Heterogeneous Face Attribute Estimation: A Deep Multi-Task Learning Approach 

2017.11.28

Introduction:

  人脸属性的识别在社会交互,提供了非常广泛的信息,包括:the person’s identity, demographic (age, gender, and race), hair style, clothing, etc. 基于人脸属性识别的场景也越来越多,如:(i)video Surveillance;

(ii)face retrieval;(iii)social media。尽管最近在属性识别上取得了很大的进展,但是,大部分 prior works 限制在预测单个属性(如:age),或者 针对每一个属性学习一个 model,进行识别。为了解决上述的局限性,已经有很多工作在尝试 joint 的预测多个属性【见文章引用 19-23】。但是这些方法都有或多或少的不足:

  1. The approaches in [19], [20], [22] used the same features for estimating all the attributes without considering the attribute heterogeneity.

  2. The sumproduct network (SPN) adopted in [21] for modeling attribute correlations may not be feasible because of the exponentially growing number of attribute group combinations.

  3. The cascade network in [23] also required learning a separate Support Vector Machine (SVM) classifier for each face attribute, and is not an end-to-end learning approach.

  

  图一展示了人脸属性的相关性以及多样性。属性之间关系要么是 pos 要么是 neg。与此同时,单个属性可以是多样的(根据 data type 或者 scale,以及 semantic meaning)。这种属性相关性以及多样性应该被编码到 属性预测模型中去(Such attribute correlation and heterogeneity should be considered in designing face attribute estimation models.)。

  

Proposed Algorithm:

  本文提出一种 Deep Multi-Task Learning (DMTL) approach 来 Jointly 的预测单张图像中的多个属性。所提出的方法,是受到现有方法的启发,但是在一个网络中,考虑到 attribute correlation 以及 attribute heterogeneity。所提出的 DMTL 有前期的共享特征提取阶段,以及 特定类型的特征学习来进行多个属性的预测。共享的特征学习自然地探索了多个 task 之间的相关性,可以更加鲁棒以及有效的进行特征的表达。

Main Contributions:

  (i) an efficient multi-task learning (MTL) method for joint estimation of a large number of face attributes;

  (ii) modeling both attribute correlation and attribute heterogeneity in a single network;

  (iii) studying the generalization ability of the proposed approach under cross-database testing scenarios;

  (iii) compiling the LFW+ database2 with face images in the wild (LFW), and heterogeneous demographic attributes (age, gender, and race) via crowdsourcing.


Proposed Approach

  1. Deep Multi-task Learning : 

  本文的目标是,用一个联合的预测模型,同时预测多个人脸属性。当大量 face attributes 给特征学习效率上带来挑战的同时,他们也提供了结合属性内部关系的机会(leveraging the attribute inter-correlations to obtain informative and robust feature representation)。例如,CelebA dataset 中的各个属性之间就有很强的 correlation,如下图所示:

  

  那么,采用 多任务的框架来学习这个东西,就变的特别直觉了。但是,外观变换的出现 以及 the heterogeneity of individual attributes, 从 face image space 到 attribute space 的映射,通常是 nonlinear。所以, the joint attribute estimation model 应该可以捕获到复杂和综合的非线性变换。CNN model 是一种有效的处理 MTL 以及 nonlinear transformation learning 的方法。所以,我们选择基于 CNN 的 多任务框架来完成该任务:

  一个传统的 DMTL model 进行联合的属性预测可以 formulated by minimizing the regularization error function

  

  上述 model 就是:重构 loss + 正则化项的标准做法。但是这种方法不是最优的,因为属性之间的关系并没有考虑到,而属性的预测应该共享某些 feature。这也是被其他 paper 所支持的【34】。但是,公式 1 当中的表达方式,并没有显示的强调了 a large portion of feature sharing during MTL。我们将上述表达式改为下面的形式:

  

  其中,Wc 控制了人脸属性共享的 feature,Wj 控制了共享 feature 的更新。Specifically, as shown in Fig. 2, a face image is first projected to a high-level representation through a shared deep network (Wc) consisting of a cascade of complex non-linear mappings, and then refined by shallow subnetworks ({Wj}M j=1) towards individual attribute estimation tasks。


Heterogeneous Face Attributes Estimation

  尽管上述 DMTL 在特征学习过程中用到了 attribute correlationsthe attribute heterogeneity 仍然需要考虑。单个 face Attribute 的异质性曾经被提出过,但没有受到足够多的关注。原因是如下两个方面:

  1. many of the public-domain face databases are labeled with a single attribute, the requirement of designing corresponding models becomes no longer urgent ;

  2. many of the published methods choose to learn a separate model for each face attribute; model learning for individual attributes does not face the attribute heterogeneity problem.

