0 HHL摘要 Hetero-Homogeneous异质LearningH
对两个不变属性进行加强:1 相机不变性、2 域联结性connectedness
- 相同相机同域数据样本对:homogeneous 同质
- 通过源域和目标域负样本对,heterogeneous 异质,train pair(t-s)
1 Intro
再次明确REID,同Camera不同场景,同场景不同camera
1 UDA: 本身t和s类别同,但是数据分布不同。person reid中t和s的ID也不同,数量也不同
2 挖掘:target中相机不变性质
motivatoin: 自适应域内变化
- 同域中相机不变性,intra-domain而不是iner-domain。因为test的时候,query的probe和gallery中的图像原则上应该是不同camera拍摄的。对于这种细粒度的分类要考虑域内相机不同导致的变化
- 利用先验知识:s和t可以组合天然的负样本对。还需要理解
use these negative pairs to learn person embeddings within a triplet loss formulation.
Deng2018只提出Gan将S生成T域的图像,且保留S的label进行训练。相当于再target域内训练有标签的数据,进行query: Image-image domain adaptation with preserved self-similarity and domain-dissimilarity for person re-identification. In: CVPR (2018)
HHL:数据:source 标签,target 相机信息(因为容易得到,所以成无target监督,更严格的称弱target supervision)
underpin支持两个属性约束:camera invariance
- target相机不变性,正样本对:来自同相机图像是同一个域,是homogeneous。用Image2image生成正样本对,每个相机的图像是一风格。
- 域联通:t和s采样负样本对挖掘 source 和 target 的底层信息特征
anchor连接s和t—triplet loss
总之:HHL通过引入triplet loss应用了相机不变性、域联通
2 Related work
UDA 最开始方法
- UDA前人通过缩小T和S的分布差异divergence。通过align 均值、方差 CORAL
- Deep CORAL 伪标签。用特征的相似性贴伪标签
- [33]用k-nearest近邻造为标签
- 用真标签和假标签共同训练。Co-training
最近CyCADA GAN将S变换到T,通过pixel和feature level
Unsupervised person REID
1 手工特征ELF,LOMO,SDALF
通过手工设计鲁棒的特征,忽略数据分布的差异,难以在大的数据集上表现良好
2 Iteratively applying k-means clustering
3 k-reciprocal nearest neigh-bors [29] and iteratively learn features for unsupervised video re-ID.
4 Wu et al.[40] propose a progressive sampling method to gradually predict reliable pseudo labels and update deep model for one-shot video-based re-ID.
5Peng et al. [28] propose to learn a discriminative representation for target domain based on asymmetric multi-task dictionary learning
6 a transferable model is proposed to jointly learn attribute-semantic and identity discriminative feature representation for target domain
上面通过GAN的都是减小S和T的差异,均值和方差。像素级别、特征级别。
intra domain in camera
3 方法
1 问题定义:
The goal of this paper is to leverage both labeled source training images and unlabeled target training images to learn discriminative embeddings for target testing set
HHL:利用labeled Source images和unlabeled Target但是有Camera ID的Images,通过triplet 中anchor正,负学习一个分类空间。
3.3 相机不变性(同域特征)学习 CycleGAN
3.4 域连接学习
3.5 HHL:异-同学习----相机不变性,s和t的连接域 学习
3.6 总结
- 相机不同的影响很大,不管同域还是S2T,Camera style transfer,下图证明了Camera transfer比随机剪裁和反转要好
- How to sample 采样 training images from target domain?:三种策略
- 随机mini batch 没有ID重叠的images
- cluster-based sampling 先K-means成 n t n_t nt类别根据当前模型,再采样每个ID一张,
- Supervised sampling:假设target有label,随机选择target 中的ID 图像不重复
-结果分析:效果差不多
4 实验结果
HHL
Baseline是IDE:ID loss 的embedding space
(S和T通过cycleGan转换减小分布差异)
PTGAN:Person transfer gan to bridge domain gap for person REID.
SPGAN:Image to Image domain adaptation with preserved self-similarity and domain-dissimilarity for person re-identification2018
PUL: Unsupervised person REID: Clustering and fine-tuning.
CAMEL: Crossview Asymmetric Metric Learning for unsupervised person REID
UMDL: Unsupervised cross-dataset transfer learning for person REID
TJ-AIDL: Transferable Joint Attribute-Identity deep learning for unsupervised person reid
LOMO: Person re-identification by LOcal Maximal Occurrence representation and metric learning. In: CVPR (2015)局部最大事件表征