Omarrran commited on
Commit
2ece2f7
·
verified ·
1 Parent(s): 6c90b9a

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +72 -3
README.md CHANGED
@@ -1,3 +1,72 @@
1
- ---
2
- license: apache-2.0
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ ---
4
+
5
+ # Vision-Language Pairs Dataset
6
+
7
+ This dataset contains metadata about image-text pairs from various popular vision-language datasets.
8
+
9
+ ## Contents
10
+
11
+ - **vision_language_data/all_vision_language_images.csv**: Combined metadata for all images (75629 records)
12
+ - **vision_language_data/all_vision_language_captions.csv**: Combined captions for all images (86676 records)
13
+ - **dataset_statistics.csv**: Summary statistics for each dataset
14
+ - **category_distribution.csv**: Distribution of image categories across datasets
15
+ - **caption_length_distribution.csv**: Distribution of caption lengths
16
+ - **caption_style_distribution.csv**: Distribution of caption styles
17
+ - **category_caption_statistics.csv**: Caption statistics by category
18
+ - **vision_language_catalog.json**: Searchable catalog with sample image-caption pairs
19
+
20
+ ## Datasets Included
21
+
22
+ - **COCO** (Common Objects in Context): COCO is a large-scale object detection, segmentation, and captioning dataset with multiple captions per image. (123287 images)
23
+ - **Flickr30K** (Flickr 30,000 Images): Flickr30K contains images collected from Flickr with 5 reference captions per image provided by human annotators. (31783 images)
24
+ - **Visual Genome** (Visual Genome): Visual Genome connects structured image concepts to language with detailed region descriptions and question-answer pairs. (108077 images)
25
+ - **Conceptual Captions** (Conceptual Captions): Conceptual Captions is a large-scale dataset of image-caption pairs harvested from the web and automatically filtered. (3300000 images)
26
+ - **CC3M** (Conceptual 3 Million): CC3M is a dataset of 3 million image-text pairs collected from the web, useful for vision-language pretraining. (3000000 images)
27
+ - **SBU Captions** (SBU Captioned Photo Dataset): The SBU dataset consists of 1 million images with associated captions collected from Flickr. (1000000 images)
28
+
29
+ ## Fields Description
30
+
31
+ ### Images Table
32
+ - **image_id**: Unique identifier for the image
33
+ - **dataset**: Source dataset name
34
+ - **image_url**: URL to the image (simulated)
35
+ - **primary_category**: Main content category
36
+ - **width**: Image width in pixels
37
+ - **height**: Image height in pixels
38
+ - **aspect_ratio**: Width divided by height
39
+ - **caption_count**: Number of captions for this image
40
+ - **license**: License under which the image is available
41
+
42
+ ### Captions Table
43
+ - **caption_id**: Unique identifier for the caption
44
+ - **image_id**: ID of the associated image
45
+ - **dataset**: Source dataset name
46
+ - **text**: Caption text
47
+ - **language**: Caption language (default: en)
48
+ - **style**: Caption style (descriptive, short, or detailed)
49
+ - **length**: Number of characters in the caption
50
+ - **word_count**: Number of words in the caption
51
+
52
+ ## Usage Examples
53
+
54
+ This metadata can be used for:
55
+
56
+ 1. Analyzing the composition of vision-language datasets
57
+ 2. Comparing caption characteristics across different datasets
58
+ 3. Training and evaluating image captioning models
59
+ 4. Studying linguistic patterns in image descriptions
60
+ 5. Developing multimodal AI systems
61
+
62
+ ## Data Generation Note
63
+
64
+ This dataset contains synthetic metadata that represents the structure and
65
+ characteristics of actual vision-language pair collections, but the specific
66
+ image and caption details are generated for demonstration purposes.
67
+
68
+ Created: 2025-04-26
69
+
70
+
71
+ ## Note:
72
+ ## All files are packaged into a ZIP archive called vision_language_pairs_data.zip for easy download, with expected size in the 150-200MB range, making it suitable for research and educational purposes.