Image Annotation 101

What does it mean to annotate an image?

Image annotation is defined as the task of annotating an image with labels, typically involving human-powered work and in some cases, computer-assisted help. Labels are predetermined by a machine learning engineer and are chosen to give the computer vision model information about what is shown in the image. The process of labeling images also helps machine learning engineers hone in on important factors that determine the overall precision and accuracy of their model. Example considerations include possible naming and categorization issues, how to represent occluded objects, how to deal with parts of the image that are unrecognizable, etc.

How do you annotate an image?

From the example image below, a person applies a series of labels by applying bounding boxes to the relevant objects, thereby annotating the image. In this case, pedestrians are marked in blue and taxis are marked in yellow, while trucks are marked in yellow. This process is then repeated and depending on the business use case and project, the quantity of labels on each image can vary. Some projects will require only one label to represent the content of an entire image (e.g., image classification). Other projects could require multiple objects to be tagged within a single image, each with a different label (e.g., bounding boxes).


Bounding boxes applied to identify vehicle types and pedestrians

What are the different types of image annotation?

In order to create a novel labeled dataset, data scientists and ML engineers have the choice between a variety of annotation types. Let’s compare and summarize the three common annotation types within computer vision: 1) classification 2) object detection and 3) image segmentation.

  • With whole-image classification, the goal is to simply identify which objects and other properties exist in an image.

  • With image object detection, you go one step further to find the position (bounding boxes) of individual objects.

  • With image segmentation, the goal is to recognize and understand what's in the image at the pixel level. Every pixel in an image belongs to at least one class, as opposed to object detection where the bounding boxes of objects can overlap.


Different types of image annotation

Whole image classification provides a broad categorization on an image and is a step up from unsupervised learning as it associates an entire image with just one label. A distinct benefit it is by far the easiest and quickest to annotate out of the other common options. Whole-image classification is also a good option for abstract information such as scene detection and time of day.

Bounding boxes, on the other hand, are the standard for most object detection use cases and requires a higher level of granularity than whole-image classification. Bounding boxes provide a balance between quick annotation speed and targeting items of interest.

For specificity, image segmentation is chosen to support use cases in a model where you need to definitively know whether or not an image contains the object of interest and also what isn’t an object of interest. This is in contrast to other annotation types such as classification or bounding boxes that may be faster in nature but less accurate.

How does a training data platform support complex image annotation?

Image annotation projects begin by identifying and instructing annotators to perform the annotation tasks. Annotators must be thoroughly trained on the specifications and guidelines of each annotation project, as every company will have different requirements.

Once the annotators are trained on how to annotate the data, they will begin annotating hundreds or thousands of images on a training data platform dedicated to image annotation. A training data platform is software that is designed to have all the necessary tools for the desired type of annotation and is commonly equipped with multiple tools which allow you to outline complex shapes for image annotation.

In addition, training data platforms typically include additional features that specifically help optimize your image annotation projects which include:

  • High-performance annotation tools:

    An important point to consider and test is whether or not the tools provided by the training data platform you are testing can support a high number of objects and labels per image without sacrificing loading times. At Labelbox, our vector pen tool allows you to draw freehand as well as straight lines. Blazingly fast and ergonomic drawing tools help reduce the time-consuming nature of having pixel-perfect labels consistently.


Labelbox's pen tool illustrated

  • Customization based on ontology requirements:

    The ability to configure the label editor to your exact data structure (ontology) requirements, with the ability to further classify instances that you have segmented. Ontology management includes classifications, custom attributes, hierarchical relationships and more.


Configure the label editor to your exact data structure (ontology) requirements.

  • A streamlined user interface which emphasizes performance for a wide array of devices:

    An intuitive design helps lower the cognitive load on labelers which enables fast labeling. Even on lower spec PCs and laptops, performance becomes critical for professional labelers who are working in an annotation editor all day.


A simple, intuitive UI reduces friction

  • Seamlessly connect your data via Python SDK or API:

    Stream data into your training data platform and push labeled data into training environments like TensorFlow and PyTorch. Labelbox was built to be developer friendly and API-first so you can use it as infrastructure to scale up and connect your ML models to accelerate labeling productivity and orchestrate active learning.


Simplified data import without writing and maintaining your own scripts

  • Benchmarks & Consensus:

Quality is measured by both the consistency and the accuracy of labeled data. The industry standard methods for calculating training data quality are benchmarks (aka gold standard), consensus, and review. As a data scientist in AI, an essential part of your job is figuring out what combination of these quality assurance procedures is right for your ML project. Quality assurance is an automated process that operates continuously throughout your training data development and improvement processes. With Labelbox consensus and benchmark features, you can automate consistency and accuracy tests. These tests allow you to customize the percentage of your data to test and the number of labelers that will annotate the test data.

Benchmarks overview

Benchmarks in action, highlighting the example labeled asset with a gold star

  • Collaboration and Performance Monitoring:

Having an organized system to invite and supervise all you labelers during an image annotation project is important for both scalability and security. A training data platform should include granular options to invite users and to review the work of each one.

With Labelbox, setting up a project and inviting new members is extremely easy, and there are many options for monitoring their performance, including statistics on seconds needed to label an image. You can implement several quality control mechanisms, including activating automatic consensus between different labelers or setting gold standard benchmarks.


Seamless collaboration between data science teams, domain experts, and dedicated external labeling teams