README.md

---
license: mit
task_categories:
- visual-question-answering
language:
- en
tags:
- visual reason
- transformation
- benchmark
- computer vision
size_categories:
- 1K<n<10K
---

# VisualTrans: A Benchmark for Real-World Visual Transformation Reasoning

[![arXiv](https://img.shields.io/badge/arXiv-2508.04043-b31b1b.svg)](http://arxiv.org/abs/2508.04043)

## Dataset Description

VisualTrans is the first comprehensive benchmark specifically designed for Visual Transformation Reasoning (VTR) in real-world human-object interaction scenarios. The benchmark encompasses 12 semantically diverse manipulation tasks and systematically evaluates three essential reasoning dimensions through 6 well-defined subtask types.

## Dataset Statistics

- **Total samples**: 497
- **Number of manipulation scenarios**: 12
- **Task types**: 6

### Task Type Distribution

- **count**: 63 samples (12.7%)
- **procedural_causal**: 86 samples (17.3%)
- **procedural_interm**: 88 samples (17.7%)
- **procedural_plan**: 42 samples (8.5%)
- **spatial_fine_grained**: 168 samples (33.8%)
- **spatial_global**: 50 samples (10.1%)


### Manipulation Scenarios

The benchmark covers 12 diverse manipulation scenarios:
- Add Remove Lid
- Assemble Disassemble Legos
- Build Unstack Lego
- Insert Remove Bookshelf
- Insert Remove Cups From Rack
- Make Sandwich
- Pick Place Food
- Play Reset Connect Four
- Screw Unscrew Fingers Fixture
- Setup Cleanup Table
- Sort Beads
- Stack Unstack Bowls


## Dataset Structure

### Files

- `VisualTrans.json`: Main benchmark file containing questions, answers, and image paths
- `images.zip`: Compressed archive containing all images used in the benchmark

### Data Format

Each sample in the benchmark contains:

```json
{
    "task_type": "what",
    "images": [
        "scene_name/image1.jpg",
        "scene_name/image2.jpg"
    ],
    "scene": "scene_name",
    "question": "Question about the transformation",
    "label": "Ground truth answer"
}
```

## Reasoning Dimensions

The framework evaluates three essential reasoning dimensions:

1. **Quantitative Reasoning** - Counting and numerical reasoning tasks
2. **Procedural Reasoning**
   - **Intermediate State** - Understanding process states during transformation
   - **Causal Reasoning** - Analyzing cause-effect relationships
   - **Transformation Planning** - Multi-step planning and sequence reasoning
3. **Spatial Reasoning**
   - **Fine-grained** - Precise spatial relationships and object positioning
   - **Global** - Overall spatial configuration and scene understanding

## Usage

```python
import json
import zipfile

# Load the benchmark data
with open('VisualTrans.json', 'r') as f:
    benchmark_data = json.load(f)

# Extract images
with zipfile.ZipFile('images.zip', 'r') as zip_ref:
    zip_ref.extractall('images/')

# Access a sample
sample = benchmark_data[0]
print(f"Question: {sample['question']}")
print(f"Answer: {sample['label']}")
print(f"Images: {sample['images']}")
```

## Citation

If you use this benchmark, please cite our work:

```bibtex
@misc{ji2025visualtransbenchmarkrealworldvisual,
      title={VisualTrans: A Benchmark for Real-World Visual Transformation Reasoning}, 
      author={Yuheng Ji and Yipu Wang and Yuyang Liu and Xiaoshuai Hao and Yue Liu and Yuting Zhao and Huaihai Lyu and Xiaolong Zheng},
      year={2025},
      eprint={2508.04043},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2508.04043}, 
}
```

## License

This dataset is released under the MIT License.

## Contact

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