ECCV 2026

PanoGrounder: Bridging 2D and 3D with Panoramic Scene Representations for VLM-based 3D Visual Grounding

1Seoul National University, 2Robotics Lab, Hyundai Motor Company, 3Pohang University of Science and Technology (POSTECH)
* Equal contribution
PanoGrounder teaser

PanoGrounder localizes referred objects in 3D scenes by leveraging panoramic renderings as a 2D–3D interface.

Explore Results

Select a scene and a text query to explore PanoGrounder's grounding. The left view shows the predicted 3D box (green) against the ground truth (blue), while the right view shows the per-camera panoramas with predicted 2D boxes and SAM masks.

Abstract

3D Visual Grounding (3DVG) is a critical bridge from vision-language perception to robotics, requiring both language understanding and 3D scene reasoning. Traditional supervised models leverage explicit 3D geometry but exhibit limited generalization, owing to the scarcity of 3D vision-language datasets and the limited reasoning capabilities compared to modern vision-language models (VLMs).

We propose a generalizable 3DVG framework, PanoGrounder, that couples multi-modal panoramic representation with pretrained 2D VLMs for strong vision-language reasoning. Panoramic renderings, augmented with 3D semantic and geometric features, serve as an intermediate representation between 2D and 3D, and offer two major benefits: (i) they can be directly fed to VLMs with minimal adaptation and (ii) they retain long-range object-to-object relations thanks to their 360-degree field of view. We devise a three-stage pipeline that places a compact set of panoramic viewpoints considering the scene layout and geometry, grounds a text query on each panoramic rendering with a VLM, and fuses per-view predictions into a single 3D bounding box via lifting.

Our approach achieves state-of-the-art results on ScanRefer and Nr3D, and demonstrates strong generalization to unseen 3D datasets and text rephrasings.

The PanoGrounder Framework

PanoGrounder pipeline

Given a renderable 3D scene (a triangle mesh or 3D Gaussian Splatting) and a text query, PanoGrounder predicts the 3D bounding box of the referred object, using panoramic renderings as a 2D–3D interface. The pipeline runs in three stages.

1. Structure-Aware Camera Placement

Since panoramas are omnidirectional, we only choose camera locations. We greedily select a compact viewpoint set—scored by ray coverage, distance-to-surface, and distance-to-trajectory—covering each scene with only a few cameras.

2. Multi-Modal Panoramic VLM

At each viewpoint we render RGB, geometric (range), and semantic feature panoramas. The geometric and semantic cues are injected into selected VLM layers through a lightweight Feature Adapter (a 2-layer MLP + a zero-initialized 1×1 convolution), preserving the VLM's pretrained behavior while adding task-relevant signal. The VLM then predicts a 2D box on each panorama. Training uses cross-entropy with an Earth Mover's Distance (EMD) loss for distance-aware supervision.

3. Visibility-Aware 3D Aggregation

Per-view 2D predictions are lifted to metric 3D using rendered depth and camera extrinsics. We pick the most reliable anchor view by cross-view consistency, fuse the visible point sets, and fit an axis-aligned 3D box.

Quantitative Results

3D Visual Grounding on Nr3D, Sr3D, and ScanRefer

PanoGrounder achieves state-of-the-art or top results across benchmarks. All methods here are trained on a single benchmark. ScanRefer reports Acc@0.25 (IoU); ReferIt3D (Nr3D/Sr3D) reports Top-1 accuracy.

Method Nr3D Sr3D ScanRefer
EasyHardVDVIDOverall EasyHardVDVIDOverall UniqueMultipleOverall
BUTD-DETR 60.748.446.058.054.6 68.663.253.067.667.0 84.246.652.2
ViL3DRel 70.257.462.064.564.4 74.967.963.873.272.8 81.640.347.9
3D-VisTA 65.949.453.759.457.5 72.163.657.970.169.6 77.438.745.9
MIKASA 69.759.465.464.064.4 78.667.370.475.475.2
MCLN 59.8 68.4 86.952.057.2
PQ3D 73.356.760.767.064.9 78.868.251.576.775.6 85.246.852.8
LIBA 57.260.364.5 70.261.775.8 88.854.459.6
VGMamba 61.468.3 74.481.3 91.954.860.0
ViewSRD 75.364.868.670.669.9 78.370.669.076.276.0 82.137.445.4
Ours 82.267.270.576.374.6 81.374.260.580.079.1 84.355.361.0

