GRIT: Large-Scale Training Corpus of Grounded Image-Text Pairs

Dataset Summary

We introduce GRIT, a large-scale dataset of Grounded Image-Text pairs, which is created based on image-text pairs from COYO-700M 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.

Supported Tasks

During the construction, we excluded the image-caption pairs if no bounding boxes are retained. This procedure resulted in a high-quality image-caption subset of COYO-700M, which we will validate in the future.

Furthermore, this dataset contains text-span-bounding-box pairs. Thus, it can be used in many location-aware mono/multimodal tasks, such as phrase grounding, referring expression comprehension, referring expression generation, and open-world object detection.

Data Instance

One instance is

{
  'key': '000373938', 
  'clip_similarity_vitb32': 0.353271484375, 
  'clip_similarity_vitl14': 0.2958984375, 
  'id': 1795296605919, 
  'url': "https://www.thestrapsaver.com/wp-content/uploads/customerservice-1.jpg", 
  'caption': 'a wire hanger with a paper cover that reads we heart our customers', 
  'width': 1024, 
  'height': 693, 
  '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]], 
  '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]]
}

• key: The generated file name when using img2dataset to download COYO-700M (omit it).
• clip_similarity_vitb32: The cosine similarity between text and image(ViT-B/32) embeddings by OpenAI CLIP, provided by COYO-700M.
• clip_similarity_vitl14: The cosine similarity between text and image(ViT-L/14) embeddings by OpenAI CLIP, provided by COYO-700M.
• id: Unique 64-bit integer ID in COYO-700M.
• url: The image URL.
• caption: The corresponding caption.
• width: The width of the image.
• height: The height of the image.
• noun_chunks: The noun chunks (extracted by spaCy) that have associated bounding boxes (predicted by 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'.
• ref_exps: The corresponding referring expressions. If a noun chunk has no expansion, we just copy it.

Download image

We recommend to use img2dataset tool to download the images.

1. Download the metadata. You can download it by cloning current repository:

git lfs install
git clone https://huggingface.co./datasets/zzliang/GRIT

2. Install img2dataset.

pip install img2dataset

3. Download images You need to replace /path/to/GRIT_dataset/grit-20m with the local path to this repository.

img2dataset --url_list /path/to/GRIT_dataset/grit-20m --input_format "parquet"\
    --url_col "url" --caption_col "caption" --output_format webdataset \
    --output_folder /tmp/grit --processes_count 4 --thread_count 64 --image_size 256 \
    --resize_only_if_bigger=True --resize_mode="keep_ratio" --skip_reencode=True \
    --save_additional_columns '["id","noun_chunks","ref_exps","clip_similarity_vitb32","clip_similarity_vitl14"]' \
    --enable_wandb False

You can adjust some parameters according to your actual needs (e.g., processes_count, thread_count, image_size, save_additional_columns). More img2dataset hyper-parameters can be found in here.

Citation Information

If you apply this dataset to any project and research, please cite our paper and coyo-700m:

@article{Kosmos2,
  title={Kosmos-2: Grounding Multimodal Large Language Models to the World},
  author={Zhiliang Peng and Wenhui Wang and Li Dong and Yaru Hao and Shaohan Huang and Shuming Ma and Furu Wei},
  journal={ArXiv},
  year={2023},
  volume={abs/2306.14824}
}

@misc{kakaobrain2022coyo-700m,
  title         = {COYO-700M: Image-Text Pair Dataset},
  author        = {Minwoo Byeon, Beomhee Park, Haecheon Kim, Sungjun Lee, Woonhyuk Baek, Saehoon Kim},
  year          = {2022},
  howpublished  = {\url{https://github.com/kakaobrain/coyo-dataset}},
}