WanderBench: Learning to Wander

Improving the Global Image Geolocation Ability of LMMs via Actionable Reasoning

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:

{
  "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:

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

See the code repository for instructions on running GeoAoT exploration, baseline prediction, and batch geocoding.

Citation

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

@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 license.