PanoEnv-QA: A Large-Scale Geometry-Grounded Panoramic VQA Benchmark for 3D Spatial Intelligence

📖 Overview

PanoEnv-QA is a large-scale Visual Question Answering benchmark designed specifically to probe 3D spatial intelligence on Equirectangular Projection (ERP) panoramas. Built from synthetic but photorealistic 3D environments (TartanAir), PanoEnv-QA offers over 14.8K questions spanning five categories that progressively require stronger 3D understanding—all grounded in precise 3D annotations (depth, semantics, and 3D bounding boxes).

Key Features

• Geometry-Grounded: All QA pairs are programmatically derived from physical ground truth (depth maps, semantic segmentation, 3D bounding boxes)
• 360° Panoramic: Targets the unique challenges of ERP images including geometric distortions and multi-view reasoning
• RL-Ready: Designed to serve both as a reliable evaluation benchmark and as a source of verifiable supervision signals for reinforcement learning
• Diverse & Balanced: 60 diverse environments with balanced question distribution across 5 major categories

📊 Dataset Statistics

Split	Environments	Images	QA Pairs
Train	60	415	10,340
Val	60	60	1,496
Test	60	120	2,991
Total	60	595	14,827

Question Distribution

Major Category	# Questions	Percentage
Intrinsic Attribute Comparison	2,975	20.1%
Object Distance Estimation	2,975	20.1%
Relative Spatial Positioning	2,975	20.1%
Environment Identification	2,965	20.0%
Camera View Source Identification	2,937	19.8%

Question Types

Type	Count	Percentage
Multiple Choice	7,552	50.9%
True/False	4,300	29.0%
Open-Ended	2,975	20.1%

Answer Characteristics

• 1,894 unique answers
• Average answer length: 10.9 characters
• Yes/No ratio: 45.3% / 54.7% (balanced to prevent shortcuts)

🎯 Five Question Categories

① Camera View Source Identification

Evaluates whether the model recognizes that an ERP image is a composite panorama stitched from six perspective views (front/back/left/right/top/bottom). Understanding this structure is essential for handling artifacts near seam boundaries.

Sub-categories: primary_view, multi_view_visibility, seam_attribution, multi_object_relationship, shared_visibility

② Object Distance Estimation

Evaluates quantitative and qualitative depth reasoning, moving beyond 2D heuristics (e.g., size as a proxy for distance) toward true 3D understanding.

Sub-categories: depth_similarity, depth_binary, depth_compare, depth_triplet_farthest, distance_description

③ Environment Identification

Evaluates high-level scene understanding and contextual reasoning, testing whether the model can classify environments based on object composition and architectural style.

Sub-categories: env_binary_judgement, env_mcq, env_confusable_pair, env_scene_judgement, env_category_identification, env_attribute

④ Relative Spatial Positioning

Assesses the model's ability to reconstruct accurate 3D spatial relationships between objects—an inherently difficult task due to ERP distortions.

Sub-categories: relpos_cardinal, relpos_binary, relpos_distance_straightline, relpos_distance_components, relpos_triplet_extreme

⑤ Intrinsic Attribute Comparison

Probes the model's understanding of intrinsic, view-independent physical properties of objects (3D shape and size), requiring inference from 2D projections and depth.

Sub-categories: volume_comparison, volume_binary, size_triplet_extreme, shape_flatness, shape_elongation

📁 Data Structure

Each sample (*_qa.json) contains:

sampled_objects (20 objects per image)

{
  "label": "building",
  "bbox": [x1, y1, x2, y2],
  "depth": 12.5,
  "area": 15000,
  "primary_camera": "front",
  "visible_cameras": ["front", "left", "top"],
  "depth_stats": {
    "p20": 10.5, "p25": 11.0, "p50": 12.5, "p75": 14.0, "p80": 14.5, "iqr": 3.0
  },
  "bbox_3d": {
    "min_x": -5.0, "max_x": 5.0,
    "min_y": 0.0, "max_y": 10.0,
    "min_z": 8.0, "max_z": 15.0
  },
  "volume": 700.0,
  "centroid_3d": [0.0, 5.0, 11.5],
  "is_seam": true,
  "seam_types": ["crosses_left_back"],
  "is_polar": false
}

questions (25 questions per image)

{
  "major_category": "relative_position",
  "sub_category": "relpos_cardinal",
  "question_type": "open_ended",
  "question": "What is the spatial relationship of the building relative to the tree in the 3D world?",
  "answer": "The building is in front of and to the right of and above the tree.",
  "related_object_ids": [1, 5],
  "question_id": 1
}

visualizations/

PNG visualizations for each question showing the relevant objects highlighted.

