2D and 3D Bounding Boxes

Annotators draw 2D or 3D boxes around objects of interest and label them, for example by type of object or attributes. For 3D boxes, also known as cuboids, annotators need to mark anchor points at the edges of objects to delineate length, width and depth.

Polygons

To precisely identify irregularly-shaped objects of interest, annotators can use different types of polygons, which they then label by type and attribute. Machine learning models can assist this process by recognizing the object and recommending corner points for the polygon.

Lines and Splines

Annotators draw straight or curved lines that mark boundaries of an object, for example lanes on a road. This type of annotation helps train AI models for autonomous vehicles to detect lanes while driving.

Landmark Annotation

For facial traits, gestures and postures, Annotators draw points on specific features or “landmarks” on the face or body. Computer vision algorithms can use this data to interpret emotions and gestures, recognize people through facial recognition or assess posture changes in sports.

Optical Character Recognition

In order to process text from still images or video, we extract text from images or video frames and convert them into machine-encoded text. These data sets can provide the basis for training systems that interpret traffic signs, license plates, serial numbers, documents and IDs.

Image Classification

For image classification, annotators assign an entire image with one label to categorize it by type. A machine learning system can use this data, for example, to learn to differentiate skin disorders such as melanoma, melanocytic nervus, actinic keratosis, benign keratosis, etc.

Semantic Segmentation

While bounding boxes and polygons annotate certain objects of interest only, semantic segmentation annotates every pixel in the image. This segments the image into categories that are identified by a color code. For example, vehicles are colored blue or pedestrians are colored red.

Video Tracking

Video tracking requires frame-by-frame annotation to help computer vision systems detect, identify and track objects and living beings accurately. After identifying objects of interest with 2D or 3D bounding boxes or semantic segmentation, annotators can label them by type or attribute.