ComparatorNetwork_PyTorch

This is an unofficial implementation of VGG's Comparator Network [1] using PyTorch.
Overall architecture is below.

Dependencies

• Python 3.5+
• PyTorch 0.4.0
• python-opencv

Implemented Part

• Basic architecture of detect, attend and compare module.
• Model train code.
• Diversity Regularization of local landmarks.

Currently ongoing part

• Using pretrained ResNet, SENet model in detect process.
• Keypoint Regularization of local landmarks.
• Hard sampling

Dataset

Download VGGFace2 dataset [link]

Train

python train.py -i "path to your local VGGFace2 train img dir"

References

1. Weidi Xie, Li Shen, Andrew Zisserman, Comparator Networks, 2018 [pdf]