README.md

VGGFace2 Extension

This repo contains a Keras implementation of the paper,

VGGFace2: A dataset for recognising faces across pose and age (Cao et al., FG 2018).

Dependencies

• Python 2.7.15
• Keras 2.2.4
• Tensorflow 1.8.0

Data

The dataset used for the experiments are

• VGGFace2 [1]

Model

Keras Model (https://drive.google.com/file/d/1AHVpuB24lKAqNyRRjhX7ABlEor6ByZlS/view?usp=sharing),

Note:

This model is trained with a slightly different tight crops, but I have also tested on the tight crops (as we did in the paper), and am able to get similar results (on both IJBB and IJBC).

Dataset	Feat dim	Pretrain	TAR@FAR = 1e-5	TAR@FAR = 1e-4	TAR@FAR = 1e-3	TAR@FAR = 1e-2	TAR@FAR = 1e-1
IJBB	512	N	0.64	0.78	0.88	0.94	0.98
IJBC	512	N	0.72	0.82	0.90	0.95	0.98

Testing the model

To test a specific model on the IJB dataset, for example, the model trained with ResNet50 trained by sgd with softmax, and feature dimension 512

• python predict.py --net resnet50 --batch_size 64 --gpu 2 --loss softmax --aggregation avg --resume ../model/resnet50_softmax_dim512/weights.h5 --feature_dim 512

Citation

@InProceedings{Cao18,
  author       = "Q. Cao, L. Shen, W. Xie, O. M. Parkhi, A. Zisserman ",
  title        = "VGGFace2: A dataset for recognising face across pose and age",
  booktitle    = "International Conference on Automatic Face and Gesture Recognition, 2018.",
  year         = "2018",
}