Abstract
In this paper, we introduce a new large-scale face database from KIST, denoted as K-FACE, and describe a novel capturing device specifically designed to obtain the data. The K-FACE database contains more than 17.5 million high-quality images of 1,000 subjects selected by considering the ratio of gender and age groups. It includes a variety of attributes, including 27 poses, 35 lighting conditions, three expressions, and occlusions by the combination of five types of accessories. As the K-FACE database is systematically constructed through a hemispherical capturing system with elaborate lighting control and multiple cameras, it is possible to accurately analyze the effects of factors that cause performance degradation, such as poses, lighting changes, and accessories. We consider not only the balance of external environmental factors, such as pose and lighting, but also the balance of personal characteristics such as gender and age group. The gender ratio is the same, while the age groups of subjects are uniformly distributed from the 20s to 50s for both genders. The K-FACE database can be extensively utilized in various vision tasks, such as face recognition, face frontalization, illumination normalization, face age estimation, and three-dimensional face model generation.
Introduction
In computer vision, researchers often face a trade-off between two types of datasets. Web-crawled "in-the-wild" data (e.g., CelebA, VGGFace2) provides massive identity counts but suffers from noisy labels, poor image quality, and demographic imbalances. Conversely, controlled "lab" data (e.g., Multi-PIE) offers precise labels and environmental control but lacks the scale required for modern deep learning models.
K-FACE bridges this gap by providing a massive, high-resolution dataset of 1,000 Korean subjects with the precision of a laboratory environment. By systematically varying pose, lighting, and accessories, K-FACE enables researchers to isolate and analyze specific causes of model failure in unconstrained real-world scenarios.
The Hemispherical Capturing System
To achieve this balance of scale and control, the team developed a specialized hemispherical capturing system. This rig is designed to simulate a wide range of environmental conditions with high repeatability.
The hardware setup includes:
- 27 FHD DSLR Cameras: Canon EOS 100D units are positioned at specific vertical and horizontal intervals to capture 27 distinct poses simultaneously.
- 10 Lighting Devices: These panels create 35 distinct illumination conditions, ranging from 0 to 1,000 lux, coming from multiple directions.
- Synchronized Triggering: All cameras fire at once, capturing a subject from every angle in a single moment to ensure temporal consistency across poses.
Dataset Characteristics and Diversity
K-FACE stands out due to its sheer volume and demographic fairness. While traditional lab datasets like Multi-PIE contain approximately 750,000 images, K-FACE offers over 17.5 million images at a high resolution of 2592 × 1728 pixels.
Demographic Balance
Unlike web-crawled datasets that skew toward specific demographics, K-FACE was built with intentionality:
- Gender: A perfect 1:1 ratio of male to female subjects.
- Age: Uniform distribution across four decades (20s, 30s, 40s, and 50s), with a 25% quota for each group to prevent age bias in AI models.
Systematic Variations
The dataset covers a comprehensive "grid" of variations:
- 27 Poses: Defined by specific yaw and roll angles.
- 35 Lighting Conditions: Varying both directionality and intensity.
- 3 Expressions: Neutral, Smile, and Disgust.
- Accessories: Six primary categories, including complex combinations like masks, glasses, and hats. This is particularly relevant for modern "masked face recognition" research.
Conclusion
K-FACE provides a robust foundation for advancing face recognition, 3D face restoration, and age estimation. While currently limited to a Korean demographic, its systematic diversity and uniformity offer a powerful tool for analyzing the impact of environmental factors on computer vision performance. Future work may involve expanding the ethnic diversity of the database and incorporating more complex real-world factors like motion blur.
Summary updated