Table 1. Evaluation on Nr3D, Sr3D, and ScanRefer, with all methods trained on a single benchmark. VD/VID denote the View-Dependent / View-Independent splits. ScanRefer reports Acc@0.25 (IoU); ReferIt3D reports Top-1 accuracy. The Ours row is highlighted; bold marks the best in each column.

Zero-Shot on Nr3D

MethodBackbone EasyHardVDVIDOverall
ZSVG3DGPT-3.5-turbo46.531.736.840.039.0
SeeGroundQwen2-VL-72B54.538.342.348.246.1
VLM-GrounderGPT-4V55.239.545.849.448.0
OursCogVLM-17B64.242.749.854.753.2

Table 2. Zero-shot on Nr3D. Without any fine-tuning, ours surpasses methods with far larger backbones. Bold marks the best in each column.

Cross-Dataset Generalization

Method ScanRefer ARKitScenes 3RScan
uniq.mult.over. uniq.mult.over. uniq.mult.over.
ViL3DRel 92.051.859.6 57.221.128.3 71.831.336.8
3D-VisTA 89.549.957.2 59.726.632.9 74.132.037.7
BUTD-DETR* 92.552.658.5 66.330.636.1
MCLN* 93.454.960.6 61.230.035.3
Ours 91.758.564.9 74.248.053.5 80.437.743.8

Table 3. Trained on ScanRefer, evaluated on unseen ARKitScenes and 3RScan. All methods use GT object segmentation; * additionally uses GT semantic labels. All baseline results are re-run by us. Bold marks the best in each column.

Robustness to Text Variants

We probe linguistic robustness on ScanRefer by generating four variants of each query with LLaMA 3.3: Para. (paraphrased), +Aff. (an affordance/functional description added), +Aff.−N (+Aff. with the target noun removed), and Mask (target noun replaced with “object”). As explicit class cues disappear, PanoGrounder's advantage grows sharply—it must reason over spatial and contextual cues rather than lexical matching.

Method Org.Para.+Aff.+Aff.−NMask
ViL3DRel 59.641.59.76.522.9
3D-VisTA 57.252.950.428.328.5
BUTD-DETR* 58.553.455.135.527.3
MCLN* 60.654.655.735.731.5
PQ3D 66.063.358.635.833.3
Ours 64.961.561.351.741.8
Δ (Ours − PQ3D) −1.1−1.8 +2.7+15.9+8.5

Table 4. Robustness to ScanRefer text variants (Top-1 accuracy). Four variants are generated with LLaMA 3.3. All methods use GT object segmentation; * additionally uses GT semantic labels. All baseline results are re-run by us. Bold marks the best in each column; the last row is PanoGrounder's gain over PQ3D.

Qualitative Results

Representative success cases on ScanRefer across a variety of room types (kitchens, bathrooms, offices, and bedrooms). PanoGrounder's predicted boxes (red) align closely with the ground-truth boxes (green), even when the target object is small (e.g., (f), (h)). Many queries require relational cues such as “A next to B” or “A above B”, which PanoGrounder reliably parses and grounds to the correct instance in 3D.

Qualitative comparison of PanoGrounder predictions (red) with ground-truth boxes (green) on ScanRefer

BibTeX

@inproceedings{jungchoi2026panogrounder,
  title={PanoGrounder: Bridging 2D and 3D with Panoramic Scene Representations for VLM-based 3D Visual Grounding},
  author={Seongmin Jung and Seongho Choi and Gunwoo Jeon and Minsu Cho and Jongwoo Lim},
  booktitle={Proceedings of the European Conference on Computer Vision (ECCV)},
  year={2026}
}