ID retrieval is a quantitative criterion to measure the performance of face stylization algorithms. Specifically, it leverages a pre-trained face recognition model to measure the similarity of face identity between stylized images and content images.
How to calculate ID retrieval: We select the first 100 images of CelebA-HQ as content images, which are not seen by the face toonify model during training. We randomly synthesize 50 stylized images for each content image, so there are 5000 stylized images in total. We use a pre-trained face recognition network to extract face identity vectors for content and stylized images. For each stylized image, we search for its nearest face in the content images and check if the nearest face matches the original content face. The distance adopts the Euclidean distance between the face identity vectors. ID retrieval is the accuracy rate calculated by the proportion of successfully matched images to all stylized images.
Please refer to the datasets folder.
tree datasets
datasets/
├── content
│ ├── id1.jpg
│ ├── id2.jpg
│ └── xxx.jpg
└── transfer
│ ├── id1
│ │ ├── xxx.jpg
│ │ ├── xxx.jpg
│ ├── id2
│ │ ├── xxx.jpg
│ │ ├── xxx.jpg
│ ├── xxx
│ │ ├── xxx.jpg
│ │ ├── xxx.jpg
└── style (optional, for fid)
├── xxx.jpg
└── xxx.jpg
Download face recognition model model_ir_se50.pth.
Put model_ir_se50.pth in cache_pretrained/pretrained.
tree cache_pretrained
cache_pretrained/
└── pretrained
└── model_ir_se50.pth
pip install -r requirements.txt
python ID_retrieval/scripts/eval.py
Results:
ID_retrieval (top1) : 100.00%
ID_retrieval (thresh 1.5) : 68.75%
FID : 229.58
ID_retrieval (top1) is the final result. Please refer to ID_retrieval/scripts/eval.py for more details.
| Title | Venue | Year |
|---|---|---|
| ArcFace: Additive Angular Margin Loss for Deep Face Recognition | CVPR | 2019 |
| FaceShifter: Towards High Fidelity And Occlusion Aware Face Swapping | CVPR | 2020 |