Multi-Agent Semi-Siamese Training for Long-tail and Shallow Face Learning

With the recent development of deep convolutional neural networks and large-scale datasets, deep face recognition has made remarkable progress and been widely used in various applications. However, unlike the existing public face datasets, in many real-world scenarios of face recognition, the depth...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:ACM transactions on multimedia computing communications and applications 2023-11, Vol.19 (6), p.1-20
Hauptverfasser: Tai, Yichun, Shi, Hailin, Zeng, Dan, Du, Hang, Hu, Yibo, Zhang, Zicheng, Zhang, Zhijiang, Mei, Tao
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page 20
container_issue 6
container_start_page 1
container_title ACM transactions on multimedia computing communications and applications
container_volume 19
creator Tai, Yichun
Shi, Hailin
Zeng, Dan
Du, Hang
Hu, Yibo
Zhang, Zicheng
Zhang, Zhijiang
Mei, Tao
description With the recent development of deep convolutional neural networks and large-scale datasets, deep face recognition has made remarkable progress and been widely used in various applications. However, unlike the existing public face datasets, in many real-world scenarios of face recognition, the depth of training dataset is shallow, which means only two face images are available for each ID. With the non-uniform increase of samples, such issue is converted to a more general case, a.k.a long-tail face learning, which suffers from data imbalance and intra-class diversity dearth simultaneously. These adverse conditions damage the training and result in the decline of model performance. Based on the Semi-Siamese Training (SST), we introduce an advanced solution, named Multi-Agent Semi-Siamese Training (MASST), to address these problems. MASST includes a probe network and multiple gallery agents, the former aims to encode the probe features, and the latter constitutes a stack of networks that encode the prototypes (gallery features). For each training iteration, the gallery network, which is sequentially rotated from the stack, and the probe network form a pair of semi-siamese networks. We give the theoretical and empirical analysis that, given the long-tail (or shallow) data and training loss, MASST smooths the loss landscape and satisfies the Lipschitz continuity with the help of multiple agents and the updating gallery queue. The proposed method is out of extra-dependency, thus can be easily integrated with the existing loss functions and network architectures. It is worth noting that, although multiple gallery agents are employed for training, only the probe network is needed for inference, without increasing the inference cost. Extensive experiments and comparisons demonstrate the advantages of MASST for long-tail and shallow face learning.
doi_str_mv 10.1145/3594669
format Article
fullrecord <record><control><sourceid>acm_cross</sourceid><recordid>TN_cdi_crossref_primary_10_1145_3594669</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>3594669</sourcerecordid><originalsourceid>FETCH-LOGICAL-a239t-c9506c45283a9701b0a2f1833c0c9eb4534507e18aae6710262471cf7df714003</originalsourceid><addsrcrecordid>eNo9kE1LAzEYhIMoWKt495Sbp2jezdfmWIrVwoqHreflbZrUyH5IsiL-e1tae5qBeRiGIeQW-AOAVI9CWam1PSMTUAqYLrU6P3llLslVzp-cC62knpDl63c7Rjbb-n6kte8iqyN2Pnu6Shj72G9pGBKthn7LRowtxX5D6w9s2-GHLtB5WnlMe-6aXARss7856pS8L55W8xdWvT0v57OKYSHsyJxVXDupilKgNRzWHIsApRCOO-vXUgmpuPFQInptgBe6kAZcMJtgQO52T8n9odelIefkQ_OVYofptwHe7B9ojg_syLsDia47Qf_hHyjTU4g</addsrcrecordid><sourcetype>Aggregation Database</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>Multi-Agent Semi-Siamese Training for Long-tail and Shallow Face Learning</title><source>ACM Digital Library</source><creator>Tai, Yichun ; Shi, Hailin ; Zeng, Dan ; Du, Hang ; Hu, Yibo ; Zhang, Zicheng ; Zhang, Zhijiang ; Mei, Tao</creator><creatorcontrib>Tai, Yichun ; Shi, Hailin ; Zeng, Dan ; Du, Hang ; Hu, Yibo ; Zhang, Zicheng ; Zhang, Zhijiang ; Mei, Tao</creatorcontrib><description>With