davidsandberg/facenet

Implements the FaceNet architecture using triplet loss and softmax training to generate compact 128-dimensional face embeddings for recognition and clustering tasks. Uses Inception-ResNet-v1 backbone with MTCNN for face detection and alignment, and provides pre-trained models trained on CASIA-WebFace and VGGFace2 datasets achieving 99.65% LFW accuracy. Supports both training from scratch and inference with TensorFlow r1.7+, enabling deployment of face verification and identification pipelines.

14,313 stars. No commits in the last 6 months.