Label export format

Updated 1 week ago by Alex Cota

This page describes what is included in the JSON file when you export your labels from Labelbox.

JSON export

The information included in "objects" and "classifications" will vary depending on the tool you use to create your labels.

{
"ID": "label_id",
"DataRow ID": "datarow_id",
"Labeled Data": "url to labeled image",
"Label": {
"objects": [],
"classifications": []
},

"Created By": "email of person to edit label",
"Project Name": "project name",
"Created At": "date time created",
"Updated At": "date time updated",
"Seconds to Label": time in seconds,
"External ID": datarow external id,
"Agreement": calculated consensus score,
"Benchmark Agreement": calculated consensus score against benchmark,
"Benchmark ID": benchmark label was scored against,
"Benchmark Reference ID": points to benchmark,
"Dataset Name": "datasest name",
"Reviews": [],
"View Label": "url to label view in interface"
}
Objects & Classifications

In our most recent version of the Image editor, we expanded the Label object to provide more information for each feature. The key change is that, instead of each class name appearing as a key at the root of Label, objects and classifications are now the keys at the root and information for each class is distributed between the two properties.

Objects

The objects property contains the 2D annotation data for each feature on the image. Objects may or may not have nested classifications.

Bounding box label exports, Polyline label exports, Polygon label exports, and Point label exports all contain XY coordinates in their exported files. The XY origin is located at the top left of the labeled image. We recently added an instanceURI field to vector polygons.

Follow the example in this page to authenticate and append a query parameter to the end of the URL given for the instanceURI to access.

Image segmentation label exports contain a white/clear PNG mask per instance which is presented as the instanceURI in the JSON or CSV export file. This format supports panoptic (instance) segmentation, so each object instance will be one row in your output and will contain its own white/clear mask. The XY coordinates are listed in the order they were created.

Classifications

The classifications property contains the global classification data for the image. The data type for the answer depends on how the image classification was entered: text input will be a string, radio input will be an object, and checkbox or dropdown input will be an array. The classifications property does not include object nested classifications.

Classification label exports do not include geometric information. The data type in the exported JSON file depends on whether the labeler uses radio, checklist, dropdown, or text classification to answer the question in the image editor.

Bounding box label format for legacy image editor.

{ 
"bounding_box_class_one": [
{
// instance one
geometry: [
{
x: 10, y: 10
},
{
x: 20, y: 10
},
{
x: 20, y: 20
},
{
x:10, y: 20
}
]
}
],
"text_question": "the text questions answer"
}

Bounding box label format in new image editor.

{ 
"objects": [
{
"featureId": "",
"schemaId": "[Unique ID of the bbox_class_one class]",
"title": "Bounding box class one",
"value": "bounding_box_class_one",
"color": "#FF0000",
"bbox": {
"top": 10,
"left": 10,
"height": 10,
"width": 10
}
}
],
"classifications": [
{
"featureId": "[Unique ID of answer instance to question]",
"schemaId": "[Unique ID of the bbox_class_one class]",
"title": "Text Question",
"value": "text_question",
"answer": "the text questions answer"
}
]
}

COCO or PASCAL VOC export

Labelbox does not offer any direct export options for COCO or PASCAL VOC. Users still using the legacy image editor can use the conversion scripts in this open source package to convert a JSON format to COCO or VOC format.


Was this page helpful?