Learning to Generalize Unseen Domains via Memory-based Multi-Source Meta-Learning for Person Re-ID

Methods an overview

Learning generates pedestrian re recognition in unknown domain through memory based multi-source meta learning .
1, This paper proposes a multi-source domain meta learning framework , Can mimic domain generation （DG） Training for - Testing process . This method enhances the domain independent representation of model learning and the generalization ability .
2, The article is equipped with memory-based modular , The loss of identity is realized in a non parametric way , It can prevent unstable optimization caused by traditional parameter methods .
3, Put forward MetaBN To generate various meta test features , These features can be directly injected into our meta learning framework , And get further improvement .

Contents summary

Title of thesis	abbreviation	meeting / Periodical	Year of publication	baseline	backbone	Data sets
Learning to Generalize Unseen Domains via Memory-based Multi-Source Meta-Learning for Person Re-Identification.	M3L	CVPR	2021	【QAConv50】Shengcai Liao and Ling Shao. Interpretable and generaliz- able person re-identification with query-adaptive convolution and temporal lifting. In ECCV, 2020. 2, 6, 7	ResNet-50 [11] and IBN-Net50 [28]	Market-1501 [46], DukeMTMC-reID [30, 48], CUHK03 [20, 49] and MSMT17 [41]

Online links ：https://openaccess.thecvf.com/content/CVPR2021/html/Zhao_Learning_to_Generalize_Unseen_Domains_via_Memory-based_Multi-Source_Meta-Learning_for_CVPR_2021_paper.html
Source link ： https://github.com/HeliosZhao/M3L

Work Overview

1, we study the problem ofmulti- source domain generalization in ReID, which aims to learn a model that can perform well on unseen domains with only several labeled source domains.
2, we propose the Memory-based Multi-Source Meta- Learning (M3L) framework to train a generalizable model for unseen domains. Specifically, a meta-learning strat- egy is introduced to simulate the train-test process of do- main generalization for learning more generalizable mod- els.
3, we propose a memory-based iden- tification loss that is non-parametric and harmonizes with meta-learning.
4, We also present a meta batch normaliza- tion layer (MetaBN) to diversify meta-test features, further establishing the advantage of meta-learning.

Summary of results

our M3L can effectively enhance the gen- eralization ability ofthe model for unseen domains and can outperform the state-of-the-art methods on four large-scale ReID datasets

Methods,

Methods the framework

Algorithm description

Concrete realization

1, As can be seen from the frame diagram , The main method of the article is meta - learning , Only source domain information is used , Divide multiple source domains into two parts: meta training and meta testing . In the network to join the innovative MetaBN and Memory_based The design of the .
2, Finally, network optimization combines meta training and testing , Formalized as a formula 1 Shown .

3, Memory_based Mainly to achieve a parameterless goal , This paper analyzes the disadvantages of the two forms of reference , It is considered that this nonparametric form is needed in meta learning , And make use of Memory Compare similarity with features , obtain Memory_based Loss of identity , As formula 3 Shown .

4, Memory The initialization of is to take the class characteristic mean , The update is a formula 2.

5, In addition to using Memory_based In addition to the loss of identity , Still used Triplet loss（ The formula 4）, This is useful in many ways , The article will not go into detail .

6, About MetaBN, The design is shown in the picture 4 Shown . What I understand is to combine the samples of meta test with the samples of meta training , And again Memory Find the similar loss . Sample a feature for each meta training domain, such as formula 5 Shown , Then it is mixed with the features of meta test samples to get the new features after mixing （ The formula 6 ）Z yes zi The combination of , Finally, the mixed features will be normalize（ The formula 7）.

7, The loss function of a single field in meta training is shown in the formula 8 Shown , Global functions such as formulas 9 Shown .

8, The loss function of meta test is as follows 10 Shown .

experimental result

Overall evaluation

1, Poor performance , But in CVPR, I think the paper is well written , from contributions Look at , The results are well packed .
2, The innovation sounds magnificent , And make outstanding contributions .
3, From the writing of the article , The thesis work is quite solid , There is also a lot of information . After reading, understand why the performance is poor , Because it has no target domain data , This is equivalent to a open-set The problem of , The work is quite valuable .

Small sample learning and intelligent frontier （ below ↓ official account ） The background to reply “M3L", The paper electronic resources can be obtained .

Citation format

@inproceedings{DBLP:conf/cvpr/ZhaoZYLLLS21,
author = {Yuyang Zhao and
Zhun Zhong and
Fengxiang Yang and
Zhiming Luo and
Yaojin Lin and
Shaozi Li and
Nicu Sebe},
title = {Learning to Generalize Unseen Domains via Memory-based Multi-Source
Meta-Learning for Person Re-Identification},
booktitle = { {CVPR}},
pages = {6277–6286},
publisher = {Computer Vision Foundation / {IEEE}},
year = {2021}
}