the recent development of deep convolutional neural networks and large-scale datasets, deep face recognition has made remarkable progress and been widely used in various applications. However, unlike the existing public face datasets, in many real-world scenarios of face recognition, the depth of training dataset is shallow, which means only two face images are available for each ID. With the non-uniform increase of samples, such issue is converted to a more general case, a.k.a long-tail face learning, which suffers from data imbalance and intra-class diversity dearth simultaneously. These adverse conditions damage the training and result in the decline of model performance. Based on the Semi-Siamese Training (SST), we introduce an advanced solution, named Multi-Agent Semi-Siamese Training (MASST), to address these problems. MASST includes a probe network and multiple gallery agents, the former aims to encode the probe features, and the latter constitutes a stack of networks that encode the prototypes (gallery features). For each training iteration, the gallery network, which is sequentially rotated from the stack, and the probe network form a pair of semi-siamese networks. We give the theoretical and empirical analysis that, given the long-tail (or shallow) data and training loss, MASST smooths the loss landscape and satisfies the Lipschitz continuity with the help of multiple agents and the updating gallery queue. The proposed method is out of extra-dependency, thus can be easily integrated with the existing loss functions and network architectures. It is worth noting that, although multiple gallery agents are employed for training, only the probe network is needed for inference, without increasing the inference cost. Extensive experiments and comparisons demonstrate the advantages of MASST for long-tail and shallow face learning.</description><identifier>ISSN: 1551-6857</identifier><identifier>EISSN: 1551-6865</identifier><identifier>DOI: 10.1145/3594669</identifier><language>eng</language><publisher>New York, NY: ACM</publisher><subject>Biometrics ; Computing methodologies</subject><ispartof>ACM transactions on multimedia computing communications and applications, 2023-11, Vol.19 (6), p.1-20</ispartof><rights>Copyright held by the owner/author(s). Publication rights licensed to ACM.</rights><lds50>peer_reviewed</lds50><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed><cites>FETCH-LOGICAL-a239t-c9506c45283a9701b0a2f1833c0c9eb4534507e18aae6710262471cf7df714003</cites><orcidid>0000-0003-3656-2593 ; 0000-0003-1066-4399 ; 0000-0002-7699-0747 ; 0000-0002-3603-2683 ; 0000-0002-9553-7358 ; 0000-0002-5990-7307 ; 0009-0006-7762-6986 ; 0000-0003-1300-1769</orcidid></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><linktopdf>$$Uhttps://dl.acm.org/doi/pdf/10.1145/3594669$$EPDF$$P50$$Gacm$$Hfree_for_read</linktopdf><link.rule.ids>314,776,780,2276,27901,27902,40172,75970</link.rule.ids></links><search><creatorcontrib>Tai, Yichun</creatorcontrib><creatorcontrib>Shi, Hailin</creatorcontrib><creatorcontrib>Zeng, Dan</creatorcontrib><creatorcontrib>Du, Hang</creatorcontrib><creatorcontrib>Hu, Yibo</creatorcontrib><creatorcontrib>Zhang, Zicheng</creatorcontrib><creatorcontrib>Zhang, Zhijiang</creatorcontrib><creatorcontrib>Mei, Tao</creatorcontrib><title>Multi-Agent Semi-Siamese Training for Long-tail and Shallow Face Learning</title><title>ACM transactions on multimedia computing communications and applications</title><addtitle>ACM TOMM</addtitle><description>With the recent development of deep convolutional neural networks and large-scale datasets, deep face recognition has made remarkable progress and been widely used in various applications. However, unlike the existing public face datasets, in many real-world scenarios of face recognition, the depth of training dataset is shallow, which means only two face images are available for each ID. With the non-uniform increase of samples, such issue is converted to a more