Generalized Inter-class Loss for Gait Recognition

2025-04-24 0 0 2.07MB 10 页 10玖币

侵权投诉

Weichen Yu∗

yuweichen16@mails.ucas.ac.cn

Center for Research on Intelligent Perception and

Computing, Institute of Automation, Chinese Academy of

Sciences

Beijing, China

Hongyuan Yu

hongyuan.yu@cripac.ia.ac.cn

Center for Research on Intelligent Perception and

Computing, Institute of Automation, Chinese Academy of

Sciences

Beijing, China

Yan Huang

yhuang@nlpr.ia.ac.cn

Center for Research on Intelligent Perception and

Computing, Institute of Automation, Chinese Academy of

Sciences

Beijing, China

Liang Wang†

wangliang@nlpr.ia.ac.cn

Center for Research on Intelligent Perception and

Computing, Institute of Automation, Chinese Academy of

Sciences

Beijing, China

ABSTRACT

Gait recognition is a unique biometric technique that can be per-

formed at a long distance non-cooperatively and has broad applica-

tions in public safety and intelligent trac systems. Previous gait

works focus more on minimizing the intra-class variance while

ignoring the signicance in constraining inter-class variance. To

this end, we propose a generalized inter-class loss which resolves

the inter-class variance from both sample-level feature distribu-

tion and class-level feature distribution. Instead of equal penalty

strength on pair scores, the proposed loss optimizes sample-level

inter-class feature distribution by dynamically adjusting the pair-

wise weight. Further, in class-level distribution, generalized inter-

class loss adds a constraint on the uniformity of inter-class feature

distribution, which forces the feature representations to approx-

imate a hypersphere and keep maximal inter-class variance. In

addition, the proposed method automatically adjusts the margin

between classes which enables the inter-class feature distribution

to be more exible. The proposed method can be generalized to

dierent gait recognition networks and achieves signicant im-

provements. We conduct a series of experiments on CASIA-B and

OUMVLP, and the experimental results show that the proposed

loss can signicantly improve the performance and achieves the

state-of-the-art performances.

CCS CONCEPTS

•Computing methodologies →Biometrics

;Supervised learning

by classication;Neural networks.

KEYWORDS

Gait recognition, metric learning

∗This work was partly done in Watrix.

†Corresponding author.

This work is licensed under a Creative Commons Attribution

International 4.0 License.

MM ’22, October 10–14, 2022, Lisboa, Portugal

ACM ISBN 978-1-4503-9203-7/22/10.

https://doi.org/10.1145/3503161.3548311

ACM Reference Format:

Weichen Yu, Hongyuan Yu, Yan Huang, and Liang Wang. 2022. Generalized

Inter-class Loss for Gait Recognition. In Proceedings of the 30th ACM Interna-

tional Conference on Multimedia (MM ’22), Oct. 10–14, 2022, Lisboa, Portugal.

ACM, New York, NY, USA, 10 pages. https://doi.org/10.1145/3503161.3548311

1 INTRODUCTION

Gait recognition is a biometric technique based on unique walk-

ing patterns of pedestrians. Compared to other biometrics such as

face, ngerprint or iris, gait can be captured at a distance without

the cooperation of subjects or intrusion to them. Therefore, it has

been applied to many applications recently, such as crime preven-

tion, forensic identication, public security, and intelligent trac

systems [8, 41, 52].

Compared to other image recognition tasks, gait recognition has

the challenge of larger intra-class variance and smaller inter-class

variance. The large intra-class variance is mainly due to the large

visual dissimilarity among silhouettes from dierent angles and the

same pedestrian wearing dierent clothes. The small inter-class

variance is mainly caused by the following reason: although dier-

ent pedestrians have dierent walking patterns, their gait silhouette

frames are very similar because pedestrians have the same body

structures (head, torso, arms and legs), similar body proportions,

and similar walking postures including raising one leg, stepping for-

ward, shifting weight from one leg to another, etc. Many eective

works addressing intra-class variance problem have been proposed,

such as cross-view gait recognition [

] and cross-cloth

gait recognition [

], and achieve promising performances.

However, most previous works overlook the importance of small

inter-class variance and only address it implicitly.

Prior gait recognition works [

] mainly focus

on resolving intra-class variance by designing network architec-

tures, while incidentally addressing the small inter-class variance

by extracting ne-grained features from the designed networks. Pre-

vious approaches in gait recogition are categorized into two types:

model-based [

] and appearance-based [

The model-based methods use three-dimensional models and con-

vey more information than two-dimensional ones, thus magnify-

ing the inter-class variance. But model-based methods are highly

arXiv:2210.06779v1 [cs.CV] 13 Oct 2022

MM ’22, October 10–14, 2022, Lisboa, Portugal Weichen Yu, Hongyuan Yu, Yan Huang, & Liang Wang

(a) Baseline (b) Ours

Figure 1: TSNE visualization on CASIA-B test set. One color

denotes a class. The red boxes denote suboptimal represen-

tations. (a) baseline method. (b) the proposed generalized

intra-class loss.

dependent on pose estimation accuracy. And appearance-based

methods including Gait Energy Image (GEI)-based, set-based, and

3DCNN-based, extract ne-grained features which enlarge inter-

class variance. But the above approaches do not explicitly constrain

inter-class feature distribution.

