pengzhiliang commited on
Commit
67c8e38
·
1 Parent(s): d324bb2

update readme

Browse files
Files changed (1) hide show
  1. README.md +48 -3
README.md CHANGED
@@ -15,17 +15,62 @@ tags:
15
  task_categories:
16
  - text-to-image
17
  - image-to-text
18
- - zero-shot-classification
19
  - object-detection
 
20
  task_ids:
21
  - image-captioning
22
  - visual-question-answering
23
  ---
24
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
25
  ### Citation Information
26
- If you apply this dataset to any project and research, please cite our code and coyo-700m:
27
  ```
28
- @article{Kosmos22,
29
  title={Kosmos-2: Grounding Multimodal Large Language Models to the World},
30
  author={Zhiliang Peng and Wenhui Wang and Li Dong and Yaru Hao and Shaohan Huang and Shuming Ma and Furu Wei},
31
  journal={ArXiv},
 
15
  task_categories:
16
  - text-to-image
17
  - image-to-text
 
18
  - object-detection
19
+ - zero-shot-classification
20
  task_ids:
21
  - image-captioning
22
  - visual-question-answering
23
  ---
24
 
25
+ # GRIT: Large-Scale Training Corpus of Grounded Image-Text Pairs
26
+
27
+ ### Dataset Description
28
+ - **Repository:** [Microsoft unilm](https://github.com/microsoft/unilm/tree/master/kosmos-2)
29
+ - **Paper:** [Kosmos-2](https://arxiv.org/abs/2306.14824)
30
+ - **Point of Contact:** [Unilm team](fuwei@microsoft.com)
31
+
32
+ ### Dataset Summary
33
+ We introduce GRIT, a large-scale dataset of Grounded Image-Text pairs, which is created based on image-text pairs from [COYO-700M](https://github.com/kakaobrain/coyo-dataset) and LAION-2B. We construct a pipeline to extract and link text spans (i.e., noun phrases, and referring expressions) in the caption to their corresponding image regions. More details can be found in the [paper](https://arxiv.org/abs/2306.14824).
34
+
35
+ ### Supported Tasks
36
+ During the construction, we exclude the image-caption pair if no bounding boxes are retained. This procedure results in a high-quality image-caption subset of COYO-700M. We will validate it in the future.
37
+
38
+ Furthermore, this dataset contains text-span-bounding-box pairs. So it can be employed in many location-aware mono/multimodal tasks, such as phrase grounding, referring expression comprehension, referring expression generation and open-world object detection.
39
+
40
+ ### Data Instance
41
+ One instance is
42
+ ```python
43
+ {
44
+ 'key': '000373938',
45
+ 'clip_similarity_vitb32': 0.353271484375,
46
+ 'clip_similarity_vitl14': 0.2958984375,
47
+ 'id': 1795296605919,
48
+ 'url': "https://www.thestrapsaver.com/wp-content/uploads/customerservice-1.jpg",
49
+ 'caption': 'a wire hanger with a paper cover that reads we heart our customers',
50
+ 'width': 1024,
51
+ 'height': 693,
52
+ 'noun_chunks': [[19, 32, 0.019644069503434333, 0.31054004033406574, 0.9622142865754519, 0.9603442351023356, 0.79298526], [0, 13, 0.019422357885505368, 0.027634161214033764, 0.9593302408854166, 0.969467560450236, 0.67520964]],
53
+ 'ref_exps': [[19, 66, 0.019644069503434333, 0.31054004033406574, 0.9622142865754519, 0.9603442351023356, 0.79298526], [0, 66, 0.019422357885505368, 0.027634161214033764, 0.9593302408854166, 0.969467560450236, 0.67520964]]
54
+ }
55
+
56
+ ```
57
+ - `key`: The generated file name when using img2dataset to download COYO-700M (omit it).
58
+ - `clip_similarity_vitb32`: The cosine similarity between text and image(ViT-B/32) embeddings by [OpenAI CLIP](https://github.com/openai/CLIP), provided by COYO-700M.
59
+ - `clip_similarity_vitl14`: The cosine similarity between text and image(ViT-L/14) embeddings by [OpenAI CLIP](https://github.com/openai/CLIP), provided by COYO-700M.
60
+ - `id`: Unique 64-bit integer ID in COYO-700M.
61
+ - `url`: The image URL.
62
+ - `caption`: The corresponding caption.
63
+ - `width`: The width of the image.
64
+ - `height`: The height of the image.
65
+ - `noun_chunks`: The noun chunks (extracted by [spaCy](https://spacy.io/)) that have associated bounding boxes (predicted by [GLIP](https://github.com/microsoft/GLIP)). The items in the children list respectively represent 'Start of the noun chunk in caption', 'End of the noun chunk in caption', 'normalized x_min', 'normalized y_min', 'normalized x_max', 'normalized y_max', 'confidence score'.
66
+ - `ref_exps`: The corresponding referring expressions. If a noun chunk has no expansion, we just copy it.
67
+
68
+ ### Download image
69
+
70
  ### Citation Information
71
+ If you apply this dataset to any project and research, please cite our paper and coyo-700m:
72
  ```
73
+ @article{Kosmos2,
74
  title={Kosmos-2: Grounding Multimodal Large Language Models to the World},
75
  author={Zhiliang Peng and Wenhui Wang and Li Dong and Yaru Hao and Shaohan Huang and Shuming Ma and Furu Wei},
76
  journal={ArXiv},