Giter VIP home page Giter VIP logo

maskinsightface's Introduction

insightface制作自己的数据及其训练

2019.10.21: 更新人脸检测模型,检测器是基于SSH,相较于原版检测,该版本主要更新了误检

2019.10.01: 公布人脸识别模型,模型基于glint和私有数据训练,在私有数据上拥有0.88的F1-score,insightface原始模型0.56

说明:算法集成insightface:https://github.com/deepinsight/insightface

改进地方:ssh(人脸检测)+prnet(68 landmark 人脸对齐, 3d人脸mask)+insightface

(1)修改人脸检测器(优化后的SSH,误检率更低,对大角度和blur的face进行过滤)
(2)使用68个点的landmark,prnet的对齐效果更准
(3)利用prnet拟合有效区域的人脸位置,抠出背景,以平均人脸像素填充背景,减少噪声影响,会在图片质量较好的情况下提高识别

MASK0

MASK1

0.安装

(1)mxnet
(2)tensorflow

1.生成对齐后的数据集

1.1.数据下载

http://trillionpairs.deepglint.com/data

cd make_rec

1.2.生成,'.lst, .rec, .idx, property'

(1)为了合并数据,可采用generate_lst.sh

(2)property是属性文件,里面内容是类别数和图像大小,例如

1000,112,112 其中1000代表人脸的类别数目,图片格式为112x112(一直都不知道怎么自动生成这个,我是自己写的)

(3)sh generate_lst.sh

1.3.生成测试文件.bin

python gen_valdatasets.py

1.4.生成数据

python3 gen_datasets.py  #完成后会output下生成train.lst

2.验证model精度

2.1.在bash

python3 -u ./src/eval/verification.py --gpu 0 --model "./models/glint-mobilenet/model,1" --target 'lfw'

2.2.快捷

sh verification.sh

3.训练

3.1.在bash里面训练

CUDA_VISIBLE_DEVICES='2,3,4,5' python3 -u train.py --network r100 --loss arcface --per-batch-size 64 2>&1 > log.log &

3.2.如果想要合并不同数据集

CUDA_VISIBLE_DEVICES=0 python3 src/data/dataset_merge.py --include 001_data,002_data --output ms1m+vgg --model ../../models/model,1

4.result

'参数设置' network backbone: r100 ( output=E, emb_size=512, prelu )

loss function: arcface(m=0.5)

batch-size:256, 4gpu, config.fc7_wd_mult = 10

lr = 0.004, lr_steps [105000, 125000, 150000], default.wd = 0.0005, end with 180001,

then retrain with lr = 0.0004, lr_steps[200000, 300000, 400000], default.wd = 0.00001

Data LFW CFP_FP AgeDB30
ACCU(%) 99.82+ 98.50+ 98.25+

5.预训练模型

  1. 人脸检测模型请参见 https://github.com/bleakie/mxnet-ssh-face-detection (在自有数据集上标定+修改部分训练参数,可在FDDB上取得98.7%)

  2. 人脸识别预训练模型(模型基于glint和私有数据训练,backbone resnet152,在私有数据上拥有0.88的F1-score,insightface原始模型0.56,因为进行了私有数据的增强训练,在开源测试集上效果一般)
    链接: https://drive.google.com/drive/folders/1zWadm9yu0rcjIQ_MnoXAQ27kA-CJYGms?usp=sharing 百度云链接: https://pan.baidu.com/s/1ySZeJWa-r7oS4E_8dpdo4w 提取码: enph

Todo

  1. 释放训练好的模型(PRNET,更新人脸检测模型基于Retina的RetinaDetection 链接:https://github.com/bleakie/RetinaDetector

  2. 近期会更新新的识别策略,可相较于现版本提高2%

maskinsightface's People

Contributors

bleakie avatar

Watchers

James Cloos avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.