2nd_Place_Solution_to_Google_Landmark_Retrieval_Competition_2020_2.pdf

2025-04-28 0 0 155.11KB 3 页 10玖币

侵权投诉

2nd Place Solution to Google Landmark Retrieval Competition 2020

Min Yang*

, Cheng Cui*

, Xuetong Xue*

, Hui Ren*

, Kai Wei*

{yangmin09, cuicheng01, xuexuetong, renhui, weikai}@baidu.com

Abstract

This paper presents the 2nd place solution to the Google

Landmark Retrieval Competition 2020. We propose a

training method of global feature model for landmark re-

trieval without post-processing, such as local feature and

spatial veriﬁcation. There are two parts in our retrieval

method in this competition. This training scheme mainly

includes training by increasing margin value of arcmargin

loss and increasing image resolution step by step. Mod-

els are trained by PaddlePaddle1framework and Pytorch2

framework, and then converted to tensorﬂow 2.2. Using

this method, we got a public score of 0.40176 and a pri-

vate score of 0.36278 and achieved 2nd place in the Google

Landmark Retrieval Competition 2020 [1].

1. Introduction

The Google Landmark Dataset(GLD) V2 is currently the

largest public image retrieval and recognition dataset [2], in-

cluding 4M training data, more than 100,000 query images

and nearly 1M index data. But This dataset also contains a

lot of noise data. So this year the GLD v2 clean dataset [3]

is proposed as the ofﬁcial data of the competition, including

1.5M training data and more than 80000 classes.

The most important change of the competition this year

is that participants can only submit tensorﬂow models while

the searching and ranking process are performed by orga-

nizers without post-process, local feature and spatial veriﬁ-

cation. So the main task is training models to extract dis-

tinguish global descriptors using large landmark datasets.

In our solution, GLDv1 and GLDv2-clean datasets are used

and models are trained by PaddlePaddle and Pytorch.

The paper is listed as follows, Section2 will give an

overview of our retrieval method and Section3 describes the

training, testing and converting strategies in detail. Scores

of our models in different stages are also presented.

*These authors contributed equally to this work.

1https://github.com/PaddlePaddle/PaddleClas

2https://github.com/feymanpriv/pymetric

Figure 1. Retrieval method overview

2. Retrieval method

Our retrieval method for this competition is depicted in

Figure 1. We mainly train two models for ﬁnal submission

and each model includes a backbone model for feature ex-

traction and a head model for classiﬁcation. ResNeSt2693

and Res2Net200 vd are selected as the backbone model

since their good performance on ImageNet. Head model

includes a pooling layer and two fully connected(fc) lay-

ers. The ﬁrst fc layer is often called embedding layer or

whitening layer whose output size is 512. While the output

size of second fc layer is corresponding to the class number

of training dataset. Instead of using softmax loss for train-

ing, we train these models with arcmargin loss [4]. Arcmar-

gin loss is ﬁrstly employed in face recognition, we found it

works well in retrieval tasks which can produce distinguish-

ing and compact descriptor in landmark.

The training process mainly consists of three steps.

Firstly, we train the two models with resolution 224x224 on

GLDv1 dataset which has total 1215498 images of 14950

classes, and GLDv2-clean dataset which has total 1580470

images of 81313 classes. Secondly, these two models

are further trained on GLDv2-clean dataset with resolution

448x448, the parameters of arcmargin loss may change dur-

ing the process. We believe that using large input size is

beneﬁcial to extract feature of tiny landmark. However, we

have to adopt the training strategy “from small to large”

mainly due to the large cost and lack of GPUs. In the ﬁnal

step, some tricks are experimented to increase the perfor-

mance. We have tried a lot of methods, such as triplet loss

ﬁnetuning, circle loss ﬁnetuning and etc but only “Gem-

Pool” [5] and “Fix” [6] strategy are helpful. Details are

described in section3.

3https://github.com/zhanghang1989/ResNeSt

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：

2ndPlaceSolutiontoGoogleLandmarkRetrievalCompetition2020MinYang*,ChengCui*,XuetongXue*,HuiRen*,KaiWei*{yangmin09,cuicheng01,xuexuetong,renhui,weikai}@baidu.comAbstractThispaperpresentsthe2ndplacesolutiontotheGoogleLandmarkRetrievalCompetition2020.Weproposeatrainingmethodofglobalfeaturemodelforlandma...

展开>> 收起<<

2nd_Place_Solution_to_Google_Landmark_Retrieval_Competition_2020_2.pdf.pdf

共3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！

2nd_Place_Solution_to_Google_Landmark_Retrieval_Competition_2020_2.pdf

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: