2019 IET之ReID：HPILN: a feature learning framework for cross-modality person re-identification

HPILN: a feature learning framework for cross-modality person re-identification
当前的问题及概述：
提出了一种新的特征学习框架：hard pentaplet loss和identity loss network (HPILN),(HPILN)。在该框架中，对现有的单模态再识别模型进行了修正以适应交叉模态场景，并采用专门设计的hard pentaplet loss和identity loss来提高修正后的交叉模态再识别模型的准确性。

模型及loss：

从上图可以看出，图a为HT loss，目的是最小化anchor xa和positive xp之间的距离，最大化anchor xa和negative xn之间的距离，图b为HP loss，既包含了HT loss的任务外，还加入了最小化cross-modality positive xcp和最大化cross-modality negative xcn，图a为选择the hardest triplet（三元组），图b为选择the hardest pentaplet （五元组）。
所以，HP loss包含两部分，原理同HT loss：
hard global triplet (HGT) loss：

a hard cross-modality triplet (HCT) loss：

本文中总loss为：

实验：
数据集：SYSU-MM01
不同特征提取方法的训练参数：

不同特征提取方法在Market1501, CUHK03, and DukeMTMC-reID数据集的比较：

不同特征提取方法在SYSU-MM01数据集上的rank 1%的比较：

用CAM方法关注不同特征提取操作的attention：

其中：
Res-Mid：The devil is in the middle: exploiting mid-level representations for cross-domain instance matching（2017）
MGN：Learning discriminative features with multiple granularities for person re-identification（2018）
PCB：Beyond part models: person retrieval with refined part pooling (and a strong convolutional baseline) （2018）
MLFN：Multi-level factorisation net for person re-identification（2018）
BFE ：Batch feature erasing for person re- identification and beyond（2018）

不同现有的性能比较：

Cross-module ReID method：
zero-padding：RGB-infrared cross-modality person re-identification（2017）
cmGAN：Cross-modality person re-identification with generative adversarial training（2018）
bi-directional dual-constrained top-ranking (BDTR)：Visible thermal person re-identification via dual-constrained top-ranking（2018）
inter-channel pair between the visible-lightand thermal images + multi-scale Retinex (IPVT-1 + MSR)：Person re-identification between visible and thermal camera images based on deep residual CNN using single input（2019）
D2RL：Learning to reduce dual-level discrepancy for infrared-visible person re-identification（2019）
bi-directional center-constrained top-ranking (eBDTR)：‘Bi-directional center-constrained top- ranking for visible thermal person re-identification（2020）
D-hypersphere manifold embedding (HSME)：HSME: hypersphere manifold embedding for visible thermal person re-identification（2019）
one-stream and two-stream networks：RGB-infrared cross-modality person re-identification（2017）
Other method：
handcrafted features such as Histograms of Oriented Gradient (HOG)：Histograms of oriented gradients for human detection（2005）
and Local Maximal Occurrence (LOMO) ：Person re-identification by local maximal occurrence representation and metric learning（2015）
cross-domain models such as Common Discriminant Feature Extraction (CDFE) ：Inter-modality face recognition（2006）
and Camera coRrelation Aware Feature augmenTation (CRAFT)：Person re-identification by camera correlation aware feature augmentation（2018）
canonical correlation analysis (CCA) ：A new approach to cross-modal multimedia retrieval（2010）
metric learning method Local Fisher Discriminant Analysis (LFDA)：Local fisher discriminant analysis for pedestrian re-identification（2013）