A Tensorflow implementation of A Light CNN for Deep Face Representation with Noisy Labels from the paper by Xiang Wu
- Jan 9, 2018
- Add cleaned training list 10K and 70K.
- Sep 20, 2017
- Add model and evaluted code.
- Add training code.
- Sep 19, 2017
- The repository was built.
- Training data
- Download face dataset MS-Celeb-1M (Aligned).
- All face images are RGB images and resize to 122x144
- Download MS-Celeb-1M cleaned image_list 10K, 70K
- Testing data
- Add
- Download LCNN-29 model, this model's performance on LFW:98.2% (100%-EER)
- Download LFW features
The Light CNN performance on lfw 6,000 pairs.
| Model | traing data | method | Acc | 100% - EER | TPR@FAR=1% | TPR@FAR=0.1% | TPR@FAR=0 |
|---|---|---|---|---|---|---|---|
| LightCNN-29 (Wu Xiang) | 70K/- | Softmax | - | 99.40% | 99.43% | 98.67% | 95.70% |
| LightCNN-29 (Tensorflow) | 10K/- | Softmax | 98.36% | 98.2% | 97.73% | 92.26% | 60.53% |
| LightCNN-29 (Tensorflow) | 10K/- | Softmax+L2+PCA | 98.76% | 98.66% | 98.36% | 97% | 79.33% |
| LightCNN-29 (Tensorflow) | 10K/- | Softmax+L2+PCA+[b] | 98.95% | 98.8% | 98.76% | 97.16% | 83.36% |
| LightCNN-29 (Tensorflow) | 10K/- | Softmax_enforce+L2+PCA+[b] | 99.01% | 98.96% | 98.96% | 95.83% | 90.23% |
| Model | traing data | method | Acc | 100% - EER | TPR@FAR=1% | TPR@FAR=0.1% | TPR@FAR=0 |
|---|---|---|---|---|---|---|---|
| LightCNN-29 (Wu Xiang) | 70K/- | Softmax | - | 99.40% | 99.43% | 98.67% | 95.70% |
| LightCNN-29 (Tensorflow) | 70K/- | Softmax_enforce+L2+PCA | 99.18% | 98.9% | 98.86% | 97.9% | 94.03% |
| LightCNN-29 (Tensorflow) | 70K/- | Softmax_enforce+L2+PCA+[a] | 99.48% | 99.43% | 99.56% | 98.26% | 94.53% |
Some improved solutions:
- [a] It can be further improved by manaully aligned these images which are mis-algined in LFW
- [b] It can be further improved by doing mutiple-crop, e.g. 25 crops for per image
- [c] It can be further improved by ensemble different models
- [d] It can be further improved by adding metric learning method for similarity caculation