Case study – Provision of thousands of photosets of respectively one person, covering a period of min. 5 years to max. 20 years.

Thousands of Clickworkers from various continents sent numerous photos with their faces clearly visible. These photos have spanned a period of a minimum of 5 years to a maximum of 20 years in the past. With these photosets, an AI system was trained to be capable of clearly recognizing and identifying faces of all ethnicities and genders over a lifetime. In this article, clickworker will present a brief overview of what goes on behind the scenes of executing a task.

The Challenge

Facial recognition systems using AI are increasingly being used to leverage the uniqueness of a face as a biometric factor for identity verification and authentication in online login processes. Biometric facial recognition is one of the most reliable authentication systems as it uses unique mathematical and dynamic patterns, unlike traditional solutions such as verification emails, passwords, fingerprints, or even simple selfies. The algorithms on which these AI systems are based must be trained with an enormous amount of data in the form of photographs and/or videos of people until they are able to identify people unambiguously and without error on the basis of their faces using a machine learning process.

During the learning process, a multilayer neural network is used to process the training data. This network adjusts its face recognition parameters until a person can be clearly identified. This learning process requires not only large amounts of photographs and videos of people but also a wide variety of people depicted, corresponding to the diversity of people in the regions where the system will be deployed. In addition, in order to train a biometric face recognition system, the training data must consist of photos of people whose faces can be seen in different sizes and from various perspectives and angles. When training the system, one must also keep in mind that when it is used to authenticate people, it must be able to recognize a face at all times, even if the face changes naturally over the years, sometimes to a greater or lesser extent.

The Solution

We closely consulted with the customer before setting up a tailored project on our in-house online platform. For our registered crowdworkers, referred to as Clickworkers, this resulted in paid jobs/tasks. 850,000 Clickworkers who fit the client's selected demographics were assigned these tasks.

Thousands of Clickworkers work on the project in accordance with the instructions obtained from the concise and descriptive task briefing. After specifying their ethnicity, the first step for each Clickworker who has accepted the task involves creating two new, short videos of themselves. In this case, they film their face- with and without glasses. While doing so, they slowly move their head in all directions and say a short sentence.

In the second step, each of the Clickworkers uploaded these two videos, as well as 60 to 200 existing digital photos of themselves — where their face is clearly recognizable — as a set to our platform. None of the photos from the set were taken on the same day, no photo is repeated, and covered in total a time period of min. 5 years to max. 20 years. The photos differ in terms of perspective or angle from which the person’s face is seen, styling (e.g. hairstyle, clothing, glasses, makeup), facial expression, and lighting conditions.

To ensure the correct implementation of the specifications, all the uploaded videos and photos were checked thoroughly by our quality management team and selected accordingly. After being checked, the flawless sets are then transferred to the customer directly via an API connection.

This quickly and effectively provided the software developer access to over 300,000 face pictures and over 6,000 high-diversity videos. In order to train an AI system to accurately detect faces until the error rate approaches zero and the system can be utilized for safe online authentication, the software company used this data as training data.

Project Data

To train an AI system of such magnitude properly, clickworker had to produce a significant amount of data.

1. Number of photos:>300.000 photos (50 – 200 per Clickworker/set)
2. Time frame of the photos: Photos should cover a period of min. 5 years to max. 20 years per set and about 10 diverse photos per year.
3. Number of videos:?>6.000 (2 pro Clickworker/set) of approx. 30 seconds each
4. Proportion of Clickworkers per ethnicity:?Africa 20%, South Asia 20%, Far East 20%, Latin America 20%, Caucasia 10%, Other 10%.
5. Photo format and size:?Minimum size of the depicted head on the photos: 200 x 200 pixels, jpg or heic, landscape or portrait format
6. Interface versions of jobs:?Clickworker App (Smartphone Version), Clickworker Workplace (Desktop Version)
7. Tasks:

• Specifications regarding own ethnicity (via dropdown selection)
• Creation of two short portrait selfie videos with and without glasses (approx. 30 seconds each)
• Provision/upload of 60 – 200 photos from the mentioned period of time, on which their face is easy to recognize
• Reading confirmation and declaration of consent to the job terms and conditions

8. Quality assurance:?Quality check by clickworker’s quality management team

9. Data transfer:?Data transfer via API

Project Workflow — In Brief

A well-planned project management workflow was carefully designed by clickworker to execute the project, increasing efficiency and enhancing outcomes.

1. ?Project meeting with the customer: The resulting tasks are defined and recorded in a precise briefing for the Clickworkers.
2. ?Setting up the project by clickworker: The photo creation tasks become visible as individual jobs on the clickworker platform for qualified Clickworkers.
3. Processing of the jobs by Clickworkers: Numerous Clickworkers of the specified ethnicities accept the jobs in parallel, create the photos and videos according to the briefing, and upload them to the clickworker platform with the required information.
4. Verification of data quality: The clickworker quality management team checks the photos and videos for compliance with the specifications/briefing.
5. Transmission of all correct photo sets with videos to the customer: Occurs via an API connection.

Benefits of using clickworker

• Quick acquisition of high diversity and high-quality application-specific AI training data
• Global data sourcing and market coverage across all continents
• Access to a broad-based crowd from all ethnic origins, ages and genders
• Customer-specific project implementation and execution
• Quality assured results
• Easy data transfer

Conclusion

Facial recognition can be used for a variety of applications that can affect our day-to-day life. The first and foremost category is of course security and law enforcement. Face recognition software can be used to identify criminals or match a person’s face to their passport at a border check. Moreover, it is a useful tool to find missing children. By adding their faces to the database, they can be identified more easily.

But these are not the only applications. Face recognition also offers great advantages in marketing: By matching faces to customers, they can be targeted in a more personalized way. This means that customers receive advertising that is catered to their person rather than generalized. Another potential field of application is mobile phone security. Face recognition can be used as the unlock feature, making it more difficult for other people to get into someone else’s phone.

All these and numerous other areas will continue to be developed and improved in the future. Our face recognition training data services play an important role in that progress.