  

  我们分别对待每一个 异质的属性类别(the heterogeneous attribute categories),但是每一个类别的 attributes 都希望能够共享  feature learning 以及 classification model。为了完成这个,我们重写了目标函数:

  

  其中,G 是异质属性类别的个数。

  

  将大量属性进行几个 heterogeneous categories 的划分,依赖于 prior knowledge。此处,我们从 data type and scale (i.e. ordinal vs. nominal)  以及 semantic meaning (i.e. holistic vs. local) 考虑 face attribute heterogeneities,然后解释我们的 特定类别的建模,来进行这些 heterogeneous attribute categories。

  

  Nominal vs. ordinal attributes .

    

论文笔记之:Heterogeneous Face Attribute Estimation: A Deep Multi-Task Learning Approach的更多相关文章

  1. 论文解读(SUBLIME)《Towards Unsupervised Deep Graph Structure Learning》

    论文信息 论文标题:Towards Unsupervised Deep Graph Structure Learning论文作者:Yixin Liu, Yu Zheng, Daokun Zhang, ...

  2. 论文笔记:Heterogeneous Memory Enhanced Multimodal Attention Model for Video Question Answering

    Heterogeneous Memory Enhanced Multimodal Attention Model for Video Question Answering 2019-04-25 21: ...

  3. 论文笔记之:Optical Flow Estimation using a Spatial Pyramid Network

    Optical Flow Estimation using a Spatial Pyramid Network   spynet  本文将经典的 spatial-pyramid formulation ...

  4. 论文笔记之:Dueling Network Architectures for Deep Reinforcement Learning

    Dueling Network Architectures for Deep Reinforcement Learning ICML 2016 Best Paper 摘要:本文的贡献点主要是在 DQN ...

  5. 论文笔记之:Pedestrian Detection aided by Deep Learning Semantic Tasks

    Pedestrian Detection aided by Deep Learning Semantic Tasks CVPR 2015 本文考虑将语义任务(即:行人属性和场景属性)和行人检测相结合, ...

  6. 论文笔记(5):Fully Convolutional Multi-Class Multiple Instance Learning

    这篇论文主要介绍了如何使用图片级标注对像素级分割任务进行训练.想法很简单却达到了比较好的效果.文中所提到的loss比较有启发性. 大体思路: 首先同FCN一样,这个网络只有8层(5层VGG,3层全卷积 ...

  7. 论文笔记:(TOG2019)DGCNN : Dynamic Graph CNN for Learning on Point Clouds

    目录 摘要 一.引言 二.相关工作 三.我们的方法 3.1 边缘卷积Edge Convolution 3.2动态图更新 3.3 性质 3.4 与现有方法比较 四.评估 4.1 分类 4.2 模型复杂度 ...

  8. 论文笔记之:Active Object Localization with Deep Reinforcement Learning

    Active Object Localization with Deep Reinforcement Learning ICCV 2015 最近Deep Reinforcement Learning算 ...

  9. 论文笔记:ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware

    ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware 2019-03-19 16:13:18 Pape ...

随机推荐

  1. html5-相对定位

    *{    margin: 0px;    padding: 0px;}div{    width: 300px;    height: 300px;}#div1{    background: rg ...

  2. 【转】win中IDLE选择virtualenv的启动方法

    从dos命令行运行.(virtualenv dir)\Scripts\activate.bat脚本激活环境,然后执行: python -m idlelib.idle 摘录:https://blog.c ...

  3. Python 学习记录之----模块 paramiko

    paramiko 一.安装 pip3.5 install paramiko pip3.5 install pycrypto # pycrypto,由于 paramiko 模块内部依赖pycrypto, ...

  4. SLAM学习笔记 - ORB_SLAM2源码运行及分析

    参考资料: DBow2的理解 单目跑TUM数据集的运行和函数调用过程 跑数据集不需要ros和相机标定,进入ORB_SLAM目录,执行以下命令: ./Examples/Monocluar/mono_tu ...

  5. VIM For Windows 1

    some tips for using vim in windows. 1,download the software vim and install it, you can go to the Of ...

  6. python二叉树的深度遍历之先序遍历流程图

  7. 今天2.4寸tft触摸屏到手--刷屏驱动小结

    2010-04-29 21:28:00 根据给的51程序改成了iccavr,结果改错了2处.导致我找原因找了n久.不过也是一件好事,让我对80i更加熟悉了. 通过protues的逻辑分析仪,找到了问题 ...

  8. Symfony2学习笔记之事件分配器

    ----EventDispatcher组件使用 简介:       面向对象编程已经在确保代码的可扩展性方面走过了很长一段路.它是通过创建一些责任明确的类,让它们之间变得更加灵活,开发者可以通过继承这 ...

  9. .NET创建一个即是可执行程序又是Windows服务的程序

    不得不说,.NET中安装服务很麻烦,即要创建Service,又要创建ServiceInstall,最后还要弄一堆命令来安装和卸载. 今天给大家提供一种方式,直接使用我们的程序来安装/卸载服务,并且可以 ...

  10. Top 5 Reasons to Get BMW ICOM A2 with Latest Software

    Top 5 Reasons to Get BMW ICOM A2 with Latest Software 1.BMW ICOM A2 Hardware Version: V2018.03 2.Sup ...