Yushuo-Zheng/WanderBench

---
license: cc-by-4.0
task_categories:
- image-to-text
- visual-question-answering
- other
tags:
- geolocation
- street-view
- navigation
- embodied-reasoning
- panorama
- benchmark
- geoaot
- cvpr2026
language:
- en
pretty_name: "WanderBench: Global Geolocation Benchmark for Actionable Reasoning"
size_categories:
- 1K<n<10K
---

# WanderBench: Learning to Wander

### Improving the Global Image Geolocation Ability of LMMs via Actionable Reasoning

<p align="center">
<a href="https://arxiv.org/abs/2603.10463"><img src="https://img.shields.io/badge/arXiv-2603.10463-b31b1b.svg" alt="arXiv"></a>
<a href="#"><img src="https://img.shields.io/badge/CVPR%20Findings-2026-4b44ce.svg" alt="CVPR Findings 2026"></a>
<a href="https://creativecommons.org/licenses/by/4.0/"><img src="https://img.shields.io/badge/License-CC%20BY%204.0-lightgrey.svg" alt="License: CC BY 4.0"></a>
</p>

## Dataset Description

**WanderBench** is the first open-access global geolocation benchmark designed for actionable geolocation reasoning in embodied scenarios. It contains **1,049 navigable panorama graphs** comprising **32,776 panoramic nodes** distributed across six continents. Each graph encodes spatial relationships between street-view panoramas, enabling multi-step interactive exploration for geolocation tasks.

This dataset accompanies the paper *"Learning to Wander: Improving the Global Image Geolocation Ability of LMMs via Actionable Reasoning"* (CVPR Findings 2026).

## Dataset Statistics

| Statistic | Value |
|-----------|-------|
| Total graphs | 1,049 |
| Total panorama nodes | 32,776 |
| Max graph size | 30 nodes |
| Navigation steps per graph | up to 10 |
| Geographic coverage | 6 continents |
| Latitude range | ~-43° to ~54° |
| Longitude range | ~-123° to ~154° |

## Data Format

Each file is a JSON graph named `{pano_id}_10_graph.json` with the following structure:

```json
{
"center_pano_id": "pano_id_123",
"max_steps": 10,
"nodes": [
{
"pano_id": "pano_id_123",
"matrix_id": 0,
"coordinate": {
"lat": 40.7128,
"lon": -74.0060,
"heading": 1.5708,
"roll": 0.017,
"pitch": -0.058
}
}
],
"adjacency_matrix": [
[-1, 1.57, 0.0],
[4.71, -1, 1.57],
[3.14, 4.71, -1]
]
}
```

### Fields

- **`center_pano_id`** — Google Street View panorama ID of the starting node.
- **`max_steps`** — Maximum navigation steps allowed in the graph.
- **`nodes`** — List of panorama nodes, each containing:
- `pano_id` — Street View panorama identifier.
- `matrix_id` — Index in the adjacency matrix.
- `coordinate` — GPS location (`lat`, `lon` in degrees) and camera orientation (`heading`, `roll`, `pitch` in radians).
- **`adjacency_matrix`** — Directional angles (radians) between connected nodes; `-1` indicates no direct connection.

## Usage

This dataset provides the graph structure for the WanderBench benchmark. To run evaluations, clone the companion code repository:

```bash
git clone https://github.com/YushuoZheng/WanderBench.git
```

See the [code repository](https://github.com/YushuoZheng/WanderBench) for instructions on running GeoAoT exploration, baseline prediction, and batch geocoding.

## Citation

If you find WanderBench useful in your research, please cite:

```bibtex
@misc{zheng2026learningwanderimprovingglobal,
title={Learning to Wander: Improving the Global Image Geolocation Ability of LMMs via Actionable Reasoning},
author={Yushuo Zheng and Huiyu Duan and Zicheng Zhang and Xiaohong Liu and Xiongkuo Min},
year={2026},
eprint={2603.10463},
archivePrefix={arXiv},
primaryClass={cs.CV},
url={https://arxiv.org/abs/2603.10463},
}
```

## License

This dataset is released under the [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/) license.