How Data Annotation Supports Facial Recognition Projects
Facial recognition technology is helping to create a safer world. Computer vision based AI models have the capacity to protect homes and secure public spaces by identifying threats and alerting security services. However, as this technology moves into the real world significant challenges are beginning to emerge.
Until now facial recognition technology has primarily been used for mobile devices, and consumer oriented apps. In these contexts errors, inaccuracies, and biases are, at worst, an annoyance. But when AI powered cameras are deployed in complex security scenarios misidentifications and malfunctions could lead to the wrong person being arrested or a crime going unsolved.
Accuracy and quality in video and image training datasets can play an important role in ensuring that facial recognition models function effectively. Keymakr has been working with a security AI company to solve developing issues with smart annotation solutions.
Identifying Needs
The security camera AI client required a dataset creation process that could address some of the key difficulties associated with security AI annotation:
- Annotating video: Computer vision models for security cameras must have the capacity to process and identify faces in video footage. Annotating video is a daunting task, even for larger technology companies. The multiplicative nature of video frames means that precise annotating hours of video can be extremely time consuming and labour intensive. This can lead to bottlenecks in development and can mean that valuable expertise and leadership are diverted from the main purpose of the project.
- Accuracy: Facial recognition technology must display high levels of accuracy if it is to be deployed for security purposes. Absolute precision in annotation is therefore essential to ensure that security AI models are not reproducing errors that could lead to serious real world consequences.
- Emotion recognition: Part of the function of facial recognition security AI is to identify emerging threats through sentiment analysis and emotion recognition. This will allow AI powered security cameras to act as early warning systems, detecting threatening behaviour and warning security services.
- Avoiding bias: Facial recognition technology is not immune from bias and the errors bias can cause. Smart security AI development must consider how models cope with difference, including age, gender, and race.
Meeting Needs Through Collaboration
Professional annotation services can play an important role in overcoming the challenges of facial recognition AI development. Keymakr collaborated with the client to address their specific needs:
- Professional video annotation: Skilled teams of annotators, managed in-house by experienced supervisors, can significantly reduce the burden of video annotation. By utilising proprietary annotation tools Keymakr’s team annotated over one thousand hours of video footage, from multiple angles and in a variety of conditions.
- Verifying classifications: Accuracy and quality are primarily assured by a robust quality control process. Keymakr provides three layers of human quality verification as well as automated quality checking features. This rigorous focus on precision leads to extremely accurate final datasets.
- Tracking facial features: Quality facial annotation is the raw material for facial and emotion recognition. Keymakr collected and annotated over twenty million images of different individuals and marked key facial features. Different attributes could then be assigned to the images, emotion being among them.
- Finding the right data: Bias in data can be overcome by collecting the right data and annotating it correctly. Keymakr assembled a diverse selection of images and videos, taking account of differences in age, gender, and race, as well as complicating factors for identification like hats or accessories.
Professional Annotation Services Help Security AI Development
Keymakr utilises proprietary technology, in-house teams of annotators, and multiple layers of quality control to provide data annotation that is precise, affordable and scalable.