VGGFace2 is a large-scale face recognition dataset developed to support research on facial recognition across variations in pose, age, illumination, and identity. It consists of 3.31 million images covering 9,131 subjects, with an average of over 360 images per subject. The dataset was collected from Google Image Search, ensuring a wide diversity in ethnicity, profession, and real-world conditions. It is split into a training set with 8,631 identities and a test set with 500 identities, making it suitable for benchmarking and large-scale model training. Alongside the dataset, the repository provides pre-trained models based on ResNet-50 and SE-ResNet-50 architectures, trained with both MS-Celeb-1M pretraining and fine-tuning on VGGFace2. These models achieve strong verification performance on benchmarks such as IJB-B and include variants with lower-dimensional embeddings for compact feature representation. The project also includes preprocessing tools, face detection scripts, and etc.
Features
- Dataset of 3.31 million images across 9,131 subjects
- Training set (8,631 identities) and test set (500 identities)
- Covers wide variations in age, pose, illumination, and ethnicity
- Pre-trained models in ResNet-50 and SE-ResNet-50 architectures
- Models available in Caffe, MatConvNet, PyTorch, and Keras
- Includes preprocessing scripts with MTCNN and evaluation examples
Project Samples
