Image annotation: Best practices for ML model training
Artificial intelligence (AI) and machine learning (ML) are fast becoming integral to every industry. From manufacturing to financial services to healthcare, these technologies rely on big data to train models that can make decisions. Effective ML models rely on large-scale data annotation.
Research suggests that the global data annotation market will expand at a CAGR of 26.6% between 2022 and 2030. Data annotation is the building block of AI.
This article will explain the image annotation best practices.
What is image annotation?
Image annotation is the process of manually labelling the images in a dataset. This helps train artificial intelligence and machine learning computer vision (CV) models.
Labelled images form a dataset from which the CV model learns. While the model may be less accurate after the first pass, its accuracy increases as we use more data to train it.
How does it help in machine learning?
CV models based on machine learning require large datasets where each image must have proper labelling. For example, image annotation in the manufacturing industry could show minor faults in the key parts and the final product with greater accuracy.
Types of image annotation
The five common types of image annotation methods are bounding box, polygon, polyline, key points, and primitives.
Bounding box –
It is a box around an object in an image. This object could be a car, a tree, or a glass on a table. The annotator applies a unique label to this object and feeds this data into a machine-learning CV model. The model uses this manual annotation to label the rest of the similar objects in the image automatically.
A polygon gives the annotator freedom to draw freehand around an object. For example, a polygon could be the right approach to annotate a tumour in a human body.
A polyline is great for annotating an object or a pattern that continues through the series of images. This could be a railway line or a road. The CV model uses this data to label the object from one image to the other.
Key points –
Key points help identify and label unique shapes, such as a human face. It outlines and pinpoints specific features. The CV model uses this key point data and applies it to other images, thus speeding up the process.
Primitives are used for specific annotation to make templates of 3D cuboids, rotating boxes, pose estimation skeletons, and other unique shapes.
Image annotation best practices
Every dataset requires unique labelling instructions. To achieve high-quality image annotation, annotators must apply these best practices.
Tight bounding boxes
The bounding boxes should be small enough to engulf the entire object, yet not too tight. A bounding box that is too tight may cut out a portion or relevant pixels of an object.
Label or tag occluded objects
An occluded object is one that is partially hidden in an image. Ensure that you label these objects as if they were in full view. Never draw a bounding box only on the visible part of the object. As long as you label all the objects, overlapping bounding boxes should not be a concern.
All objects have some degree of variation, which is why annotation requires high consistency. For example, the criteria to define a tumour as malignant must be uniform across all the images.
Tag all objects in each image
Computer vision models learn the pixel patterns and correspond them to an object in an image. Image annotators must label every occurrence of an object across images to identify it with greater precision.
Label objects of interest in their entirety
Annotators must label the full object and not leave any portion out of the bounding box. It is also important to label all objects across categories in an image. Failure to do so may result in low-quality ML models.
Keep the labelling instructions clear
Clear labelling instructions help in improving the model in the future. Other labelling team members may rely on your instructions in the future.
Use label names in images
Being specific about the labelling makes any relabelling in the future easier. For example, if you are building a commercial vehicle detector for traffic management, it is best to label the trucks based on the brand, carrying capacity, or the number of wheels. This saves you from relabelling the complete dataset in the future.
For organisations on the digital transformation journey, agility is key in responding to a rapidly changing technology and business landscape. Now more than ever, it is crucial to deliver and exceed on organisational expectations with a robust digital mindset backed by innovation. Enabling businesses to sense, learn, respond, and evolve like a living organism, will be imperative for business excellence going forward. A comprehensive, yet modular suite of services is doing exactly that. Equipping organisations with intuitive decision-making automatically at scale, actionable insights based on real-time solutions, anytime/anywhere experience, and in-depth data visibility across functions leading to hyper-productivity, Live Enterprise is building connected organisations that are innovating collaboratively for the future.
How can Infosys BPM help?
Infosys BPM helps data science teams build high-quality training data for AI at scale. The approach involves a human plus a platform model. Key sectors for which Infosys BPM supports AI/ML models are CPG, retail, media, railways, oil & gas, insurance, healthcare, and financial services.
Read more about the annotation service at Infosys BPM.