Upload 数据集说明文档
Browse files
README.md
ADDED
@@ -0,0 +1,138 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
task_categories:
|
4 |
+
- visual-question-answering
|
5 |
+
- image-classification
|
6 |
+
language:
|
7 |
+
- en
|
8 |
+
tags:
|
9 |
+
- visual-reasoning
|
10 |
+
- transformation
|
11 |
+
- benchmark
|
12 |
+
- computer-vision
|
13 |
+
size_categories:
|
14 |
+
- 1K<n<10K
|
15 |
+
---
|
16 |
+
|
17 |
+
# VisualTrans: A Benchmark for Real-World Visual Transformation Reasoning
|
18 |
+
|
19 |
+
[](http://arxiv.org/abs/2508.04043)
|
20 |
+
|
21 |
+
## Dataset Description
|
22 |
+
|
23 |
+
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.
|
24 |
+
|
25 |
+
## Dataset Statistics
|
26 |
+
|
27 |
+
- **Total samples**: 497
|
28 |
+
- **Number of manipulation scenarios**: 12
|
29 |
+
- **Task types**: 6
|
30 |
+
|
31 |
+
### Task Type Distribution
|
32 |
+
|
33 |
+
- **count**: 63 samples (12.7%)
|
34 |
+
- **procedural_causal**: 86 samples (17.3%)
|
35 |
+
- **procedural_interm**: 88 samples (17.7%)
|
36 |
+
- **procedural_plan**: 42 samples (8.5%)
|
37 |
+
- **spatial_fine_grained**: 168 samples (33.8%)
|
38 |
+
- **spatial_global**: 50 samples (10.1%)
|
39 |
+
|
40 |
+
|
41 |
+
### Manipulation Scenarios
|
42 |
+
|
43 |
+
The benchmark covers 12 diverse manipulation scenarios:
|
44 |
+
- Add Remove Lid
|
45 |
+
- Assemble Disassemble Legos
|
46 |
+
- Build Unstack Lego
|
47 |
+
- Insert Remove Bookshelf
|
48 |
+
- Insert Remove Cups From Rack
|
49 |
+
- Make Sandwich
|
50 |
+
- Pick Place Food
|
51 |
+
- Play Reset Connect Four
|
52 |
+
- Screw Unscrew Fingers Fixture
|
53 |
+
- Setup Cleanup Table
|
54 |
+
- Sort Beads
|
55 |
+
- Stack Unstack Bowls
|
56 |
+
|
57 |
+
|
58 |
+
## Dataset Structure
|
59 |
+
|
60 |
+
### Files
|
61 |
+
|
62 |
+
- `VisualTrans.json`: Main benchmark file containing questions, answers, and image paths
|
63 |
+
- `images.zip`: Compressed archive containing all images used in the benchmark
|
64 |
+
|
65 |
+
### Data Format
|
66 |
+
|
67 |
+
Each sample in the benchmark contains:
|
68 |
+
|
69 |
+
```json
|
70 |
+
{
|
71 |
+
"task_type": "what",
|
72 |
+
"images": [
|
73 |
+
"scene_name/image1.jpg",
|
74 |
+
"scene_name/image2.jpg"
|
75 |
+
],
|
76 |
+
"scene": "scene_name",
|
77 |
+
"question": "Question about the transformation",
|
78 |
+
"label": "Ground truth answer"
|
79 |
+
}
|
80 |
+
```
|
81 |
+
|
82 |
+
## Reasoning Dimensions
|
83 |
+
|
84 |
+
The framework evaluates three essential reasoning dimensions:
|
85 |
+
|
86 |
+
1. **Quantitative Reasoning** - Counting and numerical reasoning tasks
|
87 |
+
2. **Procedural Reasoning**
|
88 |
+
- **Intermediate State** - Understanding process states during transformation
|
89 |
+
- **Causal Reasoning** - Analyzing cause-effect relationships
|
90 |
+
- **Transformation Planning** - Multi-step planning and sequence reasoning
|
91 |
+
3. **Spatial Reasoning**
|
92 |
+
- **Fine-grained** - Precise spatial relationships and object positioning
|
93 |
+
- **Global** - Overall spatial configuration and scene understanding
|
94 |
+
|
95 |
+
## Usage
|
96 |
+
|
97 |
+
```python
|
98 |
+
import json
|
99 |
+
import zipfile
|
100 |
+
|
101 |
+
# Load the benchmark data
|
102 |
+
with open('VisualTrans.json', 'r') as f:
|
103 |
+
benchmark_data = json.load(f)
|
104 |
+
|
105 |
+
# Extract images
|
106 |
+
with zipfile.ZipFile('images.zip', 'r') as zip_ref:
|
107 |
+
zip_ref.extractall('images/')
|
108 |
+
|
109 |
+
# Access a sample
|
110 |
+
sample = benchmark_data[0]
|
111 |
+
print(f"Question: {sample['question']}")
|
112 |
+
print(f"Answer: {sample['label']}")
|
113 |
+
print(f"Images: {sample['images']}")
|
114 |
+
```
|
115 |
+
|
116 |
+
## Citation
|
117 |
+
|
118 |
+
If you use this benchmark, please cite our work:
|
119 |
+
|
120 |
+
```bibtex
|
121 |
+
@misc{ji2025visualtransbenchmarkrealworldvisual,
|
122 |
+
title={VisualTrans: A Benchmark for Real-World Visual Transformation Reasoning},
|
123 |
+
author={Yuheng Ji and Yipu Wang and Yuyang Liu and Xiaoshuai Hao and Yue Liu and Yuting Zhao and Huaihai Lyu and Xiaolong Zheng},
|
124 |
+
year={2025},
|
125 |
+
eprint={2508.04043},
|
126 |
+
archivePrefix={arXiv},
|
127 |
+
primaryClass={cs.CV},
|
128 |
+
url={https://arxiv.org/abs/2508.04043},
|
129 |
+
}
|
130 |
+
```
|
131 |
+
|
132 |
+
## License
|
133 |
+
|
134 |
+
This dataset is released under the MIT License.
|
135 |
+
|
136 |
+
## Contact
|
137 |
+
|
138 |
+
For questions or issues, please open an issue on our [GitHub repository](https://github.com/WangYipu2002/VisualTrans) or contact the authors.
|