fiftyone.zoo.datasets.base.COCO2017Dataset

COCO is a large-scale object detection, segmentation, and captioning dataset.

This version contains images, bounding boxes, segmentations, and keypoints for the 2017 version of the dataset.

This dataset supports partial downloads:

• You can specify subsets of data to download via the label_types, classes, and max_samples parameters
• You can specify specific images to load via the image_ids parameter

See :ref:`this page <dataset-zoo-coco-2017>` for more information about partial downloads of this dataset.

Full split stats:

• Train split: 118,287 images
• Test split: 40,670 images
• Validation split: 5,000 images

Notes:

• COCO defines 91 classes but the data only uses 80 classes
• Some images from the train and validation sets don't have annotations
• The test set does not have annotations
• COCO 2014 and 2017 use the same images, but the splits are different

Example usage:

import fiftyone as fo
import fiftyone.zoo as foz

#
# Load 50 random samples from the validation split
#
# By default, only detections are loaded
#

dataset = foz.load_zoo_dataset(
    "coco-2017",
    split="validation",
    max_samples=50,
    shuffle=True,
)

session = fo.launch_app(dataset)

#
# Load segmentations for 25 samples from the validation split that
# contain cats and dogs
#
# Images that contain all `classes` will be prioritized first, followed
# by images that contain at least one of the required `classes`. If
# there are not enough images matching `classes` in the split to meet
# `max_samples`, only the available images will be loaded.
#
# Images will only be downloaded if necessary
#

dataset = foz.load_zoo_dataset(
    "coco-2017",
    split="validation",
    label_types=["segmentations"],
    classes=["cat", "dog"],
    max_samples=25,
)

session.dataset = dataset

#
# Download the entire validation split and load both detections and
# segmentations
#
# Subsequent partial loads of the validation split will never require
# downloading any images
#

dataset = foz.load_zoo_dataset(
    "coco-2017",
    split="validation",
    label_types=["detections", "segmentations"],
)

session.dataset = dataset

Dataset size

25.20 GB

Source

http://cocodataset.org/#home