DREAM block for Pose-Robust Face Recognition

This is our implementation for our CVPR 2018 accepted paper Pose-Robust Face Recognition via Deep Residual Equivariant Mapping paper on arxiv.

The code is wriiten by Yu Rong and Kaidi Cao

Prerequisites

ubuntu 16.04 配一块1080T, cuda版本9.1
Anaconda2 + Python 2.7
opencv3.4.1
按作者git下载模型和数据

Train DREAM Block

stitch Training

Prepare the feature extracted from any face recognition model (You could use the pretrained model we prepared).
We prepared a piece of sample data (stitching.zip) which could be download from Google Drive Baidu Yun

Download the sample data

mkdir data
mv stitching.zip data
cd data
unzip stitching.zip

Train the model:

cd src/stitching
sh train_stitch.sh

end2end Training

Download the Ms-Celeb-1M Subset

mkdir data
mv msceleb.zip data
cd data
unzip msceleb.zip

Train the model:

cd src/end2end
sh train.sh

ResNet50直接训显存不够，batchsize改成64训练loss直接nan，lr改成0.001,200个epoch训练集100%，500个epoch测试集rank1=79，训练loss约0.1,测试loss约1.0；
RESNET18，batchsize=128，lr=0.01,40个epoch训练集100%,100个epoch测试集rank1=75%，训练loss=0.004,测试loss=1.4
放出来的数据没有清洗Moira_Quirk里怎么有俩男的......
训练需要提供头部偏航角yaw,做一个非线性映射到[0,1]区间，trainlabel.txt和train_list.txt按行一一对应
找不到作者的yaw coef非线性映射函数，代码里直接将yaw写死为1训练，resnet18也能到75%

trainlabel.txt:
467 10          #label比list多一行，训练样本总数=467 id数=10
0 1.43247e-13   #id=0 yaw coef=1.43-13
0 2.41652e-05   #id=0 yaw coef=2.41-05
......
9 0.000148665   #id=9 yaw coef=0.00014

evaluate CFP

Download the CFP dataset and preprocess the image. Then download the image list for evaluation

# make sure you are in the root directory of DREAM project
mkdir data
cd src/preprocess
sh align_cfp.sh     ##这里用了test_process_align， ldd看它依赖的库，缺啥装啥
cd data/CFP
unzip CFP_protocol.zip

aligin_cfp.sh里几个变量绝对地址改为自己机器上的CFP开源数据集地址
生成的estimate_pose.txt倒数第二列是yaw
Download pretrained model

# make sure you are in the root directory of DREAM project
cd ../ 
mv model.zip data
cd data
unzip model.zip

Evaluate the pretrained model on CFP dataset

# make sure you are in the root directory of DREAM project
cd src/CFP
sh eval_cfp.sh

CUDA_VISIBLE_DEVICES改成自己的卡0 eval_roc.py第96行一样的改为自己机器上的CFP开源数据集地址

evaluate IJBA

Download the IJBA dataset(contact me to get the aligned images)

# make sure you are in the root directory of DREAM project
mkdir data
mv IJBA.zip data
cd data
unzip IJBA.zip

Download pretrained models (If have downloaded the models, skip this step)

# make sure you are in the root directory of DREAM project
cd ../ 
mv model.zip data
cd data
unzip model.zip

Evaluate the pretrained model on IJBA dataset

# make sure you are in the root directory of DREAM project
cd src/IJBA
sh eval_ijba.sh

IJBA/align_image_1N/split/gallery路径下的list里都是写的绝对路径，需要改成自己的，用processIJBA.py暴力改
eval_ijba.sh跑到最后会死掉，貌似是多线程问题，待解决

tszssong / DREAM