GUIDE NLP Autolabeling with Quality Assurance 🤖

Magicwand

The Magicwand tag makes it possible to click in a region of an image a user is doing segmentation labeling on, drag the mouse to dynamically change flood filling tolerance, then release the mouse button to get a new labeled area. It is particularly effective at segmentation labeling broad, diffuse, complex edged objects, such as clouds, cloud shadows, snow, etc. in earth observation applications or organic shapes in biomedical applications.

Use with the following data types: image.

Zooming is supported for the Magic Wand, but it will not work on rotated images.

Example of the Magic Wand in use:

Animated GIF showing Magic Wand clicking on cloud and dragging, automatically segmenting and selecting
pixels to create a mask

CORS Configuration

The Magic Wand requires pixel-level access to images that are being labelled in order to do its thresholding and flood filling. If you are hosting your images to label on a third-party domain, you will need to enable CORS headers for the Magic Wand to work with cross domain HTTP GET requests in order for the Magic Wand to be able to threshold the actual image pixel data. See the Label Studio storage guide for more details on configuring CORS.

Image Tag Configuration

The Magicwand tag is configured to work with an Image tag that it will operate on for labeling. If you are storing an image cross-domain that the Image tag will reference, you will have to correctly setup the crossOrigin on the Image attribute. This attribute mimics the same crossOrigin attribute that a normal DOM img tag would have ([reference])(https://developer.mozilla.org/en-US/docs/Web/API/HTMLImageElement/crossOrigin).

If the image is on a public server or Google/AWS/Azure bucket that is publicly readable without any authentication, you should set crossOrigin to anonymous.

If the image is on a server or a private cloud bucket that requires authentication of any kind (i.e. the request must have HTTP headers that prove authentication set along with the third party request), then you should set crossOrigin to use-credentials. Note that Google’s cloud buckets do not support authenticated requests for CORS requests, which means you either need to make that Google bucket world readable to work with the Magic Wand, or use Label Studio’s signed URL support (AWS, GCP, and Azure).

If the image is on the same host as your Label Studio instance, you can simply leave off the crossOrigin attribute or set it to none.

Parameters

Param Type Default Description
name string Name of the element
toName string Name of the image to label
[opacity] float 0.6 Opacity of the Magic Wand region during use
[blurradius] number 5 The edges of a Magic Wand region are blurred and simplified, this is the radius of the blur kernel
[defaultthreshold] number 15 When the user initially clicks without dragging, how far a color has to be from the initial selected pixel to also be selected

Sample Results JSON

Name Type Description
original_width number Width of the original image (px)
original_height number Height of the original image (px)
image_rotation number Rotation degree of the image (deg)
value Object
value.format "rle" Format of the masks, only RLE is supported for now
value.rle Array.<number> RLE-encoded image

Example JSON

{
  "original_width": 1920,
  "original_height": 1280,
  "image_rotation": 0,
  "value": {
    "format": "rle",
    "rle": [0, 1, 1, 2, 3],
    "brushlabels": ["Car"]
  }
}

Example

Basic image segmentation labeling configuration, with images stored on a third-party public cloud bucket:

<View>
  <Labels name="labels" toName="image">
    <Label value="Person" />
    <Label value="Animal" />
  </Labels>
  <MagicWand name="magicwand" toName="image" />
  <Image name="image" value="$image" crossOrigin="anonymous" />
</View>

Example

Magic Wand example with zoom controls and the brush turned on:

<View>
  <Labels name="labels" toName="image">
    <Label value="Person" />
    <Label value="Animal" />
  </Labels>
  <MagicWand name="magicwand" toName="image" />
  <Brush name="brush" toName="image" />
  <Image name="image" value="$image" zoomControl="true" zoom="true" crossOrigin="anonymous" />
</View>