wujiyang/Contrastive-Loss

contrastive loss for face recognition.The idea is similar to DeepID2, but only in positive pairs.

Triplet Loss for FaceRecognition

Introduction

Modified from wjgaas/DeepID2, update the source code to fit the latest verison of BVLC/caffe.

Details:

1. add pair_size parameter for image_data_layer to shuffle training examples in each pair_size. Normally, pair_size = 2.
   therefore, most of the every two training examples are of the same label. in my experiments，I only make positive pairs.

   in the prototxt:

layer {
  name: "data"
  type: "ImageData"
  top: "data"
  top: "label"
  include { 
    phase: TRAIN
  }
  transform_param {
    scale: 0.0078125
    mirror: true
  }
  image_data_param {
    source: "/media/wujiyang/data/FaceData/train.txt"
    batch_size: 256
    shuffle: true
    pair_size: 2
  }
}   

2. add normalization_layer for feature normalization

   normalization_layer.hpp normalization_layer.cpp normalization_layer.cu

3. add identity2verify_layer for blob split
   in data_layer，the positive pairs are read in a batch, what this layer do is spliting them into two blobs.

Usage

In the prototxt  

layer {
	name: "split-layer"
	type: "Identity2Verify"
	bottom: "fc5"
	bottom: "label"
	top: "feature1"
	top: "feature2"
	top: "sim"
} 

Attention:

When training models with the CASIA-WebFace dataset，trying to set margin = 0.1 or 0.2.
In my experiments, when margin = 1, the loss didn't converge at all.