With properties, you can add complex data structures without the need to create additional classes. This adds greater granularity to the visual complexity which you can label in your training data, as well as making it easier to keep your class taxonomy organized.
After adding a label, you can now click on it to select from customizable and predefined properties.
Here's an example of using the properties feature for classifying different types of apple diseases.
While all polygon segmentation masks above are instances of the Apple class, we can also create our own Properties, such as optional disease types like Apple Scab or Bitter Rot. You can utilize single or multiple-choice properties that can be easily assigned to your annotations with just a few clicks, which are structured as questions for your annotators to answer.
Benefits of Using Properties:
Once you create a new class in your dataset’s class management tab it is possible to add multiple properties to it. After the update, you can access six different property types.
With these properties, AI training becomes more nuanced and adaptive. For example, in autonomous vehicle development, combining single select properties like ‘type: pedestrian', ‘state: 'moving', and directional vectors can provide richer training data for obstacle detection algorithms. Similarly, in healthcare, using text and single select properties can greatly enhance the quality of DICOM annotations by providing detailed context and specific classification.
In V7 Darwin, integrating properties into your annotation process is a straightforward and intuitive experience.
Start by navigating to a specific dataset and opening its class management tab (Classes). This is your control center for creating and managing classes and their properties. Whether you're creating a new class or editing an existing one, you'll find an option to add properties.
Note that our new property types, Multi Select and Single Select let you add a specific question or annotation guidelines. For example, in the screenshot above you can see a Single Select property called ‘Toxicity’ with a short instruction: ‘Determine, on a scale of 1 to 5, how toxic it [a mushroom] is.
Customize each property according to your project's specific needs. The interface allows you to define options for selection-based properties.
Once you've set up your properties, save your changes. These properties will then be available to annotators, who can apply them during the data labeling process.
Properties in V7 Darwin can be accessed in two ways: by clicking on the annotation within the image or frame, or through the annotation list located on the right side of the annotator’s panel.
For additional information you can visit our documentation on properties.
V7 properties serve as a versatile tool for data organization and contextualization, enhancing the depth of datasets. They enable users to categorize, filter, and manage annotations efficiently, linking data to external AI models for automated insights. Moreover, properties enrich data for multimodal AI, offering additional layers of context for more accurate and dynamic machine learning applications.
Use properties to sort and find specific dataset items. For instance, in automotive manufacturing, you can filter out images with surface irregularities or paint defects using custom properties. This can help you pick the right training data for models focusing on detecting specific types of defects, while still using a relatively simple class structure. You can find out more about advanced dataset filtering here.
In V7, regular Tags are image-level annotations. However, if you need to add tag-like labels to specific parts of an image, you can use multi/single-select properties or attributes. This process can also be automated with V7 workflows. For example, you can add custom attributes to different segmentation masks with foundation models.
In the picture above, you can see attributes generated automatically by ChatGPT Vision. The request prompt was to generate code conforming to Darwin JSON for each piece of clothing.
In medical diagnostics, use text properties to annotate CT scans, MRIs, or X-rays with clinical notes. This pairing can train AI models to correlate visual data with medical observations, enhancing diagnostic accuracy.
With the introduction of customizable properties in V7 Darwin, the data annotation process becomes not only more nuanced but also more adaptable to textual information. This enhancement paves the way for creating richer, more informative datasets, ultimately leading to the development of more accurate and sophisticated AI models.