Label formats

When you label a sample and press the save button, you've created a label for that sample. Labels come in different types, with the available options determined by the type of the corresponding sample.

When downloading or uploading labels using the Python SDK, the format of the attributes field depends on the type of label. The different formats are described here.

A label can additionally also contain Object attributes and Image attributes.

Image

Segmentation labels

Format of the attributes field in client.get_label():

{
  "format_version": "0.1",
  "annotations": [
    {
      "id": 1, // this is an object id. Should be > 0.
      "category_id": 1, // this is a category id
      "track_id": 1 // only required if part of a sequence sample
    },
    {
      "id": 2, 
      "category_id": 1,
      "track_id": 2,
    },
    {
      "id": 3, 
      "category_id": 4,
      "track_id": 3,
    }
  ],
  "segmentation_bitmap": {
    "url": "https://segmentsai-staging.s3.eu-west-2.amazonaws.com/assets/davy/ddf55e99-1a6f-42d2-83e9-8657de3259a1.png"
  }
}

Thesegmentation_bitmap_urlrefers to a 32-bit RGBA png image which contains the segmentation masks. The alpha channel is set to 255, and the remaining 24-bit values in the RGB channels correspond to the object ids in the annotations list. Unlabeled regions should have a value of 0. Because of the large dynamic range, these png images may appear black in an image viewer.

When downloading a label, you can use the utility function utils.load_label_bitmap_from_url(url) in the Python SDK to load the label bitmap as a numpy array containing object ids.

When uploading a label, the easiest way to transform a segmentation bitmap into this format and upload it is by using the util functionbitmap2file:

from segments.utils import bitmap2file

# segmentation_bitmap is a numpy array of type np.uint32, with values corresponding to instance_ids
file = bitmap2file(segmentation_bitmap)
asset = client.upload_asset(file, "label.png")
segmentation_bitmap_url = asset.url

For a full example of uploading model-generated labels to Segments.ai, please refer to this blogpost.

Vector labels (bounding box, polygon, polyline, keypoint)

Format of the attributes field in client.get_label():

{
  "format_version": "0.1",
  "annotations": [
    {
      "id": 1, // the object id
      "category_id": 1, // the category id
      "type": "bbox", // refers to the annotation type (bounding box)
      "points": [
        [12.34, 56.78], // x0, y0 (upper left corner of bbox)
        [90.12, 34.56]  // x1, y1 (lower right corner of bbox)
      ]
    },
    {
      "id": 2,
      "category_id": 2,
      "type": "polygon", // refers to the annotation type (polygon)
      "points": [
        [12.34, 56.78], // x0, y0 (starting point of the polygon)
        [90.12, 34.56], // x1, y1
        [78.91, 23.45], // x2, y2
        [67.89, 98.76], // x3, y3
        [54.32, 10.01]  // x4, y4
      ]
    },
    {
      "id": 3, 
      "category_id": 3,
      "type": "polyline", // refers to the annotation type (polyline)
      "points": [
        [12.34, 56.78], // x0, y0 (starting point of the polyline)
        [90.12, 34.56], // x1, y1
        [78.91, 23.45], // x2, y2
        [67.89, 98.76], // x3, y3
        [54.32, 10.01]  // x4, y4
      ]
    },
    {
      "id": 4,
      "category_id": 4,
      "type": "point", // refers to the annotation type (keypoint)
      "points": [
        [12.34, 56.78] // x, y (coordinates of keypoint)
      ]
    },
  ],
}

Image sequence

Segmentation labels

Format of the attributes field in client.get_label():

{
  "format_version": "0.1",
  "frames": [
    { ... },
    { ... },
    { ... }
  ]
}

Name	Type
format_version	string	Format version.
frames	array of segmentation labels	List of segmentation labels (one per frame in the sequence).

Vector labels (bounding box, polygon, polyline, keypoint)

Format of the attributes field in client.get_label():

{
  "format_version": "0.2",
  "frames": [
    { ... },
    { ... },
    { ... }
  ]
}

Where each frames object has the following format:

