🌍 CVSBench: Cross-View Spatial Reasoning and Dreaming Benchmark

📄 Paper | 🌐 Project Page | 🤗 Dataset | 💻 GitHub

🔗 Links

Resource	Link
📑 Paper	Arxiv
🌐 Project Page	earth-insights/CVSBench
💻 GitHub	earth-insights/CVSBench
🤗 Dataset (Hugging Face)	zlyzlyzly/CVSBench

---

🧭 What Is In This Dataset

CVSBench is a benchmark for evaluating whether vision-language models can reason across views, align satellite and street-view observations, localize corresponding objects, and imagine unseen visual content from partial evidence.

This repository contains the released dataset splits and task files. It focuses on cross-view understanding between:

• satellite imagery
• street-view imagery

The benchmark is designed to test abilities such as:

• cross-view correspondence
• spatial reasoning
• grounding and localization
• viewpoint understanding
• visual imagination

Unlike traditional cross-view benchmarks that mainly focus on retrieval or recognition, CVSBench explicitly emphasizes:

• cross-view spatial reasoning
• cross-view grounding
• view understanding and matching
• visual imagination from partial observations

---

✨ Benchmark Highlights

• Covers both satellite-to-ground and ground-to-satellite reasoning.
• Includes QA-style, grounding-style, and view-matching tasks.
• Contains multiple benchmark subsets rather than a single uniform task setting.
• Extends beyond recognition and matching to evaluate spatial reasoning and visual imagination.

---

🧩 Tasks

CVSBench currently contains two major subsets:

• cvusa/
• fov/

These subsets are complementary rather than identical, and they contain different task families.

cvusa/

• g2s: Ground-to-Satellite reasoning tasks
• s2g: Satellite-to-Ground reasoning tasks
• gs_grounding: cross-view object grounding and bounding-box localization tasks

fov/

• data: raw image and supporting resources
• g2s: Ground-to-Satellite reasoning tasks
• s2g: Satellite-to-Ground reasoning tasks
• gs_grounding: cross-view object grounding and bounding-box localization tasks
• gs_view: cross-view view-matching tasks with two settings:
  • View-Arrow: given a directional arrow or viewing direction, find the corresponding street-view image
  • View-Image: given a street-view image, find the corresponding directional arrow or viewing direction
• nanobanana: generated 3D miniature building-model images used as auxiliary visual inputs for FOV-based visual imagination experiments

---

🗂️ Dataset Structure

The released dataset is organized as follows:

CVSBench/
├── cvusa/
│   ├── data/
│   ├── g2s/
│   ├── s2g/
│   └── gs_grounding/
└── fov/
    ├── data/
    ├── g2s/
    ├── s2g/
    ├── gs_grounding/
    ├── gs_view/
    └── nanobanana/

Typical files include:

• train / test JSONL annotations
• task-specific metadata
• image path references
• grounding annotations for bbox evaluation

---

📝 Example Data Format

Below is a simplified example from a grounding-style task:

{
  "img_id": "0001119_0",
  "task": "Ground2Sat",
  "source_image": "cvusa/data/streetview/0001119.jpg",
  "target_image": "cvusa/data/bingmap/input0001119.png",
  "target_bbox": [121.0, 196.6, 153.0, 234.6],
  "questions": [
    {
      "level": 3,
      "question": "First image shows a street-view with a bounding box. In the second satellite image, provide the pixel bounding box coordinates [x_min, y_min, x_max, y_max] for the corresponding object.",
      "answer": [121.0, 196.6, 153.0, 234.6]
    }
  ],
  "dataset": "cvusa"
}

Exact fields may vary across task families.

---

⬇️ Download

You can download the dataset with:

huggingface-cli download zlyzlyzly/CVSBench \
    --repo-type dataset \
    --local-dir data/CVSBench

---

🚀 Usage

CVSBench can be used for:

• benchmarking vision-language models on cross-view reasoning
• evaluating object grounding across satellite and street-view images
• studying viewpoint alignment and cross-view matching
• studying visual imagination from sparse or partial observations

Official code and evaluation scripts are available in the GitHub repository:

• GitHub: earth-insights/CVSBench
• Dataset: zlyzlyzly/CVSBench
• Project Page: earth-insights/CVSBench

---

📚 Recommended Citation

If you use CVSBench in your research, please cite:

@article{cvsbench2026,
  title={CVSBench: A Comprehensive Benchmark for Cross-View Spatial Reasoning and Dreaming},
  author={[TBD]},
  journal={ECCV},
  year={2026}
}

The official citation will be updated after the paper metadata is finalized.

---

⚖️ License

License: CC-BY-4.0

If parts of the dataset inherit licensing or usage constraints from underlying sources, please also follow the original source terms.

---

🙏 Acknowledgements

CVSBench is built on top of existing cross-view data resources. We sincerely thank the creators and maintainers of the following datasets and projects:

• CVUSA
• University-1652
• FOV and cross-view benchmark contributors
• open-source vision-language model and benchmark tooling communities

---

📮 Contact

For questions, issues, or collaboration requests:

• Email: zlyzly@stu.xjtu.edu.cn