Large-scale CelebFaces Attributes (CelebA) Dataset

Liu, Ziwei; Luo, Ping; Wang, Xiaogang; Tang, Xiaoou

2015

CelebFaces Attributes Dataset (CelebA) is a large-scale face attributes dataset with more than 200K celebrity images, each with 40attribute annotations. The images in this dataset cover large pose variations and background clutter. CelebA has large diversities, large quantities, and rich annotations, including: 10,177 number of identities, 202,599 number of face images, and 5 landmark locations, 40 binary attributes annotations per image. Each image in the dataset captures various facial features and accessories, such as eyeglasses, smiling, or bangs. Additionally, five landmark points (e.g., eyes, nose, mouth corners) are provided per image, facilitating tasks like facial alignment. Also, a wide range of poses, expressions, and occlusions are included, reflecting real-world conditions and enhancing the robustness of models trained on this data. The dataset has a size of 25,3 kB and is organized in theree main components:

• Images:

  • In-the-Wild Images: Original images depicting celebrities in various environments and conditions.​

  • Aligned and Cropped Images: Faces have been aligned and cropped to a consistent size, facilitating standardized analysis.​

• Annotations:

  • Landmark Locations: Coordinates for five key facial points (left eye, right eye, nose, left mouth corner, right mouth corner) per image.​

  • Attribute Labels: Binary labels indicating the presence or absence of 40 distinct facial attributes for each image.​

  • Identity Labels: Each image is associated with an identity label, linking it to one of the 10,177 unique individuals.​

• Evaluation Partitions:

  • The dataset is divided into training, validation, and test sets, enabling standardized evaluation of algorithms.