Export Format Detail


JSON and CSV contain the same data in different formats. the JSON export is a list of JSON objects where each object is a label, and the CSV export has each row as a label. Below is an image of one row from the CSV and one JSON from the list:

ID: Unique internal ID for the label

DataRow ID: Unique internal ID for the row in the dataset, shared across projects **Labeled Data: ** URL linking to the image that was labeled

Label: JSON label object, as described in the label page of this documentation

Created By: Email of the last person to edit the label

Created At: Date/Time the label was first created

Updated At: Date/Time the label was most recently updated

Seconds to Label: Time it took to create the label in seconds

External ID: External Id associated with the Datarow

Agreement: Calculated consensus score for the image, if enabled

Benchmark Agreement: Calculated consensus score against a benchmark, if enabled

Benchmark ID: The benchmark that this label was scored against

Benchmark Reference ID: If this label was starred as a benchmark it is used to score other labels on the DataRow. The ID in this field points to the Benchmark.

Dataset Name: Name of the uploaded Dataset

Reviews: List of JSON reviews: [‘id’, ‘score’, ‘createdAt’, ‘createdBy’]

View Label: Direct link to this label view in the labelbox interface


Our coco export fits the described COCO image standards: http://cocodataset.org/#format-data. However, since labelbox IDs are strings, rather than integers, we populate the ID fields in our COCO export with String IDs, in order to ensure internal consistency. If you’re having trouble with this, it’s a fairly straightforward python script to migrate the IDs to a new field.

Due to the tight limitations on the COCO format, much of the data from the JSON and CSV exports isn’t available: the ExternalID, Labeler and Seconds to Label, and the reviews will all be unavailable.


VOC Exports are a compressed .tar file (.tar.gz) containing image - xml pairs. VOC exports are the only export type that downloads the images directly into the export file, with all the others only giving reference URLs. The image and .xml files are both named using the Label ID from your project, with the XML data giving the annotation to go with the image.

folder and path: these can essentially be ignored, as they’re internal references, rather than useful for reading your images. filename: The name of the corresponding image for this label database: Always reads “unknown” size: dimensions of the image data, Depth indicates the number of channels in the image Object: each polygon, box, line, or point in an image is in it’s own object box Name: The name of the object class Pose, Truncated, difficult: always Unspecified, 0 and 0 polygon or bndbox: The annotation shape polygon: set of x,y pairs going counterclockwise, with (0,0) being the top left pixel bndbox: 4 numbers defining xmin, xmax, ymin, and ymax

Polygons and Line segments are both defined by the polygon tag, and points are not exported at all in VOC.

As with the COCO format, VOC is a strictly-defined export format, and therefore loses a lot of information compared to our native JSON and CSV exports.

How did we do?