# VGG Face 2 Models

- use the VGG `SE-ResNet-50-256D` .caffemodel
- mean values and size in prototxt
- training used faces detected with MTCNN using 0.3 scale bbox expansion, thresholded at borders
- for megapixels, we are using the opencv dnn face detector which has
- variation for using OpenCV DNN face detector on test set yields

| Detector | Center | Width | Height |
|:-:|:-:|:-:|:-:|
| OpenCV DNN | 0.0271% | 3.440% | 3.306% |
| MTCNN PyTorch | - | - | - |
| MTCNN TensorFlow | - | - | - |
| MTCNN Caffe | - | - | - |

### VGG Face 2 TAR@FAR

- From <https://github.com/ox-vgg/vgg_face2>
- http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/
- TAR: True Acceptance Rate, FAR: False Acceptance Rate

| Architecture   | Feat dim | Pretrain | TAR@FAR = 0.001 | TAR@FAR = 0.01 | Model Link |
|:-:|:-:|:-:|:-:|:-:|:-:|
|   ResNet-50    | 2048 |  N  | 0.878 | 0.938 | [Caffe](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/caffe/resnet50_scratch_caffe.tar.gz), [MatConvNet](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/matconvnet/resnet50_scratch_mat.tar.gz), [TF], [PyTorch](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/pytorch/resnet50_scratch_pytorch.tar.gz) |
|   ResNet-50    | 2048 |  Y  | 0.891 | 0.947 | [Caffe](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/caffe/resnet50_ft_caffe.tar.gz), [MatConvNet](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/matconvnet/resnet50_ft_mat.tar.gz), [TF], [PyTorch](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/pytorch/resnet50_ft_pytorch.tar.gz) |
| SE-ResNet-50   | 2048 |  N  | 0.888 | 0.949 | [Caffe](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/caffe/senet50_scratch_caffe.tar.gz), [MatConvNet](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/matconvnet/senet50_scratch_mat.tar.gz), [TF], [PyTorch](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/pytorch/senet50_scratch_pytorch.tar.gz)|
| SE-ResNet-50   | 2048 |  Y  | 0.908 | 0.956 | [Caffe](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/caffe/senet50_ft_caffe.tar.gz), [MatConvNet](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/matconvnet/senet50_ft_mat.tar.gz), [TF], [PyTorch](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/pytorch/senet50_ft_pytorch.tar.gz)|
| ResNet-50-256D | 256  |  Y  | 0.898 | 0.956 | [Caffe](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/caffe/resnet50_256_caffe.tar.gz), [MatConvNet](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/matconvnet/resnet50_256_mat.tar.gz), [TF], [PyTorch](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/pytorch/resnet50_256_pytorch.tar.gz) |
| ResNet-50-128D | 128  |  Y  | 0.904 | 0.956 | [Caffe](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/caffe/resnet50_128_caffe.tar.gz), [MatConvNet](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/matconvnet/resnet50_128_mat.tar.gz), [TF], [PyTorch](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/pytorch/resnet50_128_pytorch.tar.gz) |
| SE-ResNet-50-256D|  256    |    Y   | 0.912 | 0.965 | [Caffe](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/caffe/senet50_256_caffe.tar.gz), [MatConvNet](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/matconvnet/senet50_256_mat.tar.gz), [TF], [PyTorch](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/pytorch/senet50_256_pytorch.tar.gz)|
| SE-ResNet-50-128D|  128    |    Y   | 0.910 | 0.959 | [Caffe](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/caffe/senet50_128_caffe.tar.gz), [MatConvNet](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/matconvnet/senet50_128_mat.tar.gz), [TF], [PyTorch](http://www.robots.ox.ac.uk/~vgg/data/vgg_face2/models/pytorch/senet50_128_pytorch.tar.gz) |