The Problem
As facial recognition systems become more widely deployed across security systems and identity verification, AI teams face a significant challenge: ensuring AI models can distinguish between genuine human faces and increasingly sophisticated presentation attacks and spoofing attempts.
A global technology company approached DATAmundi to support the creation of a large-scale, highly controlled video dataset designed to improve the performance of its computer vision applications.
The goal was to collect authentic video footage that reflected the wide range of real world conditions an AI system might encounter, training the machine to recognize real and spoof scenarios using both synthetic and real identities.
The project required far more than simple video capture. Every recording had to follow strict specifications covering lighting conditions, camera distance, viewing angles, accessories, movement, environmental context and metadata. Consistency in these conditions were critical to ensuring the resulting dataset represented realistic attack scenarios and everyday environments.
For the real and synthetic identities, the client required a diverse pool of contributors across age, gender and ethnicity, with carefully controlled environmental conditions and consistent data quality across collection sites in Europe and North America.
The Solution
DATAmundi designed and managed an end-to-end data acquisition program, coordinating contributor recruitment, onsite collection, quality assurance and metadata generation.
Using its DATAtalent global contributor network, DATAmundi recruited more than 100 participants to create authentic facial video data. Alongside real contributors, hundreds of synthetic identities were prepared. For both real and synthetic identities, masks were created using a variety of materials, enabling the creation of realistic attack scenarios for model training.
Contributors were trained using detailed guidelines before attending managed collection sessions in controlled indoor environments. Each contributor participated in a recording session, while additional spoof scenarios using mask-based facial presentations were captured by our trained moderators to ensure consistent, high-quality data across all test conditions. Including:
- Different lighting conditions (such as bright, dim and low light)
- Variable camera distances and viewing angles
- Accessories such as hats, glasses, wigs, and head coverings
- Different mask materials and presentation methods
- Natural movement including head rotation and changing facial orientation
All video capture followed predefined technical specifications using standard mobile devices to better represent authentic deployment environments.
Alongside video collection, DATAmundi generated and annotated metadata describing every recording. This included environment, lighting and lux levels, camera positioning, and other attributes required for downstream machine learning workflows.
Throughout the project, QA ensured every recording met the client’s exact specifications. Where recordings failed validation, contributors were rebriefed and sessions repeated to maintain dataset consistency and accuracy.
The program was delivered across multiple collection locations using dedicated data acquisition specialists responsible for coordinating equipment, contributor management, logistics and QC.
The Result
The resulting dataset provided the client with thousands of high-quality video assets representing both genuine and spoof identity scenarios.
The project demonstrated DATAmundi’s ability to manage complex multimodal AI data program that combine:
- DATAtalent global contributor recruitment
- Controlled real world data acquisition
- Synthetic and real data
- Rich metadata generation and annotation
- Human-led QA
The dataset reflected the variability AI systems encounter in production environments rather than ideal laboratory conditions, crucial in developing facial recognition and anti-spoofing applications.
If you want to discuss how to create and collect multimodal egocentric or exocentric training data for your AI model
