Recent advances in face recognition modelling have been achieved thanks to the availability of large datasets and deep learning models. However, large-scale data collection via the internet raises legal, ethical and privacy concerns. In this way, alternative methods such as generating synthetic face datasets and using synthetic images to train face recognition models are being considered. Nevertheless, the generation of synthetic datasets with sufficient variations remains an active area of research.
This approach has been pursued at the Winter Conference on Applications of Computer Vision (WACV) by organising a Face Recognition Challenge in the Era of Synthetic Data (FRCSyn). This is the first international challenge that aims to explore the use of synthetic data in face recognition to address existing limitations in technology. Specifically, the FRCSyn addresses issues related to data privacy, demographic biases and performance limitations in complicated scenarios (significant age disparities, pose variations and occlusions). The results achieved in the FRCSyn contribute significantly to the application of synthetic data to improve face recognition technology. Participants developed and applied intelligent strategies for using synthetic datasets to train face recognition models. All models submitted by the participants were evaluated using benchmark datasets and ranked according to their performance on these datasets.
The potential of synthetic data in identity verification: Facephi’s results
We at Facephi participated in the FRCSyn to test whether synthetic data could replace real data in face recognition training, whether it could mitigate the known limitations of facial biometrics, and the ethical and legal challenges associated with large-scale data collection. This is done using state-of-the-art neural network architectures and various data augmentation techniques.
Our success has been reflected in the results, standing out from the other participants and demonstrating exceptional performance in the first quality task of the synthesised dataset. When assessing the ability to mitigate bias and adapt to different datasets, we managed, through the use of synthetic data to train the FR systems, to be among the best in mitigating and reducing bias. Moreover, in the second task of the challenge, in which we were provided freedom of choice for model, dataset and training, we also achieved competitive results by proving the versatility and effectiveness of our approach in different scenarios. The judges and experts praised the originality and quality of the methods used, recognising our innovative approach and significant contribution to the advancement of research in the generation of synthetic datasets for face recognition.