{
  "format_version": "0.1",
  "timestamp": "00001", // this field is only included if the sample has a timestamp
  "annotations": [
    {
      "id": 1, // the object id
      "category_id": 1, // the category id
      "track_id": 6, // this id is used to links objects across frame
      "is_keyframe": true, // whether this frame is a keyframe
      "type": "bbox", // refers to the annotation type (bounding box)
      "points": [
        [12.34, 56.78], // x0, y0 (upper left corner of bbox)
        [90.12, 34.56]  // x1, y1 (lower right corner of bbox)
      ]
    },
    {
      "id": 2,
      "category_id": 2,
      "track_id": 5, // this id is used to links objects across frame
      "is_keyframe": true, // whether this frame is a keyframe
      "type": "polygon", // refers to the annotation type (polygon)
      "points": [
        [12.34, 56.78], // x0, y0 (starting point of the polygon)
        [90.12, 34.56], // x1, y1
        [78.91, 23.45], // x2, y2
        [67.89, 98.76], // x3, y3
        [54.32, 10.01]  // x4, y4
      ]
    },
    {
      "id": 3, 
      "category_id": 3,
      "track_id": 4, // this id is used to links objects across frame
      "is_keyframe": true, // whether this frame is a keyframe
      "type": "polyline", // refers to the annotation type (polyline)
      "points": [
        [12.34, 56.78], // x0, y0 (starting point of the polyline)
        [90.12, 34.56], // x1, y1
        [78.91, 23.45], // x2, y2
        [67.89, 98.76], // x3, y3
        [54.32, 10.01]  // x4, y4
      ]
    },
    {
      "id": 4,
      "category_id": 4,
      "track_id": 3, // this id is used to links objects across frame
      "is_keyframe": true, // whether this frame is a keyframe
      "type": "point", // refers to the annotation type (keypoint)
      "points": [
        [12.34, 56.78] // x, y (coordinates of keypoint)
      ]
    },
  ],
}

3D point cloud

Segmentation label

The annotations array contains the different objects ("annotations") in the label with their category (the category_id should correspond to an id defined in the categories).

The point_annotations array contains the object/annotation id for each point in the point cloud. The order of the ids in this array is the same as the order of the points in the point cloud.

{
  "format_version": "0.1",
  "annotations": [
    {
      "id": 1, // the object id
      "category_id": 1 // the category id
    },
    {
      "id": 2, 
      "category_id": 1
    },
    {
      "id": 3, 
      "category_id": 4
    }
  ],
  "point_annotations": [0, 0, 0, 3, 2, 2, 2, 1, 3...], // refers to object ids
}

Cuboid label

{
  "format_version": "0.2",
  "annotations": [ // list of cuboid annotations, see below
    {
      "id": 1,
      "category_id": 1,
      "type": "cuboid",
      ...
    },
    { 
      ... 
    }
  ]
}

Name	Type	Description
format_version	string	Format version.
annotations	array of cuboid annotations	List of the cuboid annotations.

Cuboid annotation

A cuboid annotation represents a single cuboid in a point cloud (frame).

{
  "id": 1,
  "category_id": 1,
  "type": "cuboid",
  "position": {
    "x": 0.0,
    "y": 0.2,
    "z": 0.5
  },
  "dimensions": {
    "x": 1.2,
    "y": 1,
    "z": 1
  },
  "yaw": 0.63,
  "rotation": {
    "qx": 0,
    "qy": 0.0491566,
    "qz": 0.3096865,
    "qw": 0.9495672
  },  // only when 3D cuboid rotation is enabled in dataset settings
  "track_id": 1,  // only in sequences
  "is_keyframe": true,  // only in sequences
  "index": 0,  // only in sequences 
}

Name	Type	Description
id	integer	Object id.
category_id	integer	Category id.
type	string	Object type, which is always "cuboid" for cuboid annotations.
position	object: { "x": float, "y": float, "z": float }	XYZ position of the center of the cuboid in world coordinates.
dimensions	object: { "x": float, "y": float, "z": float }	Dimensions of the cuboid. "x" corresponds to the length, "y" to the width, and "z" to the height. See diagram 1.
yaw	float	Cuboid rotation along the z-axis in radians between [-π, π]. 0 yaw corresponds to a cuboid aligned with the x-axis pointing to increasing x-values. The yaw value increases with a counter-clockwise rotation up to π, and decreases with a clockwise rotation up to -π. See diagram 2.
rotation	object: { "qx": float, "qy": float, "qz": float, "qw": float }	Cuboid rotation defined as a rotation quaternion. Only available when 3D rotation is enabled in the dataset settings (under "Labeling").
track_id	integer	Track ID of the object. This ID is used to track an object over multiple frames. Only relevant for sequences.
is_keyframe	boolean	Whether this cuboid annotation is a keyframe or an interpolated frame. Only relevant for sequences.
index	integer	The frame index. Only relevant for sequences.

Diagram 1: x and y attributes of the cuboid dimensions. The red arrow shows the cuboid heading.

Diagram 2: yaw rotation of a cuboid. The red arrow shows the cuboid heading. yaw = π/2 corresponds to a heading in the direction of increasing y values, while yaw = -π/2 corresponds to a heading in the direction of decreasing y values.