To address the small inter-class variance, explicitly constraining

the inter-class feature distribution is benecial. From sample-level

perspective, some samples of the same viewpoint from dierent

classes are close to each other due to visual similarity. They need to

be emphasized to increase their distances and thus the inter-class

variance increases. However, previous gait works [

] usually treat pairs from dierent classes inexibly, where the

penalty strength on pair scores is restricted to be equal. From class-

level perspective, constraining the inter-class feature distribution

to be more uniform can increase inter-class variance. Previous

works [

] seldom have constraints on inter-

class distribution, resulting in lack of spatial symmetry, which is

not optimal in keeping maximal mutual information. Nevertheless,

margin aims to constrain the distance between classes, but prior

works treat all pedestrians equally with the same given margin

[

], which lacks exibility for optimization. Also, dierent

classes with the same given margin lack ability to discriminate

between each other.

To this end, we propose a generalized inter-class loss to resolve

the inter-class variance problem from both sample-level and class-

level. From sample-level perspective, the proposed generalized inter-

class loss treats dierent pairs with dynamic and automatic coe-

cients, which enables dierent inter-class samples to dynamically

adjust their distances from the anchor class.

Further, from class-level perspective, the proposed generalized

inter-class loss adds a constraint on uniformity of inter-class feature

representation and has advantages threefold. Firstly, uniformity

prefers the inter-class feature distribution that preserves maximal

information. The proposed similarity cross entropy (SimCE) in

generalized inter-class loss can be regarded as a variation of von

Mises-Fisher kernel density estimation [

], and forces the

inter-class feature distribution to approximate a hypersphere in

high dimension space. Thus, inter-class uniformity enables maxi-

mal inter-class variance. Secondly, the proposed loss is robust with

respect to inter-class feature representation dierences in its local

area. Thirdly, to address the xed given margin between dierent

classes, the proposed generalized inter-class loss enables automat-

ically adjusting margins between dierent classes and forces a

exible inter-class feature distribution.

Fig.1 is the TSNE visualization of test features. It can be clearly

seen that the feature distribution is more uniform, those hard exem-

plars are eectively optimized and the suboptimal representations

in red boxes are decreasing.

The contributions of the proposed method are summarized as

follows:

•

We propose a unied method to resolve the inter-class vari-

ance of gait features from both sample-level and class-level,

which dynamically and automatically adjusts the penalty

strength on pair scores and margins between dierent classes.

•

We further analyze the properties of the proposed method

from three aspects, namely inter-class hard mining, unifor-

mity and robustness of inter-class feature distribution, and

dynamic margin. And we illustrate how these properties

constrain a better inter-class feature distribution.

•

The proposed gait recognition method improves the perfor-

mance regardless of model structure. Experimental results

on public datasets CASIA-B and OUMVLP achieve state-of-

the-art performances, especially with an improvement (6.2%)

in dierent cloth (CL) condition.

2 RELATED WORKS

2.1 Gait Recognition

Gait recognition [

] is to learn the unique spatio-

temporal pattern about the human gait characteristics to obtain its

identity information. The gait model input is bipartite: 3D based

methods [

–

] reconstructing the human 3D models from dier-

ent cameras views, while 2D gait data [

] is more convenient

and easier to achieve. In early gait recognition, to deal with the

large variance in gait representation of same identity, hand-crafted

view-invariant feature [

] and View Transformation Model

(VTM) [

] are proposed. Recent deep gait recognition networks

in CNN are mostly used to capture gait information. GEInet [

] and

siamese gait network [

] work on GEI input with CNN. Temporal

information capturing includes compressing the gait sequence into

one frame using order-consistent statistic operations along tempo-

ral dimensions [

]. Temporal information is also captured by

LSTM or GRU to aggregate pose features in time series to generate

the nal gait feature [68].

To further improve the spatial temporal gait representation,

Zhang et al. [

] utilizes a temporal attention mechanism and adap-

tively adjusts the weights of dierent frames. GaitNet [

] and

ICDNet [

] emphasize disentangled representation learning. GAN

is also utilized [

] to generate more data and help with fea-

ture constructing. SelfGait [

] uses self-supervised learning to

perform gait recognition. Gait in the wild attracts researchers’ at-

tention [

], which focuses on real gait conditions and provides

datasets in the wild. However, most of the works above focus more

on addressing large intra-class variance and seldom consider small

inter-class variance, which is of the same importance as well.

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：

GeneralizedInter-classLossforGaitRecognitionWeichenYu∗yuweichen16@mails.ucas.ac.cnCenterforResearchonIntelligentPerceptionandComputing,InstituteofAutomation,ChineseAcademyofSciencesBeijing,ChinaHongyuanYuhongyuan.yu@cripac.ia.ac.cnCenterforResearchonIntelligentPerceptionandComputing,InstituteofAutom...

展开>> 收起<<

Generalized Inter-class Loss for Gait Recognition.pdf

共10页,预览2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！

Generalized Inter-class Loss for Gait Recognition

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: