AAAI|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification

时间:2024-04-03 17:51:48

1 Abstract & Introduction

  • A novel domain adaptive attention model:(两个部分)
    Domain-shared feature map(DSH)具有迁移性,用于辅助目标域上的重识别
    Domain-specific feature map(DSP)用于消除由于域间差异引起的负迁移

  • A novel problem formulation of the domain adaptation task as one-class classification task is proposed with a domain similarity loss.
    利用域间相似度损失,将域自适应问题看成一个单分类任务。

  • A novel unsupervised person Re-ID loss is proposed to the unlabeled target domain based on a clustering process and the weighted cross-entropy loss.
    基于聚类过程和加权交叉熵损失,在无标签目标域上提出一种新的无监督行人重识别损失。

2 Related Works

3 Methodology

整个网络包括四个部分:a backbone networka domain adaptive attention modulea domain-shared brancha domain-specific branch
backbone network选用ResNet-50。
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification

3.1 Domain Adaptive Attention Module

任意图像x的特征映射Fx都可以被分成domain-shared discriminative feature map FxshF^{sh}_x和domain-specific feature map FxspF^{sp}_x,因此我们要学习一个domain-shared attention map A(Fx)(0,1)hwcA(F_x)\isin{(0,1)^{h*w*c}}:
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
根据文献1,A(Fx)A(F_x)可以分解成空间注意力S(Fx)(0,1)hw1S(F_x)\isin{(0,1)^{h*w*1}}和通道注意力C(Fx)(0,1)1cC(F_x)\isin{(0,1)^{1*c}}:
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
Spatial Attention Module
空间注意力模块的输入是特征映射FxF_x,输出是空间注意力S(Fx)S(F_x),通过挖掘特征中所有通道共有的像之间的关系来得到。因此先对所有通道进行全局平均池化过程,再经过一个卷积层(3×33\times3,stride 22)和上采样层,最后为了有效和通道信息结合,添加了一个卷积层(1×11\times1)来自动学习一个adaptive scale。
Channel Attention Module
通道注意力模块的输入是特征映射FxF_x,输出是空间注意力C(Fx)C(F_x),通过挖掘特征中通道之间的关系来获得。
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification

3.2 Domain-Shared Branch

给定任意图像的FxshF^{sh}_x,DSH的作用就是提取一个能用于目标域上行人重识别的特征。通过Global Average Pooling,Batch Normalized和ReLU active之后,将FxshF^{sh}_x投影成一个256维的特征向量fxshf^{sh}_x。为了使fxshf^{sh}_x既具有在不同域上的可迁移性,又对不同行人具有判别性,引入两种损失:domain similarity lossperson re-ID loss
Domain Similarity Loss
一个自然的思路是使得fxshf^{sh}_x的分布在源域和目标域上相似。将源域和目标域中的图像投影到一个统一的新的nominal domain中,可以看成一个one-class classification(OCC)问题。
一个256×1256\times1的全连接层和**层,用来预测图像xx属于新domain的概率po(x)p^o(x),domain similarity loss:
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification

  • OCC vs. GRL: GRL试图模糊域之间的差异,而OCC试图拉近两种域的分布。
  • OCC vs. MDD: MDD要求源域和值域具有相同的类别标签,但re-id中源域和目标域上具有完全不同的行人标签。
    Person Re-ID Loss
    对于带标签的源域来说,和有监督学习一样采用交叉熵损失:
    AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
    对于无标签的目标域来说,用在源域上训练好的模型对其提取特征后,用k平均聚类的方法为其打上弱标签:
    AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
    采用加权交叉熵损失:
    AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
    由于标签是估计的,不准确,越靠近聚类中心的样本标签具有更大的置信度,权重可由下式决定:
    AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification

3.3 Domain-Specific Branch

该分支的主要目的是为了保证最后输出的特征fxspf^{sp}_x对不同的域具有判别性,因此引入一个256×2256\times2的FC层和softmax function,用于预测任意图像xx属于源域或是目标域的概率ps(x)p^s(x)pt(x)p^t(x)
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
最后,为了保证fxshf^{sh}_xfxspf^{sp}_x完全互斥且独立,使其具有软正交限制:
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification

4 Learning

整个网络的总损失:
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
算法流程:
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification

5 Experiments

和现有无监督域适应方法比较
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
Ablation Studies

(1) Effectives of network modules
验证网络各个模块的有效性
AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
(2) Effectiveness of the proposed losses

  • 验证用于拉近域间分布的损失LDSL^{DS}的有效性
    与两种情况比较:一种是直接去掉LDSL^{DS},一种是换成GRL方法。

  • 验证目标域上加权交叉熵损失的有效性
    将权重都设成1时,准确率大大降低。
    AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
    (3) Influences of hyper-parameters

  • 迭代次数
    均在五次以后达到收敛AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification

  • 聚类类别数K
    在两个数据集上测试K=650时效果最好AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
    VisualisationAAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification
    domain-shared feature map更注重行人身体部位,而domain-specific feature map更注重背景和其他因素。AAAI2020|Domain Adaptative Attention Learning for Unsupervised Cross-Domain Person Re-Identification


  1. W. Li, X. Zhu, and S. Gong. Harmonious attention network for person re-identification. In CVPR, pages 2285–2294, 2018. 3 ↩︎