摘要
针对行人重识别(re-ID)中背景簇、遮挡、视点变化和姿态变化等因素的负面影响,提出了一种新的算法,称为多粒度融合生成对齐网络(MLFGAN).首先,提出身份(Identity,ID)不变的姿态标准化生成对抗网络(IIPN)来生成8种标准姿态的行人图像;然后,通过全局对齐网络实现图像对齐,利用局部对齐网络进行精细化行人匹配,用多粒度特征融合来整合不同粒度的信息,防止小尺度鉴别性信息丢失.MLFGAN能够有效地提取有区分性和鲁棒性的行人嵌入表示,其有效性在Market1501和DukeMTMC-reID等广泛使用的数据集上得到了清楚的验证.
Aiming at the negative influence of background cluster,occlusion,viewpoint change,posture change and other factors in person re-identification(re-ID),a novel algorithm called multi-level fused generated alignment network(MLFGAN)was presented.First,identity(ID)-invariant pose normalized generative adversarial networks(IIPN)was proposed to obtain pedestrian images with eight standard postures.Then,pedestrian images were aligned on a global alignment network,and the partial alignment network was used for finer matching of pedestrian's parts.The multi-level feature fusion was used to integrate information of multiply granularity and prevent the subtle scale but vital discriminative cues loss.MLFGAN was effectively able to extract discriminatory and robust pedestrian embedding representations,and the effectiveness of MLFGAN was clearly demonstrated on two widely used datasets,including Market1501 and DukeMTMC-reID.
作者
种衍文
章郴
冯文强
潘少明
CHONG Yanwen;ZHANG Chen;FENG Wenqiang;PAN Shaoming(State Key Laboratory of Information Engineering in Surveying,Mapping and Remote Sensing,Wuhan University,Wuhan 430079,China)
出处
《华中科技大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2022年第4期64-70,共7页
Journal of Huazhong University of Science and Technology(Natural Science Edition)
基金
国家自然科学基金资助项目(62072345,41671382).
关键词
行人重识别
姿态标准化
多粒度融合
对齐网络
图像检索
person re-identification
pose normalization
multi-level fused
alignment network
image retrieval
作者简介
种衍文(1972-),男,教授,E-mail:chenzhang1@whu.edu.cn.
