Image Annotation for Deep Learning

May 19, 2021

Defining image annotation

Image annotation gives machine learning models information about what a given image is showing. In most cases this information is applied to datasets by human annotators who highlight and label relevant parts of images in accordance with the needs of machine learning developers. Image annotation allows researchers to establish goals for their models, and to focus on how best to label images in order to produce the most accurate and functional results. This might mean addressing the best naming practices for labels, or how to cope with edge cases, such as unclear image sections.

Image annotation in practice

Image annotation is primarily carried out by human annotators. Annotators can use bounding boxes to highlight relevant objects in a given image. In the case of image annotation for livestock management AI this might mean locating each animal in a picture with a box and an associated label. Often annotators are required to label multiple objects, this would mean bounding boxes of one colour for animals and a separate colour for humans or vehicles. Different AI projects demand different types of annotation, with more or less detail added to images by human annotators.

Image annotation | Keymakr

Annotation types in focus

Image annotation can often mean straightforwardly describing an image. This is known as image classification. An image of livestock in a field might be labeled solely as “sheep”, for example.

Object detection goes further by adding bounding boxes to every object in an image. In the case of livestock AI training images this would mean boxing each animal, as well as any other salient objects in the frame, such as farm machinery.

For additional layers of detail annotators can segment every pixel in an image into a class. This process is defined by two terms:

  • Semantic segmentation: This annotation type achieves a higher degree of granularity than image classification and object detection. Annotators use an outlining tool to highlight the precise shape of a particular object. For busy street traffic images this might mean tracing the outline of a bus and assigning all of the pixels within it to a particular class e.g. “bus” or “large vehicle”. This process is repeated for each relevant object. The space between these shapes is also classified. In the traffic images this would mean the sidewalk, the sky, nearby buildings, and the road itself. All would be assigned a colour and a label. This annotation type is considerably more time consuming but has a significant advantage in terms of detail.
Semantic segmentation | Keymakr
  • Instance segmentation: This annotation type extends semantic segmentation by highlighting each occurrence of a particular object. For images of street traffic that means that each car is outlined and labeled with a different colour and name, e.g. “car 1”, “car 2”, etc. This additional information is useful for developers who are looking to distinguish between objects of interest in their training image data.
Instance segmentation | Keymakr

Specific image annotation techniques

In order to apply annotation methods to training images, annotators make use of a variety of annotation techniques. These techniques can be deployed to accurately label a wide range of images, helping training data to precisely model the complexity of the real world:

Image annotation techniques | Keymakr
  • Bounding boxes: Annotation platforms allow operators to drag boxes around items of interest. This is the most common annotation technique due to its speed and simplicity. However, it does not capture the exact shape of an object so is inherently less detailed.
  • Polygon annotation: This annotation type allows operators to capture complex shapes. It works by connecting together small lines with vertices in order to precisely trace the outline of any object. This technique is essential for semantic segmentation as it allows all pixels to be accurately assigned to a specific class.
  • Skeletal annotation: This technique is used to locate the position of limbs in images. Lines are drawn onto human or animal limbs and then connected at points of articulation, such as knee joints. This can allow machine learning models to interpret movement and analyse body positions.
  • Key points annotation: This technique uses points to locate important features. One common example of key point annotation is facial feature labeling, where points are used to identify noses, eyes, lips, etc. It can also be used to pinpoint important parts of buildings and other structures.
  • Lane annotation: This annotation method is primarily deployed to demarcate roads, railway lines, pipelines, and other linear structures. Annotators trace the shape of these features using annotation platform tools.
Keymakr Demo

High performance annotation platforms

The wide variety of annotation types and techniques empowers AI developers by allowing for the creation of bespoke datasets that fit the needs of any project. Realising the full potential of annotation often requires access to the best annotation platforms and tools. Keymakr’s proprietary annotation platform is an example of a system designed to optimize annotation. Essential platform features include:

  • Automatic object labeling accelerates annotation by transforming box labels into polygon outlines.
  • Classification options make annotation workflows efficient and precise by providing annotators with easy access to label maps, classes, attributes, and specific annotation types.
  • Workflows can be made visible through user friendly interfaces where managers can track the progress of tasks and specific workers. Jobs can be assigned through the platform and detailed productivity metrics can provide a holistic view of the status of any part of the project.
  • Keymakr’s platform allows managers to review annotations, flag errors, and grade accuracy. These grades can then be used to assign annotation tasks in the future.
Data annotataion | Keymakr

Optimizing image annotation with smart outsourcing

Image annotation is an essential part of most AI development cycles. However, it can be challenging for many companies to access enough of the high quality training data their models demand. Finding the right data and assembling a competent annotation workforce can often be a time consuming distraction. Outsourcing annotation to dedicated service providers, like Keymakr, can allow AI innovators to focus on what they do best whilst ensuring a steady flow of precisely labeled images for computer vision training. Keymakr offers developers a number of key advantages:

  • The image annotation workforce: Crowdsourced annotations are reliably cheap but are less reliably accurate. It can be difficult to ensure precise annotation across a large, distributed workforce. On the other hand building an in-house annotation operation can be expensive and a management challenge for AI companies, distracting from the core development mission. Keymakr lets developers enjoy the benefits of centrally located teams of annotators, without the burden of hiring. These teams are overseen by experienced managers, leading to exceptional levels of accuracy in image annotations.
  • Scalability and cost effectiveness: Professional annotation services allow AI companies to scale annotation in line with changing training data needs. In house annotation can be hard to increase when more images are required and can be inefficient when data needs are reduced. Outsourcing is the most cost effective option for AI companies. Keymakr is able to respond to increased image annotation demands quickly, whilst maintaining quality throughout datasets.
  • Troubleshooting and quality control: Crowdsourced annotation, spread across multiple time zones and cultural contexts, can lead to problems with communication that impede troubleshooting. Keymakr’s annotation teams are centrally managed allowing for quick feedback and responsive quality control procedures.
Keymakr Demo

Unlock image annotation with Keymakr

Outsourcing image annotation to Keymakr is a cost effective way of ensuring that your computer vision project has access to the best quality training data. Our expertise and experience allow us to create bespoke datasets that meet the most demanding of needs and schedules. Contact a team member to book your personalized demo today.

Great! You've successfully subscribed.
Great! Next, complete checkout for full access.
Welcome back! You've successfully signed in.
Success! Your account is fully activated, you now have access to all content.