Choosing the right images for training, validating, and testing computer vision algorithms will significantly affect your AI project’s success. To train an AI model for object detection, segmentation, and classification with human-like performance, each image in the dataset must be labeled thoughtfully and accurately. Annotation quality makes a significant difference in the performance of machine-learning models, so the better the annotations, the better the results. Image annotation is the process of labeling images in a dataset in order to train a machine learning algorithm. Photo and video annotations are manual, time-consuming, and very important. There are several tools available on the AIEX platform that make this process easier and faster. AIEX supports three main CV tasks: object detection, segmentation, and classification, and provides three different types of annotations. Let’s review the supported annotation types.
Annotating images with bounding boxes is quick and easy. A bounding box (bbox) annotation requires specific objects in an image to be enclosed within a bounding box. Object detection algorithms require these annotations to denote the boundaries of objects. However, they do not provide as much accuracy as segmentation or polygonal annotations, but they are sufficient for use in detectors. A bounding box is shown in the figure below, its coordinates are stored as [Xmin, Ymin, W, H].
Polygon annotations are made by connecting points as vertices around the object. They capture lines and angles more accurately than bounding boxes. Polygon annotation gives annotators the flexibility to change directions to accurately represent an object’s shape. In general, polygon masks are more precise than bounding boxes. As shown in the figure below, we specify the desired object area using a set of points, which are stored as the location of the object.
Segmentation involves categorizing regions by their class or label in an Image. In segmentation, we assign regions selected by polygons at the pixel level to categories. The annotation of a sample segmentation project on the AIEX platform is shown below.
The object-detection method places a bbox around each object. Objects inside the bboxes are categorized based on their names and placed in corresponding bboxes. We need a bbox around each object in a photo specified by the coordinates of its 4 points. The regression algorithm determines these 4 points, and the classification algorithm determines what is inside the bbox. Object-Detection annotation on the AIEX platform is shown below:
Classification models aim to find out what the image represents as a whole. Classification is concerned with identifying and categorizing a class of images rather than a specific object in an image. Classification is generally applied to images that contain one object. The process of annotation for images in a classification project is straightforward, we just need to assign a label to the entire image. Below you can see how image classification annotation is done on the AIEX platform:
You can enter your email address and subscribe to our newsletter and get the latest practical content. You can enter your email address and subscribe to our newsletter.