WangYipu2002/CrossPoint-Bench

---
task_categories:
- visual-question-answering
- image-to-text
language:
- en
size_categories:
- 1K<n<10K
pretty_name: CrossPoint-Bench
tags:
- cross-view
- spatial-understanding
configs:
- config_name: default
data_files: CrossPoint-Bench.jsonl
---

# CrossPoint-Bench

[](https://arxiv.org/abs/2512.04686)
[](https://github.com/WangYipu2002/CrossPoint)

<p align="center">
<img src="CrossPoint-Bench.png" alt="CrossPoint-Bench Overview" width="70%">
</p>

**CrossPoint-Bench** is a comprehensive benchmark for evaluating Vision-Language Models (VLMs) on cross-view point correspondence tasks. It assesses models' abilities to spatial understanding, and correspondence between different viewpoints.

### Dataset Structure

```
CrossPoint-Bench/
├── CrossPoint-Bench.jsonl # Main benchmark data file
└── image/
├── origin_image/ # Original scene images organized by scene ID
│ ├── scene0000_02/
│ ├── 0bd6b209/
│ └── ...
└── visual_image/ # Annotated visualization images
├── scene0000_02/
├── 0bd6b209/
└── ...
```

### Data Fields

Each instance in `CrossPoint-Bench.jsonl` contains:
- `idx`: Unique identifier
- `type`: Task type (e.g., "Correspondence-Pointing", "Fine-grained Grounding")
- `images`: List of image paths
- `question`: Question text
- `answer`: Ground truth answer

### Task Types

CrossPoint-Bench contains **1,000 samples** across **4 task types**:

1. **Fine-grained Grounding** (161 samples): Input an image and an instruction; output the coordinates of the referred target.

2. **Visibility Reasoning** (220 samples): Input an image and a point; output whether the point is visible in another view.

3. **Correspondence-Judgement** (156 samples): Input an image and a point; select the correct correspondence from multiple candidates in another view.

4. **Correspondence-Pointing** (463 samples): Input an image and a point; predict the exact coordinates of the corresponding point in another view.

### Evaluation

For evaluation scripts and detailed instructions, please visit the [CrossPoint GitHub repository](https://github.com/WangYipu2002/CrossPoint).

## Citation

If you find CrossPoint-Bench useful for your research, please cite:

```bibtex
@article{wang2025crosspoint,
title={Towards Cross-View Point Correspondence in Vision-Language Models},
author={Wang, Yipu and Ji, Yuheng and Liu, Yuyang and Zhou, Enshen and Yang, Ziqiang and Tian, Yuxuan and Qin, Ziheng and Liu, Yue and Tan, Huajie and Chi, Cheng and Ma, Zhiyuan and Zeng, Daniel Dajun and Zheng, Xiaolong},
journal={arXiv preprint arXiv:2512.04686},
year={2025}
}
```

## Contact

For questions or issues, please open an issue on our [GitHub repository](https://github.com/WangYipu2002/CrossPoint).