general case, a.k.a long-tail face learning, which suffers from data imbalance and intra-class diversity dearth simultaneously. These adverse conditions damage the training and result in the decline of model performance. Based on the Semi-Siamese Training (SST), we introduce an advanced solution, named Multi-Agent Semi-Siamese Training (MASST), to address these problems. MASST includes a probe network and multiple gallery agents, the former aims to encode the probe features, and the latter constitutes a stack of networks that encode the prototypes (gallery features). For each training iteration, the gallery network, which is sequentially rotated from the stack, and the probe network form a pair of semi-siamese networks. We give the theoretical and empirical analysis that, given the long-tail (or shallow) data and training loss, MASST smooths the loss landscape and satisfies the Lipschitz continuity with the help of multiple agents and the updating gallery queue. The proposed method is out of extra-dependency, thus can be easily integrated with the existing loss functions and network architectures. It is worth noting that, although multiple gallery agents are employed for training, only the probe network is needed for inference, without increasing the inference cost. Extensive experiments and comparisons demonstrate the advantages of MASST for long-tail and shallow face learning.</description><subject>Biometrics</subject><subject>Computing methodologies</subject><issn>1551-6857</issn><issn>1551-6865</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2023</creationdate><recordtype>article</recordtype><recordid>eNo9kE1LAzEYhIMoWKt495Sbp2jezdfmWIrVwoqHreflbZrUyH5IsiL-e1tae5qBeRiGIeQW-AOAVI9CWam1PSMTUAqYLrU6P3llLslVzp-cC62knpDl63c7Rjbb-n6kte8iqyN2Pnu6Shj72G9pGBKthn7LRowtxX5D6w9s2-GHLtB5WnlMe-6aXARss7856pS8L55W8xdWvT0v57OKYSHsyJxVXDupilKgNRzWHIsApRCOO-vXUgmpuPFQInptgBe6kAZcMJtgQO52T8n9odelIefkQ_OVYofptwHe7B9ojg_syLsDia47Qf_hHyjTU4g</recordid><startdate>20231130</startdate><enddate>20231130</enddate><creator>Tai, Yichun</creator><creator>Shi, Hailin</creator><creator>Zeng, Dan</creator><creator>Du, Hang</creator><creator>Hu, Yibo</creator><creator>Zhang, Zicheng</creator><creator>Zhang, Zhijiang</creator><creator>Mei, Tao</creator><general>ACM</general><scope>AAYXX</scope><scope>CITATION</scope><orcidid>https://orcid.org/0000-0003-3656-2593</orcidid><orcidid>https://orcid.org/0000-0003-1066-4399</orcidid><orcidid>https://orcid.org/0000-0002-7699-0747</orcidid><orcidid>https://orcid.org/0000-0002-3603-2683</orcidid><orcidid>https://orcid.org/0000-0002-9553-7358</orcidid><orcidid>https://orcid.org/0000-0002-5990-7307</orcidid><orcidid>https://orcid.org/0009-0006-7762-6986</orcidid><orcidid>https://orcid.org/0000-0003-1300-1769</orcidid></search><sort><creationdate>20231130</creationdate><title>Multi-Agent Semi-Siamese Training for Long-tail and Shallow Face Learning</title><author>Tai, Yichun ; Shi, Hailin ; Zeng, Dan ; Du, Hang ; Hu, Yibo ; Zhang, Zicheng ; Zhang, Zhijiang ; Mei, Tao</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-a239t-c9506c45283a9701b0a2f1833c0c9eb4534507e18aae6710262471cf7df714003</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2023</creationdate><topic>Biometrics</topic><topic>Computing methodologies</topic><toplevel>peer_reviewed</toplevel><toplevel>online_resources</toplevel><creatorcontrib>Tai, Yichun</creatorcontrib><creatorcontrib>Shi, Hailin</creatorcontrib><creatorcontrib>Zeng, Dan</creatorcontrib><creatorcontrib>Du, Hang</creatorcontrib><creatorcontrib>Hu, Yibo</creatorcontrib><creatorcontrib>Zhang, Zicheng</creatorcontrib><creatorcontrib>Zhang, Zhijiang</creatorcontrib><creatorcontrib>Mei, Tao</creatorcontrib><collection>CrossRef</collection><jtitle>ACM transactions on multimedia computing communications and applications</jtitle></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext</fulltext></delivery><addata><au>Tai, Yichun</au><au>Shi, Hailin</au><au>Zeng, Dan</au><au>Du, Hang</au><au>Hu, Yibo</au><au>Zhang, Zicheng</au><au>Zhang, Zhijiang</au><au>Mei, Tao</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>Multi-Agent