Explore popular image annotation types and select the correct one for your AI project!
A label is simply a designation of a part of an image with a classification. There are a number of different annotation and labeling types and it can be confusing to choose the appropriate one for your project. This article will define the most common types of labels, their properties, and when it is appropriate to use them. LabelFlow currently supports bounding boxes and polygons, but we are expanding to include more label types. Vote on the label type you’d like us to introduce next!
The simplest and most rudimentary type of label is the bounding box. A bounding box is a rectangle that surrounds the object of interest in an image. It is defined by two vertices which are diagonal to each other.
The simplicity of the label makes it easy for human labelers and smart tools to draw. It is also an easy output for computer vision AI models used in video annotation services and fast-paced systems like autonomous driving. This is because it doesn’t take the exact shape of the object into consideration and thus reduces computational complexity.
The polygon label is another popular type of label. It allows the labeler to draw an approximate shape of the object using a number of vertices. The quality of this label type is much better than a simple bounding box because it cuts down on the noise of the background behind the object.
The polygon labeling process takes more time and effort than a bounding box. This is true for a human labeler and a labeling machine in terms of computing time and power. Allocating more resources may seem trivial for a few labels, but it adds up quickly when we want to label hundreds or possibly thousands of images. However, the quality of the label more than makes it worth your while with increased AI performance for some models! Polygon labels are for example commonly used in the energy industry to label equipment defects. They are also used in the geospatial industry to label houses, blocks, monuments, and more.
Image tagging is different from the other types of image labels because it doesn’t label an area of interest within an image. Instead, it tags the whole image with a classification and doesn’t differentiate between pixels. This type of classification is very simple for annotators to label.
A cuboid is a rectangular prism where all sides are rectangles. It is basically a bounding box with a third dimension. It is relatively easy to draw and labeling platforms allow labelers to move the vertices to precisely cover the object of interest. The information from the third dimension is used to train AI models’ ability to perceive depth. This type of labeling is sometimes used in the automotive industry to help vehicles visually determine distances.
Lines and splines are unidimensional annotations that are used to indicate paths in an image. They are easy to draw and they provide very useful information. These labels are mostly used to train AI models that move along aisles, lanes, pathways, and more. The largest use case of this type of labeling is to train autonomous vehicles to recognize lanes on the road.
Landmark annotation is a type of label that places points on significant features of the object of interest. The annotation by itself is very simple and doesn’t require much effort, but placing it precisely could be a challenge. This type of annotation is commonly used to train facial recognition models.
Image segmentation, or segmentation annotation, is a complex type of image labeling where objects of interest are labeled down to the last pixel. Labeling with pixelwise annotation tools is difficult, time-consuming, and resource-intensive. However, it produces high-quality outputs that are very useful for medical care and computer vision training. There are basically two types of image segmentation techniques.
This type of segmentation annotates all objects of a certain class into one label. It won’t be possible to distinguish distinct objects from the label even if the objects are far apart or overlap in the image. This method is typically used in the medical industry for cancer diagnoses, where cells can be classified pixel by pixel and "cancerous" or "non-cancerous".
In this type of segmentation, each object (or instance of a class) will be labeled separately. Although similar objects will have the same class, it will be possible to distinguish between separate objects. Instance segmentation annotation tools are especially useful for AI models that specialize in counting objects.
All these label types have their pros and cons, but the most important thing is to select the most appropriate one for your AI project. The LabelFlow team is working on bringing more label types to our open annotation platform. Tell us which ones you want first here!