合成面部识别基准数据集

本研究开发了一个大规模的合成面部识别基准数据集，包含48,000对合成面部图像（10,200个独特的合成面部）和555,000个人类注释。数据集通过神经面部生成器生成，每个感兴趣的属性独立修改，其他属性保持不变。人类观察者提供感知身份相似性的基本事实。该数据集用于验证面部识别算法在种族和性别偏见方面的性能，揭示了算法在黑人和东亚人群子集中的准确性较低。此外，数据集还能量化属性感知变化对模型报告的面部身份距离的影响。该数据集的应用领域主要集中在面部识别技术的公平性和准确性评估，旨在解决算法偏见问题。

This study developed a large-scale synthetic facial recognition benchmark dataset, which consists of 48,000 pairs of synthetic facial images (covering 10,200 unique synthetic faces) and 555,000 human annotations. The dataset is generated using a neural facial generator, where each attribute of interest is independently modified while all other attributes remain unchanged. Human observers provided ground truth labels for perceived identity similarity. This dataset is designed to validate the performance of facial recognition algorithms with respect to racial and gender bias, revealing that the algorithms achieve lower accuracy in the Black and East Asian population subsets. Additionally, the dataset enables quantification of the impact of attribute-perceived changes on the facial identity distances reported by models. The primary application scenarios of this dataset focus on fairness and accuracy evaluation of facial recognition technologies, with the goal of addressing algorithmic bias issues.