2D Bounding Box Annotation
A perfectly bounding 2D box is used to label the objects in a given 2D image. Mostly used in creating datasets for autonomous vehicle training.
Polygon/Contour Annotation
Objects in an image are labeled by drawing an accurate contour around it. Used in creating datasets for training precise application models.
Semantic Segmentation
The image is segmented semantically at its pixel level. Based on the pixel labeling, semantic segmentation is of 2 types:
- Full pixel/Standard segmentation.
- Instance segmentation.
Cuboidal Annotation
Labeling of objects in an image using cuboids for processing 3D training data. This technique can be used in both 2D images and 3D point cloud data.
Key Point Annotation
Objects in an image are labeled using points to determine the shape of it. Key point annotation are used to label facial/skeletal features, automotive parts etc.
Polyline Annotation
It is the labeling of lanes using splines. Exclusively used for validating ADAS systems in autonomous vehicle training.
Frequently Asked Questions
Image annotation services involve labeling objects, regions, boundaries, or keypoints in images to create structured datasets for training and evaluating AI models. Common applications include object detection, image segmentation, and visual recognition.
Common image annotation techniques include 2D bounding box annotation, polygon annotation, semantic segmentation, instance segmentation, cuboid annotation, keypoint annotation, and polyline annotation. The appropriate technique depends on the target objects and the intended AI application.
2D bounding box annotation uses rectangular boxes to mark object locations in images. It is commonly used to prepare datasets for object detection models in applications such as autonomous driving, retail analytics, and manufacturing inspection.
Polygon annotation traces object boundaries using connected points to create an outline. Semantic segmentation assigns class labels at the pixel level. Both can represent object shapes, but they use different annotation approaches and may suit different project requirements.
Semantic segmentation assigns category labels to pixels, while instance segmentation also distinguishes individual objects within the same category. For example, semantic segmentation labels pixels as cars, whereas instance segmentation separates each individual car.
Cuboid annotation uses three-dimensional bounding boxes to represent an object's position, dimensions, and orientation. It can support 3D object detection and spatial perception datasets. The workflow differs depending on whether the source data consists of 2D images or 3D point clouds.
