metadata

dataset_info:
  features:
    - name: image
      dtype: image
  splits:
    - name: train
      num_bytes: 1052463508.623
      num_examples: 103273
    - name: validation
      num_bytes: 123787922.896
      num_examples: 4016
  download_size: 1170594310
  dataset_size: 1176251431.519
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: validation
        path: data/validation-*
task_categories:
  - image-to-image
size_categories:
  - 100K<n<1M
license: cc
tags:
  - coco
  - colorization
  - image colorization
pretty_name: COCO 2017 Image Colorization 224x224

COCO 2017 for Image Colorizaion

Dataset Summary

COCO (Common Objects in Context) is a large-scale object detection, segmentation, and captioning dataset. It includes complex, everyday scenes with common objects in their natural context. This dataset is a version of the COCO 2017 dataset, specifically processed and cleaned for image colorization tasks. It contains the images from the original dataset but they have been resized and center cropped to 224x224. It is further filtered to remove grayscale images, heavily filtered images, and other artifacts not suitable for training a natural colorization model.

Dataset structure

The dataset contains the following field image, containing a color image. There are two splits:

Train split
Validation split

Note: All original annotations (including bounding boxes, segmentations, and captions) have been removed from this version of the dataset, as they are not required for image colorization tasks. During training, the grayscale version of the image can be generated on-the-fly from the color image.

The final number of images in the dataset after filtering are 103,273 in the train split, compared to the original 118,287 images and 4016 in the validation split, compared to the original 5000 images.

Curation Rationale

This dataset was curated with the goal of training deep learning models to convert grayscale images into realistic color images, and advancing the state of the art in image colorization tasks. Since there are no dedicated datasets for image colorization, this dataset can be very helpful in such task. Existing datasets like nickpai/coco2017-colorization have many black & white, heavily filtered, and other images not suitable for image colorization tasks.

License

The images in this dataset are sourced from the COCO 2017 dataset, which were originally published on Flickr. The creators of COCO do not own the copyright for the images. Use of the images must abide by the Flickr Terms of Use. For more details, please refer to the original COCO dataset page.

Citation Information

@inproceedings{lin2014microsoft,
title={Microsoft coco: Common objects in context},
author={Lin, Tsung-Yi and Maire, Michael and Belongie, Serge and Hays, James and Perona, Pietro and Ramanan, Deva and Doll{'a}r, Piotr and Zitnick, C Lawrence},
booktitle={Computer Vision--ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part V 13},
pages={740--755},
year={2014},
organization={Springer}
doi={10.1007/978-3-319-10602-1_48},
url={https://arxiv.org/abs/1405.0312}
}