Vector label (polygon, polyline, keypoint)

  "format_version": "0.1",
  "annotations": [
    {
      "id": 2,
      "category_id": 2,
      "type": "polygon", // refers to the annotation type (polygon)
      "points": [
        [12.34, 56.78, 0], // x0, y0, z0 (starting point of the polygon)
        [90.12, 34.56, 0], // x1, y1, z1
        [78.91, 23.45, 0], // x2, y2, z2
        [67.89, 98.76, 0], // x3, y3, z3
        [54.32, 10.01, 0]  // x4, y4, z4
      ],
      "track_id": 2, // only in sequences
      "is_keyframe": true, // only in sequences
      "index": 0 // only in sequences 
    },
    {
      "id": 3, 
      "category_id": 3,
      "type": "polyline", // refers to the annotation type (polyline)
      "points": [
        [12.34, 56.78, 0], // x0, y0, z0 (starting point of the polyline)
        [90.12, 34.56, 0], // x1, y1, z1
        [78.91, 23.45, 0], // x2, y2, z2
        [67.89, 98.76, 0], // x3, y3, z3
        [54.32, 10.01, 0]  // x4, y4, z4
      ],
      "track_id": 1, // only in sequences
      "is_keyframe": false, // only in sequences
      "index": 1 // only in sequences 
    },
    {
      "id": 4,
      "category_id": 4,
      "type": "point", // refers to the annotation type (keypoint)
      "points": [
        [12.34, 56.78, 0] // x, y, z (coordinates of keypoint)
      ],
      "track_id": 3, // only in sequences
      "is_keyframe": false, // only in sequences
      "index": 2 // only in sequences 
    }
  ]
}

3D point cloud sequence

Segmentation label

{
  "format_version": "0.2",
  "frames": [
    { ... },
    { ... },
    { ... }
  ]
}

Where each frames object has the following format:

{
  "format_version": "0.2",
  "annotations": [
    {
      "id": 1, // the object id
      "category_id": 1, // the category id
      "track_id": 3 // this id is used to link objects across frames
    },
    {
      "id": 2, 
      "category_id": 1,
      "track_id": 4
    },
    {
      "id": 3, 
      "category_id": 4,
      "track_id": 5
    },
  ],
  "point_annotations": [0, 0, 0, 3, 2, 2, 2, 1, 3...], // refers to object ids
}

Cuboid label

{
  "format_version": "0.2",
  "frames": [
    { ... },
    { ... },
    { ... }
  ]
}

Name	Type
format_version	string	Format version.
frames	array of cuboid labels	List of cuboid labels (one per frame in the sequence).

Vector label (polygon, polyline, keypoint)

Format of the attributes field in client.get_label():

{
  "format_version": "0.2",
  "frames": [
    { ... },
    { ... },
    { ... }
  ]
}

Name	Type
format_version	string	Format version.
frames	array of vector labels	List of vector labels (one per frame in the sequence).

Multi-sensor sequence

{
  "sensors": [
    {
      "name": "Lidar", 
      "task_type": "pointcloud-cuboid-sequence",
      "attributes": { ... }
    },
    {
      "name": "Camera 1", 
      "task_type": "image-vector-sequence",
      "attributes": { ... } 
    },
    ...
  ]
}

Name	Type	Description
sensors	array of sensors	List of the sensors with label attributes.

Sensor

Name	Type	Description
name	string	Name of the sensor.
task_type	string	The task type of the sensor. Currently, pointcloud-cuboid-sequence and image-vector-sequence are supported.
attributes	object	The label attributes for the sensor. Currently, 3D point cloud sequence and image sequence are supported.

Object attributes

Objects in the annotations list can optionally also contain an attributes field to store object-level attributes. Make sure to properly configure the label editor if you're using object-level attributes.

{
  "format_version": "0.1",
  "annotations": [
    {
      "id": 1, 
      "category_id": 1,
      "attributes": { // object-level attributes
        "is_crowd": "1",
        "color": "red"
      }
    },
    {
      "id": 2, 
      "category_id": 1,
      "attributes": {
        "is_crowd": "0",
        "color": "blue"
      }
    },
    {
      "id": 3, 
      "category_id": 4,
      "attributes": {
        "is_crowd": "1",
        "color": "yellow"
      }
    }
  ],
  ...
}

Image attributes

You can also define image-level attributes. These can be useful in image classification tasks. Make sure to properly configure the label editor if you're using image-level attributes.

{
  "format_version": "0.1",
  "annotations": [...],
  "image_attributes": { // sample-level attributes
    "scene_type": "crossroads",
    "weather": "sunny"
  }
}