ORAL EMNLP 2026

PruneGround:
Plug-and-play Spatial Pruning for 3D Visual Grounding

Reason over the right region, not the entire scene. PruneGround narrows the 3D search space, enriches the query with multi-view spatial cues, and grounds the target with a repurposed spatial LLM.

★ Co-first authors

Affiliations
1 Knovel Engineering Lab 2 University of Toronto 3 Vector Institute 4 MPI for Intelligent Systems, Tübingen 5 ELLIS Institute Tübingen 6 University of Oxford
#1

Overall grounding accuracy on ScanRefer, Nr3D and Sr3D among all methods compared in Tables 1–2.

9/10

Leading evaluation settings across Nr3D and Sr3D at IoU ≥ 0.5, covering easy/hard and viewpoint-dependent splits.

9.4×

Faster per VLM stage than Gemini 3 Pro: 1.4s vs. 13.1s, with just 1 percentage point lower target recall (94.7% vs. 95.7%).

Qwen2.5-VL-3B-Instruct · Table 6

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.

Full-scene reasoning

Clutter creates ambiguity

Same-class distractors and imperfect segmentation enlarge the candidate set, increase computation, and propagate errors.

PruneGround

Language narrows the search

A frozen VLM keeps only the relevant spatial region, then adds visual context to make the query easier to ground.

Figure 1: Intuition of our proposed method, inspired by human spatial cognition. Given a language description, the grounding model, as humans do, should infer the target object by attending only to spatially relevant regions guided by anchor objects and spatial relations, rather than reasoning over all candidate objects in the scene (e.g., all chairs).

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.

Interactive 3D — PruneGroundExplore the real scan

Same room. Different descriptions.

Several chairs look alike. Choose a description to see which one its spatial relationship identifies.

Room · scene0011_00
Full room · reference highlights
Loading 3D scene… Preparing mesh
A · chair near TV B · chair near doorway
Drag to orbit · scroll/pinch to zoom
Mesh statistics & data sources
Scenescene0011_00
Full triangles—
Pruned triangles—
Geometry retained—

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.

01Pruning

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
02Multi-view

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
03Spatial LLM

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
Figure 2: Overview of our PruneGround pipeline, consisting of three plug-and-play components: ① LGSP to remove irrelevant regions, ② MCDR to simplify the description while adding relational spatial information, and ③ LLM-Grounder that repurposes a detection-pretrained spatial LLM for language-conditioned grounding on the pruned scene and reformulated description.

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.

63.8%

ScanRefer overall accuracy at IoU ≥ 0.25, up 2.8 percentage points over PanoGrounder.

75.1%

Nr3D overall accuracy at IoU ≥ 0.5 on human-written descriptions with same-class distractors.

81.5%

Sr3D overall accuracy at IoU ≥ 0.5 on descriptions that distinguish objects through spatial relations.

Table 1 · ScanReferAccuracy (%) · ↑ higher is better
ScanRefer grounding accuracy from paper Table 1, page 7. Acc@0.25 and Acc@0.5 (IoU).
MethodVenueUniqueMultipleOverall
Acc@0.25Acc@0.5Acc@0.25Acc@0.5Acc@0.25Acc@0.5
ScanReferChen et al., 2020ECCV67.646.232.121.340.026.1
MVTHuang et al., 2022aCVPR77.766.531.925.340.833.3
3D-SPSLuo et al., 2022CVPR84.166.740.329.848.837.0
EDAWu et al., 2023bCVPR85.868.649.137.654.642.3
ConcreteNetUnal et al., 2023ECCV86.482.142.438.450.646.5
Chat-SceneHuang et al., 2023NeurIPS89.682.547.842.955.550.2
VPP-NetShi et al., 2024CVPR86.167.150.339.055.743.3
G³-LQWang et al., 2024CVPR88.673.350.239.756.044.7
MA²TransVGXu et al., 2024aCVPR86.374.153.841.457.945.7
MCLNQian et al., 2024ECCV86.972.752.040.857.245.5
PQ3DZhu et al., 2024bECCV86.778.351.546.257.051.2
VGMambaZhu et al., 2025bICCV91.979.654.847.660.053.9
PanoGrounderJung et al., 2025arXiv84.3—55.3—61.0—
PruneGroundOurs—90.284.757.649.463.856.1

Bold: best · Underline: second-best · —: not reported. Paper Table 1 ↗

Table 6 · Accuracy–speed trade-off

Small model.
Near-frontier target recall.

Compare the frozen VLMs used in LGSP and MCDR.

9.4× faster

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
Table 6: frozen VLM comparison on the pruning task.
Frozen VLMArea retained (%)Target recall (%)Runtime (s / stage)
Gemini 3 Pro12.195.713.1
ChatGPT-511.795.311.2
Qwen2.5-VL-72B-Instruct16.395.112.6
Qwen2.5-VL-3B-Instruct (Ours)15.194.71.4
Source: paper Table 6 · 512 × 512 top-view area ↗

A closer look · Three cases

Qualitative Results

Follow the language. Focus the scene. Find the target.

Chair · Spatial Relations01 / 04 · Original Scene

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 analysis
“Close to the TV” is relative: an incomplete description can leave the target ambiguous. Noisy views or incorrect pruning can compound that ambiguity and lead to an incorrect prediction.

06 · Citation

Build on
PruneGround.

If this work helps your research, please cite the arXiv paper and visit the official repository for code and models.

BibTeX
@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}
}

Expanded research figure