Centerloss
Source code and models for the paper of "A Comprehensive Study on Center Loss for Deep Face Recognition" in IJCV
Install / Use
/learn @ydwen/CenterlossREADME
This branch is the implementation of the paper "A Comprehensive Study for Center Loss".
Implement_Details
The overall pipeline is the same as center loss (https://github.com/ydwen/caffe-face).
In this paper, we use CAISA-WebFace, VGG-Face2 For the details of removing overlapping ID, please refer to https://github.com/happynear/FaceDatasets
Preprocessing We use MTCNN (https://github.com/kpzhang93/MTCNN_face_detection_alignment) to detect five facial keypoints and use them to align the faces. For the alignment details, please see https://github.com/ydwen/caffe-face/blob/caffe-face/face_example/extractDeepFeature.m
Training The training details can be found on according prototxt. Note that the batch size we used is 512.
Loss layer
Center Loss & Generalized Center Loss
layer {
name: "generalized_center_loss"
type: "CenterLoss"
bottom: "fc5"
bottom: "label"
top: "center_loss"
top: "count"
param {
lr_mult: 1
decay_mult: 1
}
center_loss_param {
num_output: 7994
margin:5 ##radius##
center_filler {
type: "xavier"
}
}
loss_weight: 0.01
loss_weight: 0.0
}
Advanced Center Loss & Generalized Center Loss
layer {
name: "advanced_center_loss"
type: "SharedCenterLoss"
bottom: "fc5"
bottom: "label"
top: "shared_center_loss"
top: "count"
param {
name:"center"
lr_mult: 1
decay_mult: 1
}
param {
lr_mult: 1
decay_mult: 0
}
shared_center_loss_param {
num_output: 7994
margin:5 ##radius##
gamma_shared: True ##share weights##
center_filler {
type: "xavier"
}
gamma_filler{
type:"constant"
value:1
}
}
loss_weight: 0.01
loss_weight: 0.0
}
Files
- caffe
- caffe.proto
- center_loss_layer.hpp
- center_loss_layer.cpp
- center_loss_layer.cu
- shared_center_loss_layer.hpp
- shared_center_loss_layer.cpp
- shared_center_loss_layer.cu
- deploy_prototxt
- resnet4.prototxt
- resnet10.prototxt
- resnet20.prototxt
- resnet36.prototxt
- resnet64.prototxt
- exp4_2
- Parameter sharing
- Loss Weight
- Radius
- Training_set
- Depth
- exp4_3
- softmax
- softmax + contrastive
- normface
- coco
- SphereFace
- softmax + CL
- softmax + ACL
- softmax + ACL-γ
- coco + ACL-γ
- sphere + ACL-γ
- exp4_4 & exp4_5 & exp4_6
- softmax
- softmax + CL
- softmax+ ACL-γ
- softmax + CL (ρ=5)
- softmax+ ACL-γ (ρ=5)
- sphere+ ACL-γ (ρ=5)
- training_list
Trained_Model
Contact
License
Copyright (c) Yandong Wen, Kaipeng Zhang All rights reserved. MIT License
