類似 ReLU 功用

It is from the author of mxnet, Bing Xu, Naiyan Wang et al.

http://arxiv.org/abs/1505.00853

reference: - happynear/FaceVerification

@happynear
I am not very clearly about fine-tuning the pre-trained model with DeepID2, you mean that i can use any one pre-trained model as the initialization of DeepID2?

Ans: Sure, DeepID1 model and DeepID2 model can share the same trained parameters.

In train_val.prototxt:

layer {
  name: "accuracy_top_5"
  type: "Accuracy"
  bottom: "fc8"
  bottom: "label"
  top: "accuracy_top_5"
  accuracy_param {
    top_k: 5
  }
  include {
    phase: TEST
  }
} 

reference: delphifirst/FaceX

A high performance (real-time) open source face landmarks detector (face alignment), based on explicit shape regression algorithm.

forked by joyhuang-machine-learning/FaceX

reference: happynear/FaceVerification

@happynear
hi, as in webface datasets is very unbalanced, it has 10575 persons, but the amount of same person is small, have you made some stategies to make the generated pair images balanced such as the probability of been the same person or the different person is almost equal?

Ans: Face++ has done a experiment. The conclusion is to delete all identity with less than 10 images.
Paper: http://arxiv.org/pdf/1501.04690.pdf .

Hello, I see your test experiment, but I find your test accuracy on LFW is low comparing the paper test. Why?

reference: https://zh.wikipedia.org/wiki/%E7%BB%B4%E5%A5%A5%E6%8B%89-%E7%90%BC%E6%96%AF%E7%9B%AE%E6%A0%87%E6%A3%80%E6%B5%8B%E6%A1%86%E6%9E%B6

Viola–Jones object detection framework => 這個方法在OpenCV中被實現為cvHaarDetectObjects()。

author: AlfredXiangWu

reference: https://github.com/AlfredXiangWu/face_verification_experiment

related: #5

• caffe framework
• CASIA-WebFace dataset: contains 493456 face images of 10575 identities.
• feature extraction code: https://github.com/AlfredXiangWu/python_misc/blob/master/caffe/caffe_ftr.py
• accuracy: (The single convolution net testing is evaluated on unsupervised setting only computing cosine similarity for lfw pairs. )
  • Dataset A: 97.77%
  • Dataset B: 98.13%

reference: happynear/FaceVerification#14

happynear:

Recently I am working on face alignment. As a friend of mine said, he successfully trained a DeepID2 model, by setting the loss_weight of contrastive loss to a very low value, say 1e-5. You may try it again.

happynear:

• His net is the same with mine, and is described in the CASIA-webface paper.
• I use matlab script to generate the list of image pairs. They are in ./dataset folder of this repo.

reference: happynear/FaceVerification

how much is the best threshold for face authentication.

Three probably answers:

1. You can use a linear svm to calculate the threshold.
2. The threshold is determined by 10-fold validation. Using 5400 pair samples for training and 600 pair test, then recycled. L2 and cosine are used to determine the threshold.
3. I used all training data(944612 images) to train a joint bayesian classifier, then using all 6000 pair samples for testing.

reference: happynear/FaceVerification

OK, I can tell you something you should be attention to in order to use Xiang Wu's model.

1. Wu's alignment method is different from mine. He used the five-point landmark detection algorithm provided by CUHK, and he aligned the face image both horizontally and vertically, while in my implementation, only vertical alginment is applied.

2. Wu's face image size is 128x128, so the aligned image should also be resized to 128x128.

3. Wu's network was trained without mean file, so there is no need to find it.

4. The distance between two facial images can be calculated by caffe directly, these configurations are in CASIA_demo.prototxt, including a Normalization layer and a Euclidean layer.

5. The threshold of same / different person can be got by cross-validation on LFW or your own dataset. Since it is only one number, it is all ok to use whether 1000 samples or 10000 samples.

reference: happynear/FaceVerification

剪去 mean 的 data 不是用 scale 1/255 ?

Ans:

happynear:

But i don't understand why you set 0.0078125 scale in CASIA_train_test.prototxt. Why is not 0.00390625.Should it normalized -1 to 1 not 0 to 1? But input image is 0 - 255 , if scale set 0.0078125 it get 0-2 isn't it? So why scale is 0.0078125. Do you do any image process before you train model?
Thank your answer!

Since I have substracted the mean image, divide by 128 (1/128=0.0078125) scales the pixels' graylevel to [-1, 1].
I have also minused the mean image. So every pixel's range is [-1 1].