Image Annotation

Unlike humans, a computer only takes pixels of an image as input and tries to “guess” what an object is from a machine learning model. In supervised machine training, training data, which is also called “ground truth”, is the key to training your machine learning model to allow it to make the best possible prediction.  

Bounding Box

A bounding box is a rectangular box that is closely fitted around the target object. It is an intuitive, simple, and low-cost labeling technique that is widely used in the annotation field. It involves detecting the locations and attributes of the target persons or objects (pedestrians, vehicles, traffic signs, buildings, and other moving or static objects) from the input images.

Polygon Annotation

Polygon annotation is a more precise annotation technique than the bounding box. It is designed to output more accurate object position information for machine learning and involves a cost slightly higher than that of drawing bounding boxes.

Semantic Segmentation

Image semantic segmentation is an important part of image understanding in machine vision technology. It allows computers to classify images at the pixel level based on their understanding of image content, so that different types of objects in the image can be segmented from the perspective of pixels. Semantic segmentation is often used to segment people and objects. It requires extremely high labeling accuracy, so the cost is relatively high.

Human Skeleton Points

The joint points of human skeletons are very important in describing the postures of humans and predicting their behaviors. Therefore, the detection of key points of human skeletons is a foundation for some machine vision tasks such as motion classification, unusual behavior detection, and autonomous driving.

Point-Cloud Annotation

This annotation technique is specifically used in the automotive industry to test how well multi-LIDAR could detect objects such as other cars or pedestrians. Annotators will work frame by frame on continuous video streams collected by a LIDAR or a camera and use 3D boxing or semantic segmentation to label objects such as cars or people. When an annotator finishes one frame and moves on to the next one, the tools for semi-automatic tracking are used to ensure the re-identification of objects and thus to achieve higher working efficiency