COCO Format Export

If using Windows, we recommend using 7-Zip to unzip the exported labels.

All coordinates in the COCO format correspond directly to pixel values. Refer to the COCO documentation for details on the COCO format.

The export contains a list of JSON objects, with each JSON object representing a single data unit and containing detailed information for all exported labels.

The example below describes all the keys for a single exported data unit.

Key	Description	Notes
annotations	Contains all label information
area	The area contained by the label	in units of square pixels
attributes	A list of attributes belonging to a label
classifications	All radio and checklist attributes and their answers
#relation	The hash of the ontology object the label is related to	The hash is the object’s unique identifier
encord_track_uuid	A unique identifier for the object instance	The same as the objectHash in a JSON export.
manual_annotation	Indicates whether the annotation was created manually.	True is the annotation is manual, False if created using automated labeling techniques
track_id	The labels’s number in the list of labels, starting from 1	The value is the same for all annotations belonging to the same object instance
bbox	Pixel coordinates for the top-left corner of a bounding box that encloses the label. This is provided for all shape types.
category_id	A numerical ID linking this annotation to a field in the categories list containing Ontology information.	For example, the category_id value 1 in an annotation links it to the corresponding ontology shape with the ID 1 in the categories field of the export.
id	The labels’s number in the list of labels, starting from 0
image_id	The ID of the image or frame the label is on	For example, ID 0 indicates that the label is on the first image in the images list of the export
iscrowd	The iscrowd field specifies whether the annotation represents a single object (iscrowd=0, using polygons) or a collection of objects (iscrowd=1, using RLE). Single objects may require multiple polygons if occluded, while crowd annotations are used for large groups of objects (e.g., a crowd of people)
keypoints	The keypoints field is a list of coordinates and visibility flags for specific points on an object
num_keypoints	The num_keypoints field indicates the number of labeled keypoints (v > 0) for an object. Objects without labeled keypoints (e.g., small or crowd objects) have num_keypoints=0	Keypoints and their connections (skeletons) are defined in the categories section, which includes the names and connectivity of the keypoints for visualization
segmentation	Contains the (x,y) pixel coordinates for all label vertices
categories	Contains Ontology information
name	The object’s name in the Ontology
supercategory	The name of the object’s class
keypoints	keypoint for keypoint labels. Omitted for all other labels
skeleton	Defines connectivity via a list of keypoint edge pairs and is used for visualization
images	Contains file information for all modalities
coco_url	URL to the file if it is hosted on the COCO dataset’s official servers	Only meaningful when working with the official COCO tasks. In all other cases the coco_url field is empty.
file_name	The name of the file
height	The height of the image or frame	Measured in pixels
image_title	The title of the file
width	The width of the image or frame	Measured in pixels
info	Contains additional information