Semi-Siamese Training for Long-tail and Shallow Face Learning</atitle><jtitle>ACM transactions on multimedia computing communications and applications</jtitle><stitle>ACM TOMM</stitle><date>2023-11-30</date><risdate>2023</risdate><volume>19</volume><issue>6</issue><spage>1</spage><epage>20</epage><pages>1-20</pages><issn>1551-6857</issn><eissn>1551-6865</eissn><abstract>With the recent development of deep convolutional neural networks and large-scale datasets, deep face recognition has made remarkable progress and been widely used in various applications. However, unlike the existing public face datasets, in many real-world scenarios of face recognition, the depth of training dataset is shallow, which means only two face images are available for each ID. With the non-uniform increase of samples, such issue is converted to a more general case, a.k.a long-tail face learning, which suffers from data imbalance and intra-class diversity dearth simultaneously. These adverse conditions damage the training and result in the decline of model performance. Based on the Semi-Siamese Training (SST), we introduce an advanced solution, named Multi-Agent Semi-Siamese Training (MASST), to address these problems. MASST includes a probe network and multiple gallery agents, the former aims to encode the probe features, and the latter constitutes a stack of networks that encode the prototypes (gallery features). For each training iteration, the gallery network, which is sequentially rotated from the stack, and the probe network form a pair of semi-siamese networks. We give the theoretical and empirical analysis that, given the long-tail (or shallow) data and training loss, MASST smooths the loss landscape and satisfies the Lipschitz continuity with the help of multiple agents and the updating gallery queue. The proposed method is out of extra-dependency, thus can be easily integrated with the existing loss functions and network architectures. It is worth noting that, although multiple gallery agents are employed for training, only the probe network is needed for inference, without increasing the inference cost. Extensive experiments and comparisons demonstrate the advantages of MASST for long-tail and shallow face learning.</abstract><cop>New York, NY</cop><pub>ACM</pub><doi>10.1145/3594669</doi><tpages>20</tpages><orcidid>https://orcid.org/0000-0003-3656-2593</orcidid><orcidid>https://orcid.org/0000-0003-1066-4399</orcidid><orcidid>https://orcid.org/0000-0002-7699-0747</orcidid><orcidid>https://orcid.org/0000-0002-3603-2683</orcidid><orcidid>https://orcid.org/0000-0002-9553-7358</orcidid><orcidid>https://orcid.org/0000-0002-5990-7307</orcidid><orcidid>https://orcid.org/0009-0006-7762-6986</orcidid><orcidid>https://orcid.org/0000-0003-1300-1769</orcidid><oa>free_for_read</oa></addata></record>
fulltext fulltext
identifier ISSN: 1551-6857
ispartof ACM transactions on multimedia computing communications and applications, 2023-11, Vol.19 (6), p.1-20
issn 1551-6857
1551-6865
language eng
recordid cdi_crossref_primary_10_1145_3594669
source ACM Digital Library
subjects Biometrics
Computing methodologies
title Multi-Agent Semi-Siamese Training for Long-tail and Shallow Face Learning
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-02-05T04%3A59%3A39IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-acm_cross&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=Multi-Agent%20Semi-Siamese%20Training%20for%20Long-tail%20and%20Shallow%20Face%20Learning&rft.jtitle=ACM%20transactions%20on%20multimedia%20computing%20communications%20and%20applications&rft.au=Tai,%20Yichun&rft.date=2023-11-30&rft.volume=19&rft.issue=6&rft.spage=1&rft.epage=20&rft.pages=1-20&rft.issn=1551-6857&rft.eissn=1551-6865&rft_id=info:doi/10.1145/3594669&rft_dat=%3Cacm_cross%3E3594669%3C/acm_cross%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true