论文笔记（一）Heterogeneous Face Attribute Estimation: A Deep Multi-Task Learning Approach
1.论文出自PAMI2017，下载地址：https://arxiv.org/abs/1706.00906
2.该文主要是阐述如何能训练有效人脸属性分类模型。
Abstract
we present a Deep Multi-Task Learning (DMTL) approach to jointly estimate multiple heterogeneous attributes from a single face image. In DMTL, we tackle attribute correlation and heterogeneity with convolutional neural networks (CNNs) consisting of shared feature learning for all the attributes, and category-specific feature learning for heterogeneous attributes.
1.Introduction
Figure 1 shows that a face image portrays awide variety of attributes, which are both correlated and heterogeneous.Attribute correlation can be either positive or negative. For example, a personwith goatee and mustache is more likely to be a male, and is less likely towear lipstick. Meanwhile, individual attributes can be heterogeneous in termsof data type and scale [24], and semantic meaning [25]. While attributes likeage and hair length are ordinal, attributes like gender and race are nominal;these two categories of attributes are heterogeneous in terms of data type andscale. Similarly, while attributes such as age, gender, and race describe thecharacteristics of the whole face, attributes such as pointy nose and big lips,mainly describe the characteristics of local facial components; these twocategories of attributes are heterogeneous in terms of semantic meaning. Suchattribute correlation and heterogeneity should be considered in designing faceattribute estimation models.
1.1 Proposed Approach
We present a Deep Multi-Task Learning(DMTL) approach to jointly estimate multiple heterogeneous attributes from a singleface image. The proposed approach is motivated by recent advances in face attributeprediction, but takes into account both attribute correlation and attributeheterogeneity in a single convolutional neural network (CNN). The proposed DMTLconsists of an early-stage shared feature learning for all the attributes,followed by category-specific feature learning for heterogeneous attributecategories (see Fig. 2). The shared feature learning naturally exploits therelationship between tasks to achieve robust and discriminative feature representation.The category-specific feature learning aims at fine-tuning the shared featurestowards the optimal estimation of each heterogeneous attribute category. Giventhe effective shared feature learning and category-specific feature learning,the proposed DMTL achieves promising attribute estimation accuracy whileretaining low computational cost, making it of value in many face recognitionapplications.
The main contributions of this paperinclude:
(1) an efficient multi-task learning (MTL)method for joint estimation of a large number of face attributes;
(2) modeling both attribute correlation andattribute heterogeneity in a single network;
(3) studying the generalization ability ofthe proposed approach under cross-database testing scenarios;
(4) compiling the LFW+ database2 with faceimages in the wild(LFW), and heterogeneous demographic attributes (age, gender,and race) via crowdsourcing.
3 PROPOSED APPROACH3.1 Deep Multi-task Learning
3.2 Heterogeneous Face Attribute Estimation