introduction
we formulate a method for joint learning of local and global feature selection losses designed to optimise person re-id when using only generic matching metrics such as the L2 distance.即聯(lián)合學(xué)習(xí)局部和全局特征。作者認(rèn)為learning any matching distance metric is intrinsically learn- ing a global feature transformation across domains各吨，所以其實(shí)特征的度量用簡單的比如L2就可以了壹瘟，主要應(yīng)該聚焦于特征的提取和表達(dá)雷袋。
傳統(tǒng)的手工提取特征主要提取的是局部特征排截，比如把圖像切分成水平的條狀來處理居凶。而dl（deep learning）的方法主要提取的是圖像的全局特征界逛。但是作者認(rèn)為這兩種處理方式得到的特征都不是最優(yōu)的李请，兩者結(jié)合才好瞧筛，因?yàn)槿说囊曈X系統(tǒng)是同時處理這兩種特征（global (contextual) and local (saliency) information）的。仔細(xì)想想导盅，還是有那么點(diǎn)道理的较幌。
作者的網(wǎng)絡(luò)設(shè)計也是從這個角度出發(fā)，有兩個branch白翻，分別提取局部特征和全局特征乍炉，但是這個兩個branch不是獨(dú)立的，而是相互影響滤馍，共同學(xué)習(xí)的岛琼。這樣一個網(wǎng)絡(luò)的好處在于，不但可以同時提取局部和全局的特征巢株，還可以學(xué)習(xí)局部和全局的關(guān)系槐瑞，兩者相互補(bǔ)足，來解決局部錯位等reID中的典型問題阁苞。
此外困檩，作者還introduce a structured sparsity based feature selection learning mechanism for improving multi- loss joint feature learning robustness w.r.t. noise and data co- variance between local and global representations.意思大概就是這是一種基于稀疏性的正則化的手段，用來解決噪聲影響那槽。
related work
1.saliency learning based models悼沿。這些方法不考慮全局特征，主要modelling localised part im- portance. However, these existing methods consider only the patch appearance statistics within individual locations but no global feature representation learning, let alone the correla- tion and complementary information discovery between local and global features as modelled by the JLML.
2.Spatially Constrained Similarity (SCS) model和Multi-Channel Parts (MCP) network 骚灸。這兩個方法倒是同時考慮了全局特征糟趾。SCS主要聚焦于 supervised metric learning。但是SCS不考慮hand-crafted local and global features之間的關(guān)系甚牲。MCP主要用triplet ranking loss（不懂）來優(yōu)化义郑，而JLML主要用multiple classification loss，前者存在一定壞處：Critically, this one-loss model learning is likely to impose negative influ- ence on the discriminative feature learning behaviour for both branches due to potential over-low pre-branch independence and over-high inter-branch correlation. This may lead to sub- optimal joint learning of local and global feature selections in model optimisation, as suggested by our evaluation in Section4.3
3.HER model丈钙。主要用了regression loss魔慷，而JLML主要用的是classification loss。
4.DGD著恩。這篇文章我仔細(xì)看過，它用的也是classification loss。和JLML的區(qū)別在于 他是one-loss classification 而JLML是 multi-loss classifi- cation
模型設(shè)計
（Note that, the ReLU喉誊，rectification non-linearity [Krizhevsky et al., 2012] after each conv layer is omitted for brevity.）
兩個分支分別提取局部和全局特征邀摆。聯(lián)合學(xué)習(xí)體現(xiàn)在下面兩個方面：
1.low level的特征共享。有兩個好處伍茄，第一栋盹，共享特征，第二敷矫，減少參數(shù)例获，防止過擬合，尤其是在reID這個問題上曹仗，因?yàn)閞eID的數(shù)據(jù)集比較小
2.最后把兩個512維的特征向量疊加（local and global）
損失函數(shù)
這里他們的損失函數(shù)的選擇不同于大多數(shù)現(xiàn)存的deep reID方法榨汤，他們的損失函數(shù)主要用的是 cross- entropy classification loss function。顯存的deep reID方法主要用的contrastive loss怎茫，designed to exploit pairwise re-id labels de- fined by both positive and negative pairs, such as the pairwise verification收壕。代表之一是An improved deep learning architecture for person re- identification. In CVPR, 2015.
這么選擇損失函數(shù)的理由如下（不翻譯了，說的還挺有道理的）：The motivations for our JLML classification loss based learning are: (i) Significantly simplified training data batch construc- tion, e.g. random sampling with no notorious tricks required, as shown by other deep classification methods [Krizhevsky et al., 2012]. This makes our JLML model more scalable in real-world applications with very large training population sizes when available. This also eliminates the undesirable need for carefully forming pairs and/or triplets in preparing re-id training splits, as in most existing methods, due to the inherent imbalanced negative and positive pair size distribu- tions. (ii) Visual psychophysical findings suggest that rep- resentations optimised for classification tasks generalise well to novel categories [Edelman, 1998]. We consider that re- id tasks are about model generalisation to unseen test iden- tity classes given training data on independent seen identity classes. Our JLML model learning exploits this general clas- sification learning principle beyond the strict pair-wise rela- tive verification loss in existing re-id models.大意就是不要用正負(fù)樣本這種形式轨蛤，直接用正樣本蜜宪。DGD這篇文章也是用的一樣的思想。
其他
最后就是一些訓(xùn)練細(xì)節(jié)祥山，以及對模型各種方法有和沒有的比較圃验，證明這些方法是有好處的。好處最明顯的就是聯(lián)合global和local特征了：
還有就是兩個分支單獨(dú)學(xué)習(xí)比一起學(xué)習(xí)要好：
其他的比如有沒有l(wèi)ow level的shared feature和metric learning的選擇缝呕，以及selective feature learning（就是那個看不懂的正則化）澳窑，作用甚微。
