Export data

How to export the labeled data.

In the projects home tab the "Export labels" button is located below the full description. A click on this button creates a new JSON file containing the labeled data. This button is only enabled if labeled data is available.

The exported JSON file starts with an overview of the label classes defined within the label configuration followed by the list of tasks.

The used example:

For this documentation the following label configuration was used.

Label classes:

The label classes may look like the following snippet:

label_classes.json

{ 
  "label_classes" : [ {
    "class_name": "bus",
    "geometry_type": "rectangle"
  }, {
    "class_name": "car",
    "geometry_type": "polygon"
  }, {
    "class_name": "car_type",
    "classification_type": "select"
  }, {
    "class_name": "weather",
    "classification_type": "radio"
  }, {
    "class_name" : "daytime",
    "classification_type" : "select"
  } ]
}

The class_names are export keys from the classifications and geometries defined in the label configuration. They are also used in the labels and global_classifications sections for each task. The geometry_type or classification_type describe the type of that classes.

Classifications of type option are exported as type radio if they contain up to three options otherwise they are exported as type select.

A single tasks:

Every exported task consists of the following properties:

task_snippted.json

{
  "internal_media_ID" : "datagym-image-id",
  "external_media_ID" : "awesome-image.jpg",
  "labeler" : "labeler-id",
  "status" : "completed",
  "segmentations": [],
  "global_classifications" : {},
  "labels" : {}
}

Property	Description
internal_media_ID	The internal id to identify the media file. This ID can be used to view or download the image.
external_media_ID	The media's file name.
labeler	To identify the assigned labeler.
status	One of the task's states. See the documentation about the task's lifecycle.
segmentations	This section is optional and appears only if at least one image segmentation is configured. This list contains the urls to export the semantic maps. See the API documentation for more details.
global_classifications	Uses the label_classes to describe the image.
labels	Uses the class names listed in the label_classes section to describe a geometry within the image.

Global classifications:

The global_classifications section uses the class_names defined in the label_classes section to describe the image. This section may look like the following:

global_classifications.json

{
  "global_classifications" : {
    "weather" : [ "sunny", {
      "daytime" : [ "day" ]
    } ]
  }
}

Nested classifications are possible.

The Labels:

The label object contains an entry for each geometry type defined in the label_classes section with the class_name as key. Every entry contains a geometry of the assigned geometry_type. Each entry in that list consists of the geometry and classifications sections. The content of the geometry depends on the geometry_type and the classifications are treated like the global_classifications. The labels section may look like the following snippet:

labels_snippet.json

{
  "labels" : {
    "bus": [],
    "car": []
  }
}

The geometries:

For every geometry_type defined in the label_classes sections, a list of objects is created. Every entry describes an area of interest in the image and is defined by its geometry and classifications.

geometry_snippet.json

{
  "car": [ {
    "geometry": [{
      "x" : 57,
      "y" : 412
    }, {
      "x" : 54,
      "y" : 463
    }, {
      "x" : 130,
      "y" : 524
    }, {
      "x" : 1148,
      "y" : 684
    } ],
    "classifications": { 
      "car_type": [ null ]
    }
  } ]
}

Image segmentation geometries:

Image segmentation geometries are basically treated the same way as polygon geometries, except for the fact that they have multiple arrays of points.

seqmentation_snipped.json

{
    "seqmentation" : [ {
      "geometry" : [ [ {
        "x" : 254,
        "y" : 173
      }, {
        "x" : 177,
        "y" : 174
      }, {
        "x" : 182,
        "y" : 77
      }, {
        "x" : 833,
        "y" : 50
      }, {
        "x" : 842,
        "y" : 160
      }, {
        "x" : 780,
        "y" : 161
      }, {
        "x" : 784,
        "y" : 322
      }, {
        "x" : 250,
        "y" : 313
      }, {
        "x" : 254,
        "y" : 173
      } ], [ {
        "x" : 579,
        "y" : 165
      }, {
        "x" : 685,
        "y" : 163
      }, {
        "x" : 688,
        "y" : 278
      }, {
        "x" : 579,
        "y" : 279
      }, {
        "x" : 579,
        "y" : 165
      } ], [ {
        "x" : 400,
        "y" : 169
      }, {
        "x" : 496,
        "y" : 167
      }, {
        "x" : 500,
        "y" : 275
      }, {
        "x" : 399,
        "y" : 272
      }, {
        "x" : 400,
        "y" : 169
      } ], [ {
        "x" : 285,
        "y" : 172
      }, {
        "x" : 332,
        "y" : 171
      }, {
        "x" : 331,
        "y" : 250
      }, {
        "x" : 286,
        "y" : 249
      }, {
        "x" : 285,
        "y" : 172
      } ] ],
      "classifications" : { }
    } ]

The first array of points are the outer points of the segmentation, while every following array is an inner segment that was cut out. The above JSON Snippet corresponds to the Segmentation in the image below.

Segmentation with three inner cutouts