COCO 2017 is distributed by the COCO Consortium. The annotations are licensed under CC BY 4.0. The COCO Consortium does not own the copyright of the images: the images come from Flickr and remain under their respective Creative Commons licenses, and about 70% of them are non-commercial. Check the license of each image (license/license_url) before using it, and cite the original work.

COCO 2017

COCO 2017 instance detection, packed once with the official annotations kept whole.

What is here

train: 118,287 images (train2017)
validation: 5,000 images (val2017)
80 categories with COCO's official ids, 1 to 90 (12, 26, 29, ... are unused)
Every official annotation column: annotation id, category id, bbox, area, iscrowd, segmentation
Every image's official license and source URLs, because the images come from Flickr and are not all under the same license
All 5,000 validation images are kept, including the 48 that have no annotations; their objects lists are empty

test2017 is not included: the official test annotations are not public (the test server only), so there is nothing to pack.

Schema

column	type	meaning
`id`	string	image id, the zero-padded file name stem (e.g. `000000000139`)
`image`	Image	the original JPEG bytes
`width`, `height`	int32	image size
`license`, `license_url`	string	the image's official Flickr license name and URL
`flickr_url`, `coco_url`	string	the image's original source URLs
`objects.annotation_id`	int64	official annotation id
`objects.category`	ClassLabel	the 80 names in official id order; the index is 0 to 79
`objects.category_id`	int32	COCO's official category id, 1 to 90 with gaps
`objects.bbox`	list of 4 float64	official `[x, y, w, h]`, absolute pixels, top-left origin
`objects.area`	float64	official area in pixels
`objects.iscrowd`	int8	1 for a crowd region the official evaluator ignores
`objects.segmentation`	string	official segmentation as JSON: polygon lists, or the RLE dict for crowd boxes

category is the readable class name; category_id keeps the official number so a consumer can map straight back to the COCO category ids. No annotation is filtered, merged or renumbered.

Usage

from datasets import load_dataset

ds = load_dataset("<user>/coco2017", "detection")

Replace <user> with the repo owner.

Source and license

Official COCO 2017: https://cocodataset.org/.

The annotations in this dataset are licensed under a Creative Commons Attribution 4.0 License. The COCO Consortium does not own the copyright of the images. Use of the images must abide by the Flickr Terms of Use. The users of the images accept full responsibility for the use of the dataset, including but not limited to the use of any copies of copyrighted images that they may create from the dataset.

Each image's license is in license/license_url. The images come from Flickr under different Creative Commons licenses, and about 70% of train and validation images are non-commercial (BY-NC, BY-NC-SA, BY-NC-ND): check the per-image license before using an image, especially for commercial work. This packaging adds no rights; use it under the COCO terms of use and cite the original work.

@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={European Conference on Computer Vision (ECCV)},
  pages={740--755},
  year={2014},
  organization={Springer}
}

Reproducing this dataset

create_dataset.py in this repo builds the parquet from an unpacked official COCO 2017 folder:

python create_dataset.py --dry-run --coco-root /path/to/coco --limit 20
python create_dataset.py --repo <user>/coco2017 --coco-root /path/to/coco

It checks the packed box counts, a random sample of records and the image bytes against the official instances_*.json files before uploading.

Downloads last month: 4,958

Total file size:

20.5 GB