25 | Acc@0.5 | ||
| ScanReferChen et al., 2020 | ECCV | 67.6 | 46.2 | 32.1 | 21.3 | 40.0 | 26.1 |
| MVTHuang et al., 2022a | CVPR | 77.7 | 66.5 | 31.9 | 25.3 | 40.8 | 33.3 |
| 3D-SPSLuo et al., 2022 | CVPR | 84.1 | 66.7 | 40.3 | 29.8 | 48.8 | 37.0 |
| EDAWu et al., 2023b | CVPR | 85.8 | 68.6 | 49.1 | 37.6 | 54.6 | 42.3 |
| ConcreteNetUnal et al., 2023 | ECCV | 86.4 | 82.1 | 42.4 | 38.4 | 50.6 | 46.5 |
| Chat-SceneHuang et al., 2023 | NeurIPS | 89.6 | 82.5 | 47.8 | 42.9 | 55.5 | 50.2 |
| VPP-NetShi et al., 2024 | CVPR | 86.1 | 67.1 | 50.3 | 39.0 | 55.7 | 43.3 |
| G³-LQWang et al., 2024 | CVPR | 88.6 | 73.3 | 50.2 | 39.7 | 56.0 | 44.7 |
| MA²TransVGXu et al., 2024a | CVPR | 86.3 | 74.1 | 53.8 | 41.4 | 57.9 | 45.7 |
| MCLNQian et al., 2024 | ECCV | 86.9 | 72.7 | 52.0 | 40.8 | 57.2 | 45.5 |
| PQ3DZhu et al., 2024b | ECCV | 86.7 | 78.3 | 51.5 | 46.2 | 57.0 | 51.2 |
| VGMambaZhu et al., 2025b | ICCV | 91.9 | 79.6 | 54.8 | 47.6 | 60.0 | 53.9 |
| PanoGrounderJung et al., 2025 | arXiv | 84.3 | — | 55.3 | — | 61.0 | — |
| PruneGroundOurs | — | 90.2 | 84.7 | 57.6 | 49.4 | 63.8 | 56.1 |
Bold: best · Underline: second-best · —: not reported. Paper Table 1 ↗
| Method | Venue | Easy | Hard | View-D | View-I | Overall |
|---|---|---|---|---|---|---|
| ReferIt3DAchlioptas et al., 2020 | ECCV | 43.6 | 27.9 | 32.5 | 37.1 | 35.6 |
| ViL3DRelChen et al., 2022a | NeurIPS | 70.2 | 57.4 | 62.0 | 64.5 | 64.4 |
| 3D-VisTAZhu et al., 2023 | ICCV | 65.9 | 49.4 | 53.7 | 59.4 | 57.5 |
| MIKASAChang et al., 2024 | CVPR | 69.7 | 59.4 | 65.4 | 64.0 | 64.4 |
| G³-LQWang et al., 2024 | CVPR | — | 50.7 | — | — | 58.4 |
| MA²TransVGXu et al., 2024a | CVPR | — | 57.6 | — | — | 65.2 |
| PQ3DZhu et al., 2024b | ICCV | 73.3 | 56.7 | 60.7 | 67.0 | 64.9 |
| LIBAWang et al., 2025b | AAAI | — | 57.2 | 60.3 | — | 64.5 |
| ViewSRDHuang et al., 2025 | ICCV | 75.3 | 64.8 | 68.6 | 70.6 | 69.9 |
| VGMambaZhu et al., 2025b | ICCV | — | 61.4 | — | — | 68.3 |
| PanoGrounderJung et al., 2025 | arXiv | 82.2 | 67.2 | 70.5 | 76.3 | 74.6 |
| PruneGroundOurs | — | 84.6 | 68.8 | 71.2 | 76.6 | 75.1 |
Bold: best · Underline: second-best · —: not reported. Paper Table 2 ↗ View-D: view-dependent; View-I: view-independent.
| Method | Venue | Easy | Hard | View-D | View-I | Overall |
|---|---|---|---|---|---|---|
| ReferIt3DAchlioptas et al., 2020 | ECCV | 44.7 | 31.5 | 39.2 | 40.8 | 40.8 |
| ViL3DRelChen et al., 2022a | NeurIPS | 74.9 | 67.9 | 63.8 | 73.2 | 72.8 |
| 3D-VisTAZhu et al., 2023 | ICCV | 72.1 | 63.6 | 57.9 | 70.1 | 69.6 |
| MIKASAChang et al., 2024 | CVPR | 78.6 | 67.3 | 70.4 | 75.4 | 75.2 |
| G³-LQWang et al., 2024 | CVPR | — | 66.3 | — | — | 73.1 |
| MA²TransVGXu et al., 2024a | CVPR | — | 69.3 | — | — | 73.9 |
| PQ3DZhu et al., 2024b | ICCV | 78.8 | 68.2 | 51.5 | 76.7 | 75.6 |
| LIBAWang et al., 2025b | AAAI | — | 70.2 | 61.7 | — | 75.8 |
| ViewSRDHuang et al., 2025 | ICCV | 78.3 | 70.6 | 69.0 | 76.2 | 76.0 |
| VGMambaZhu et al., 2025b | ICCV | — | 74.4 | — | — | 81.3 |
| PanoGrounderJung et al., 2025 | arXiv | 81.3 | 74.2 | 60.5 | 80.0 | 79.1 |
| PruneGroundOurs | — | 83.7 | 76.5 | 61.7 | 82.4 | 81.5 |
Bold: best · Underline: second-best · —: not reported. Paper Table 2 ↗ View-D: view-dependent; View-I: view-independent.
Table 6 · Accuracy–speed trade-off
Small model.
Near-frontier target recall.
Compare the frozen VLMs used in LGSP and MCDR.
1.4s vs. 13.1s per VLM stage
−1.0 pp recall vs. Gemini 3 Pro
Loading Table 6 comparison…
One measured point per model; dashed lines are reading guides. Recall axis: 94–96%. Recall counts crops retaining >75% of the target. Runtime is per VLM stage, not the 4.1s full pipeline.
View all Table 6 values
| Frozen VLM | Area retained (%) | Target recall (%) | Runtime (s / stage) |
|---|---|---|---|
| Gemini 3 Pro | 12.1 | 95.7 | 13.1 |
| ChatGPT-5 | 11.7 | 95.3 | 11.2 |
| Qwen2.5-VL-72B-Instruct | 16.3 | 95.1 | 12.6 |
| Qwen2.5-VL-3B-Instruct (Ours) | 15.1 | 94.7 | 1.4 |
A closer look · Three cases
Qualitative Results
Follow the language. Focus the scene. Find the target.
Click or tap images to enlarge · Use ← → when focused
05 · Limitation
Early errors
can travel downstream.
A three-stage pipeline can propagate errors.
Low-quality renderings, occlusion, incomplete point clouds, or mistakes in LGSP and MCDR can remove the target or distort its relations. Those early errors may then be amplified by the final grounder.
Read the failure analysis06 · Citation
Build on
PruneGround.
If this work helps your research, please cite the arXiv paper and visit the official repository for code and models.
@article{dinh2026pruneground,
title = {PruneGround: Plug-and-play Spatial Pruning
for 3D Visual Grounding},
author = {Dinh, Duc Cao and Le-Duc, Khai and Draye,
Florent and Ngo, Chris and Zhang, Terry Jingchen
and Sch{\"o}lkopf, Bernhard and Jin, Zhijing},
journal = {arXiv preprint arXiv:2606.31148},
year = {2026}
}