https://github.com/happynear/normface
NormFace: L2 HyperSphere Embedding for Face Verification, 99.21% on LFW
https://github.com/happynear/normface
acm-mm caffe face-verification mirror-face normface
Last synced: 4 months ago
JSON representation
NormFace: L2 HyperSphere Embedding for Face Verification, 99.21% on LFW
- Host: GitHub
- URL: https://github.com/happynear/normface
- Owner: happynear
- License: mit
- Created: 2017-03-29T23:28:07.000Z (about 9 years ago)
- Default Branch: master
- Last Pushed: 2017-12-13T02:15:01.000Z (over 8 years ago)
- Last Synced: 2025-10-23T18:48:12.832Z (8 months ago)
- Topics: acm-mm, caffe, face-verification, mirror-face, normface
- Language: Matlab
- Homepage:
- Size: 2.02 MB
- Stars: 369
- Watchers: 29
- Forks: 106
- Open Issues: 5
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# NormFace
NormFace: L2 HyperSphere Embedding for Face Verification
http://www.cs.jhu.edu/~alanlab/Pubs17/wang2017normface.pdf
| Baseline Model | Original Accuracy | Finetune Using Normalization |
| ------------------- |:-----------------:|:----------------------------:|
| [Light CNN(MaxOut)](https://github.com/AlfredXiangWu/face_verification_experiment) | 98.41%\* |98.78% |
| [Center Face(ResNet)](https://github.com/ydwen/caffe-face) | 99.03% |99.21% |
\* It is 98.13% on [Light CNN's project page](https://github.com/AlfredXiangWu/face_verification_experiment). After applying [the mirror face trick](./MirrorFace.md), it becomes 98.41%.
# Requirements
1. My Caffe (https://github.com/happynear/caffe-windows/tree/ms). It also works on Linux. Anyway, if you want to use your own Caffe, please transplant the `inner_product_layer`, `inner_distance_layer`, `normalize_layer`, `general_contrastive_layer`, `flip_layer` and `accuracy_layer` to your Caffe. Since there are too many layers to transplant, I suggest you to use my Caffe directly:)
2. Matlab to draw some figures.
3. GPU with CUDA support.
4. MTCNN face and facial landmark detector(https://github.com/kpzhang93/MTCNN_face_detection_alignment).
5. Baseline model such as [Center Face](https://github.com/ydwen/caffe-face) or [Light CNN](https://github.com/AlfredXiangWu/face_verification_experiment) or your own model trained by softmax loss. I don't know whether my loss helps or not if your model is trained by other loss functions. You may have a try. I believe it will be effective if there is no normalization term in your model.
# Train
1. The dataset used in this paper is [CASIA-Webface](http://www.cbsr.ia.ac.cn/english/CASIA-WebFace-Database.html). Note that there are 3 identities overlap between CASIA-Webface and LFW. They are `0166921`, `1056413` and `1193098`. For fair evaluation, it is recommended to remove them from CAISA-Webface.
Update: Actually there are 17 overlap identities, the list is in [my FaceDatasets repository](https://github.com/happynear/FaceDatasets/blob/master/CASIA/webface_lfw_overlap_detail.txt).
2. Align all face images using MTCNN. The script can be found in [my FaceVerification repository](https://github.com/happynear/FaceVerification/blob/master/dataset/general_align.m).
3. Replace the final inner-product layer and softmax layer with layers defined in [scaled_cosine_softmax.prototxt](./prototxt/scaled_cosine_softmax.prototxt) or [normalized_Euclidean_contrastive.prototxt](./prototxt/normalized_Euclidean_contrastive.prototxt).
4. **Fine-tune** the network based on the original model using a small learning rate, say 0.001 or 0.0001.
# Evaluation
Evaluation codes are in [my another github repository](https://github.com/happynear/FaceVerification). Please refer to the second paragraph of the Update section.
A trick called mirror face is used during extracting the features. A sample code is in `./prototxt/example_of_mirror_face.prototxt`.
**About the mirror face:** See [./MirrorFace.md](./MirrorFace.md)
**About the histogram feature for video face verification:** Details are in [./YTF](https://github.com/happynear/NormFace/tree/master/YTF).
# Trained Models
Light CNN B model(98.78%): [Google Drive](https://drive.google.com/open?id=0B0OhXbSTAU1HT3I5V3ZLd0JDaW8) or [Baidu Yun](https://pan.baidu.com/s/1gfklrrl).
ps: I used to get a model of 98.8%, but I lost it....
Center Face (ResNet-27) model(99.21%): [Google Drive](https://drive.google.com/open?id=0B0OhXbSTAU1HM2NWcWFiN2lvbTg) or [Baidu Yun](https://pan.baidu.com/s/1i4Q4vD7).
# License
This code is distributed under MIT LICENSE. The released models are only allowed for non-commercial use.
# Citation
If you find my paper or the codes useful in your research, please cite
```
@inproceedings{wang2017normface,
title={NormFace: L2 Hypersphere Embedding for Face Verification},
author={Wang, Feng and Xiang, Xiang and Cheng, Jian and Yuille, Alan L.},
booktitle={Proceedings of the 25th ACM international conference on Multimedia},
year={2017},
organization={ACM},
doi = {https://doi.org/10.1145/3123266.3123359}
}
```
# Contact
Feng Wang [feng.wff(at)gmail.com], please replace (at) with @.