File size: 4,668 Bytes
ef653eb
 
 
 
 
 
 
 
 
 
 
40c36b1
ef653eb
a5ca571
3f99af6
 
 
 
ef653eb
 
 
 
 
a5ca571
 
ccb89c4
 
 
 
d89f333
 
 
11272db
 
 
 
 
 
 
3dcdcc1
11272db
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
d89f333
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: question
    dtype: string
  - name: answer
    sequence: int64
  splits:
  - name: valid
    num_bytes: 7094843893.125
    num_examples: 14631
  - name: train
    num_bytes: 140854221157.57
    num_examples: 289911
  download_size: 51389456693
  dataset_size: 147949065050.695
configs:
- config_name: default
  data_files:
  - split: valid
    path: data/valid-*
  - split: train
    path: data/train-*
license: cc-by-4.0
language:
- en
pretty_name: QA Patches Task Dataset
task_categories:
- image-text-to-text
- visual-question-answering
---
# Dataset Card for Patch-Based Visual Question Answering Dataset

## Dataset Details

### Dataset Description

This dataset contains approximately 305,000 triplets of `question`, `answer`, and `image` designed for patch-based visual reasoning tasks.  

A standard question in this dataset is formatted as follows:

> Image Grid: The image is divided into a 4x4 grid of 16 equal-sized patches. Patches are numbered sequentially from the top-left corner and moving right, then down to the next row.  
> Task: Identify the patch number(s) that contain a potted plant.  
> Response Format: Provide only the relevant patch number(s) as a list (e.g., [3], [5, 12], or [] if none are found).

The dataset is built on top of **COCO-2017**, from which object bounding boxes (bboxes) are used to generate questions and answers.

- **Curated by:** Yurii Potapov
- **Language(s) :** English
- **License:** Annotations and code: CC BY 4.0 (COCO), Images: Flickr Terms of Use

### Dataset Sources

- **Repository:** [Not yet published]
- **Paper:** [Not yet published]
- **Demo:** [More Information Needed]

## Uses

### Direct Use

- Training and evaluating **visual-language models (VLMs)** or other multimodal models.  
- Patch-based object detection and reasoning.  
- Research in **image question answering**, **visual reasoning**, and **multimodal representation learning**.

### Out-of-Scope Use

- Direct commercial redistribution of original COCO images without following Flickr Terms of Use.  
- Use cases where original images are required to be displayed in full, due to copyright restrictions.  

## Dataset Structure

- **question**: A textual description of the task referring to a 4x4 patch grid.  
- **answer**: List of integers representing the patch indices containing the target object(s).  
- **image**: Corresponding COCO-2017 image (PIL Image object or file path).  

The dataset contains no explicit splits; users can generate their own train/validation/test splits as needed.

## Dataset Creation

### Curation Rationale

The dataset was created to facilitate **patch-level visual question answering** and to improve the training of visual-language models using real-world images with structured spatial queries.

### Source Data

The dataset is based on COCO-2017 images and annotations. Bounding boxes from COCO are used to determine which patches contain specific objects (e.g., potted plants).

#### Data Collection and Processing

- Images are sourced from COCO-2017 (Flickr) respecting their Terms of Use.  
- Bounding boxes from COCO are used to automatically generate 4x4 grid questions.  
- Each question asks which patch(es) contain a specific object.  
- Answers are stored as lists of patch indices.  

#### Who are the source data producers?

Original images were contributed by Flickr users and annotated by the COCO Consortium.

### Annotations

#### Annotation process

Annotations (bounding boxes) are sourced from COCO-2017. Patch assignments and questions were automatically generated programmatically based on bounding box locations.

#### Who are the annotators?

Annotations are from COCO annotators; patch-level questions are generated automatically.

#### Personal and Sensitive Information

The dataset does **not contain personal or sensitive information**.

## Bias, Risks, and Limitations

- Images reflect the distribution and biases present in COCO-2017.  
- Models trained on this dataset may inherit biases from the original dataset.  
- Limited to the objects annotated in COCO-2017.  

### Recommendations

Users should be aware of the **copyright limitations of the original images** and provide attribution for COCO annotations. Use transformed or model-generated outputs rather than raw images for publication if possible.

## Glossary

- **Patch**: One of 16 equally sized blocks in a 4x4 grid over an image.  
- **VLM (Visual-Language Model)**: A model that learns joint representations of images and text.  

## Dataset Card Authors

Yurii Potapov

## Dataset Card Contact

yurii.a.potapov@gmail.com