🌍 60 Diverse Environments

Click to expand full environment list

Industrial & Infrastructure

• AbandonedCable, AbandonedFactory, AbandonedFactory2, CarWelding, CoalMine, ConstructionSite, FactoryWeather, IndustrialHangar, OldIndustrialCity, Sewerage, UrbanConstruction

Urban & City

• CyberPunk, CyberPunkDowntown, Downtown, HongKong, JapaneseAlley, JapaneseCity, ModernCityDowntown, ModularNeighborhood, ModularNeighborhoodIntExt, ModUrbanCity, Rome, SoulCity, VictorianStreet

Historical & Cultural

• AncientTowns, Antiquity3D, CastleFortress, GothicIsland, HQWesternSaloon, MiddleEast, OldTownFall, OldTownNight, OldTownSummer, OldTownWinter, Ruins, WesternDesertTown

Residential & Interior

• AmericanDiner, ArchVizTinyHouseDay, ArchVizTinyHouseNight, CountryHouse, Hospital, House, Office, OldBrickHouseDay, OldBrickHouseNight, Prison, Restaurant, RetroOffice, Supermarket

Nature & Special

• AbandonedSchool, AmusementPark, Apocalyptic, DesertGasStation, Fantasy, NordicHarbor, Ocean, PolarSciFi, SeasideTown, WaterMillDay, WaterMillNight

🔬 Benchmark Results

We evaluated 14 state-of-the-art VLMs on our test set:

Model	Total Acc. (%)	T/F (%)	MCQ (%)	OE (%)
Qwen2.5-VL-7B	49.34	65.19	57.24	6.39
Qwen2.5-VL-32B	42.70	62.47	44.96	8.36
InternVL2.5-26B	47.07	64.51	54.33	3.44
Qwen3-VL-8B	47.91	62.85	55.24	7.70
DeepSeek-VL2-Base	38.86	57.30	40.36	8.36
GRPO-Balanced (Ours)	52.93	68.78	58.90	14.83

Key Findings:

• Best zero-shot accuracy is only 49.34%, revealing significant gaps in 3D spatial understanding
• Open-ended accuracy collapses to < 9% for all baselines
• Our GRPO-trained 7B model achieves SoTA performance, outperforming 32B models
• OE accuracy improves from 6.39% to 14.83% (+132% relative gain)

🚀 Usage

Loading the Dataset

from datasets import load_dataset

# Load the full dataset
dataset = load_dataset("7zkk/PanoEnv")

# Access splits
train_data = dataset["train"]
val_data = dataset["val"]
test_data = dataset["test"]

Example: Accessing a Sample

sample = dataset["train"][0]
print(f"Environment: {sample['env']}")
print(f"Image ID: {sample['image_id']}")
print(f"Number of objects: {len(sample['sampled_objects'])}")
print(f"Number of questions: {len(sample['questions'])}")

For RL Training (GRPO)

PanoEnv-QA is designed to support reinforcement learning with ground-truth-guided rewards:

# Example reward routing based on question type
def get_reward(question_type, prediction, ground_truth):
    if question_type == "true_false":
        return 1.0 if prediction.lower() == ground_truth.lower() else 0.0
    elif question_type == "multiple_choice":
        return mcq_matching_reward(prediction, ground_truth)
    elif question_type == "open_ended":
        return spatial_reward(prediction, ground_truth)  # axis-wise matching

📐 Technical Details

ERP to 3D Projection

For any pixel $(p_x, p_y)$ in an ERP image of size $W \times H$:

Spherical coordinates: $$\lambda =(\frac{p_x}{W}-0.5)\cdot 2\pi ,\varphi =-(\frac{p_y}{H}-0.5)\cdot \pi \backslash lambda\; =\; \backslash left(\backslash frac\{p\_x\}\{W\}\; -\; 0.5\backslash right)\; \backslash cdot\; 2\backslash pi,\; \backslash quad\; \backslash phi\; =\; -\backslash left(\backslash frac\{p\_y\}\{H\}\; -\; 0.5\backslash right)\; \backslash cdot\; \backslash pi$$

3D Cartesian coordinates: $$x=-d\cdot \cos (\varphi )\cdot \sin (\lambda )x\; =\; -d\; \backslash cdot\; \backslash cos(\backslash phi)\; \backslash cdot\; \backslash sin(\backslash lambda)$$ $$y=d\cdot \sin (\varphi )y\; =\; d\; \backslash cdot\; \backslash sin(\backslash phi)$$ $$z=-d\cdot \cos (\varphi )\cdot \cos (\lambda )z\; =\; -d\; \backslash cdot\; \backslash cos(\backslash phi)\; \backslash cdot\; \backslash cos(\backslash lambda)$$

where $d$ is the depth value. The coordinate system is right-handed with +Y upward, +X rightward, and –Z forward.

📚 Citation

If you use PanoEnv-QA in your research, please cite:

@inproceedings{panoenv2026,
  title={PanoEnv-QA: A Large-Scale Geometry-Grounded Panoramic VQA Benchmark for 3D Spatial Intelligence},
  author={Zekai Lin, Xu Zheng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}

🙏 Acknowledgments

This dataset is built upon TartanAir, a synthetic dataset providing precise 3D ground truth (depth and segmentation).

📄 License

This dataset is released under the CC BY 4.0 license.