centerloss
所属分类:模式识别(视觉/语音等)
开发工具:C++
文件大小:141KB
下载次数:0
上传日期:2019-01-29 22:05:21
上 传 者:
sh-1993
说明: IJCV“深度人脸识别中心损失综合研究”论文的源代码和模型
(Source code and models for the paper of "A Comprehensive Study on Center Loss for Deep Face Recognition" in IJCV)
文件列表:
caffe (0, 2019-01-30)
caffe\caffe.proto (1104, 2019-01-30)
caffe\center_loss_layer.cpp (6286, 2019-01-30)
caffe\center_loss_layer.cu (6361, 2019-01-30)
caffe\center_loss_layer.hpp (1593, 2019-01-30)
caffe\shared_center_loss_layer.cpp (8613, 2019-01-30)
caffe\shared_center_loss_layer.cu (9507, 2019-01-30)
caffe\shared_center_loss_layer.hpp (1657, 2019-01-30)
deploy_prototxt (0, 2019-01-30)
deploy_prototxt\resnet10.prototxt (5584, 2019-01-30)
deploy_prototxt\resnet20.prototxt (11068, 2019-01-30)
deploy_prototxt\resnet36.prototxt (20047, 2019-01-30)
deploy_prototxt\resnet4.prototxt (2290, 2019-01-30)
deploy_prototxt\resnet64.prototxt (34481, 2019-01-30)
exp4_2 (0, 2019-01-30)
exp4_2\Depth (0, 2019-01-30)
exp4_2\Depth\ACL (0, 2019-01-30)
exp4_2\Depth\ACL\prototxt (0, 2019-01-30)
exp4_2\Depth\ACL\prototxt\resnet10.prototxt (7050, 2019-01-30)
exp4_2\Depth\ACL\prototxt\resnet20.prototxt (12535, 2019-01-30)
exp4_2\Depth\ACL\prototxt\resnet36.prototxt (21547, 2019-01-30)
exp4_2\Depth\ACL\prototxt\resnet4.prototxt (3757, 2019-01-30)
exp4_2\Depth\ACL\prototxt\resnet64.prototxt (35981, 2019-01-30)
exp4_2\Depth\ACL\prototxt\solver1.prototxt (312, 2019-01-30)
exp4_2\Depth\ACL_gamma (0, 2019-01-30)
exp4_2\Depth\ACL_gamma\prototxt (0, 2019-01-30)
exp4_2\Depth\ACL_gamma\prototxt\resnet10.prototxt (7049, 2019-01-30)
exp4_2\Depth\ACL_gamma\prototxt\resnet20.prototxt (12534, 2019-01-30)
exp4_2\Depth\ACL_gamma\prototxt\resnet36.prototxt (21546, 2019-01-30)
exp4_2\Depth\ACL_gamma\prototxt\resnet4.prototxt (3756, 2019-01-30)
exp4_2\Depth\ACL_gamma\prototxt\resnet64.prototxt (35980, 2019-01-30)
exp4_2\Depth\ACL_gamma\prototxt\solver.prototxt (312, 2019-01-30)
exp4_2\Depth\CL (0, 2019-01-30)
exp4_2\Depth\CL\prototxt (0, 2019-01-30)
exp4_2\Depth\CL\prototxt\resnet10.prototxt (6854, 2019-01-30)
exp4_2\Depth\CL\prototxt\resnet20.prototxt (12339, 2019-01-30)
exp4_2\Depth\CL\prototxt\resnet36.prototxt (21353, 2019-01-30)
exp4_2\Depth\CL\prototxt\resnet4.prototxt (3561, 2019-01-30)
... ...
This branch is the implementation of the paper "A Comprehensive Study for Center Loss".
* [Files](https://github.com/ydwen/centerloss/blob/master/#files)
* [Trained_Model](https://github.com/ydwen/centerloss/blob/master/#trained_model)
* [Contact](https://github.com/ydwen/centerloss/blob/master/#contact)
### Implement_Details
The overall pipeline is the same as center loss (https://github.com/ydwen/caffe-face).
In this paper, we use [CAISA-WebFace](https://github.com/ydwen/centerloss/blob/master/http://www.cbsr.ia.ac.cn/english/CASIA-WebFace-Database.html), [VGG-Face2](https://github.com/ydwen/centerloss/blob/master/http://www.robots.ox.ac.uk/~vgg/data/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
* resnet***.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
* [link](https://github.com/ydwen/centerloss/blob/master/https://drive.google.com/open?id=1RGchdWY-Yjz4kqB2kqj15jseK90NB3Rn)
### Trained_Model
- exp4_2
* [link](https://github.com/ydwen/centerloss/blob/master/https://drive.google.com/open?id=1w-Tx-N8jDEXsOi_akPPTN-jZcRC-21FP)
- exp4_3
* [link](https://github.com/ydwen/centerloss/blob/master/https://drive.google.com/open?id=1WMPbY_dwqs1jeyVu6wy1OqdykfTgkCWw)
- exp4_4 & exp4_5 & exp4_6
* [link](https://github.com/ydwen/centerloss/blob/master/https://drive.google.com/open?id=1yZdA-CGVgb07brunz8reJVE2jRaKzhIL)
### Contact
- [Yandong Wen](https://github.com/ydwen/centerloss/blob/master/http://ydwen.github.io/)
- [Kaipeng Zhang](https://github.com/ydwen/centerloss/blob/master/http://kpzhang93.github.io/)
### License
Copyright (c) Yandong Wen, Kaipeng Zhang
All rights reserved.
MIT License
近期下载者:
相关文件:
收藏者: