PARALLEL AUGMENTATION AND DUAL ENHANCEMENT FOR OCCLUDED PERSON RE-IDENTIFICATION Zi Wang1 Huaibo Huang2 Aihua Zheng3 Chenglong Li3 Ran He2

2025-04-24 0 0 1.69MB 5 页 10玖币

侵权投诉

PARALLEL AUGMENTATION AND DUAL ENHANCEMENT

FOR OCCLUDED PERSON RE-IDENTIFICATION

Zi Wang1, Huaibo Huang2, Aihua Zheng3, Chenglong Li3, Ran He 2

1School of Computer Science and Technology, Anhui University 2MAIS & CRIPAC, CASIA

3IMIS Laboratory of Anhui Province, Anhui Provincial Key Laboratory of MCC, Anhui University

ABSTRACT

Occluded person re-identiﬁcation (Re-ID), the task of search-

ing for the same person’s images in occluded environments,

has attracted lots of attention in the past decades. Recent ap-

proaches concentrate on improving performance on occluded

data by data/feature augmentation or using extra models to

predict occlusions. However, they ignore the imbalance prob-

lem in this task and can not fully utilize the information from

the training data. To alleviate these two issues, we propose

a simple yet effective method with Parallel Augmentation

and Dual Enhancement (PADE), which is robust on both

occluded and non-occluded data and does not require any

auxiliary clues. First, we design a parallel augmentation

mechanism (PAM) to generate more suitable occluded data

to mitigate the negative effects of unbalanced data. Second,

we propose the global and local dual enhancement strategy

(DES) to promote the context information and details. Ex-

perimental results on three widely used occluded datasets and

two non-occluded datasets validate the effectiveness of our

method. The code is available at PADE (GitHub).

Index Terms—Person Re-identiﬁcation, Data Augmen-

tation, Feature Enhancement

1. INTRODUCTION

Occluded person Re-ID, which incorporates the data obscured

by various obstacles, has recently gained popularity. And the

occlusions are uncommon in the training set [1,2] while abun-

dant in the test set (especially in query), as illustrated in Fig. 1

(a). Training with such unbalanced data increases the chal-

lenge for the network while testing on unknown data. Efforts

in data and feature augmentation are emerging to eliminate

the imbalance between training and testing. Most methods

[3,4,5,6] employ standard data augmentation such as ran-

dom ﬂipping, random deleting, random cropping, and so on.

Furthermore, FED [7] provides feature augmentation strate-

gies to improve the network’s adaptability to occluded data.

The widely used data/feature augmentation mechanisms take

one image/feature as the input and output only one changed

Corresponding author: Aihua Zheng (ahzheng214@foxmail.com)

Fig. 1. (a) & (b): Imbalance problem. (c) & (d): Global and

local information have their advantages, respectively.

image/feature to the subsequent network for training. How-

ever, as illustrated in Fig.1(b), practically all occlusions oc-

cur in the query, and the gallery images almost have no ob-

structions in occluded Re-ID datasets [8,9]. The methods

mentioned that focus on data/feature augmentations ignore

the unbalanced occlusion between query and gallery. To in-

crease the robustness of the network on both the non-occluded

data (in the gallery) and the occluded data (in the query), we

propose a data augmentation method called the Parallel Aug-

mentation Mechanism (PAM). Our PAM consists of three in-

dependent components: Base Augmentation (BA), Erasing

Augmentation (EA), and Cropping Augmentation (CA). In

our parallel augmentation mechanism, EA only implements

the erase operation, and CA only crops the original image.

We will obtain an image triplet after the PAM, as shown in

Fig. 2(left). Then the ViT-based feature extractor takes the

image triplet as the input.

Additionally, both details and context information are cru-

cial for the Re-ID task. As illustrated in Fig. 1(c), we can

simply identify ID1 and ID2 by local details while ﬁnding it

hard to distinguish them based on their outward appearance.

[10,11,12] propose using additional clues by leveraging fore-

ground segmentation and pose estimation models. [4,13,14]

propose to split the global feature into several parts and use

ﬁner features with detailed information for training. In some

cases, the global information becomes more crucial when the

body is hindered by unknown impediments or the details are

arXiv:2210.05438v3 [cs.CV] 10 Jan 2024

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

PARALLELAUGMENTATIONANDDUALENHANCEMENTFOROCCLUDEDPERSONRE-IDENTIFICATIONZiWang1,HuaiboHuang2,AihuaZheng3,ChenglongLi3,RanHe21SchoolofComputerScienceandTechnology,AnhuiUniversity2MAIS&CRIPAC,CASIA3IMISLaboratoryofAnhuiProvince,AnhuiProvincialKeyLaboratoryofMCC,AnhuiUniversityABSTRACTOccludedpersonre-...

展开>> 收起<<

PARALLEL AUGMENTATION AND DUAL ENHANCEMENT FOR OCCLUDED PERSON RE-IDENTIFICATION Zi Wang1 Huaibo Huang2 Aihua Zheng3 Chenglong Li3 Ran He2.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

PARALLEL AUGMENTATION AND DUAL ENHANCEMENT FOR OCCLUDED PERSON RE-IDENTIFICATION Zi Wang1 Huaibo Huang2 Aihua Zheng3 Chenglong Li3 Ran He2

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: