자료유형  

 학위논문

Control Number  

0017161963

International Standard Book Number  

9798382652702

Dewey Decimal Classification Number  

004

Main Entry-Personal Name  

Zhu, Haidong.

Publication, Distribution, etc. (Imprint  

[S.l.] : University of Southern California., 2024

Publication, Distribution, etc. (Imprint  

Ann Arbor : ProQuest Dissertations & Theses, 2024

Physical Description  

169 p.

General Note  

Source: Dissertations Abstracts International, Volume: 85-11, Section: B.

General Note  

Advisor: Nevatia, Ramakant.

Dissertation Note  

Thesis (Ph.D.)--University of Southern California, 2024.

Summary, Etc.  

요약Finding the identity of a person from a non-overlapping input image or video, known as person reidentification, is a classic and important task in biometric understanding. Identifying the corresponding identity requires extracting the representations of the person and distinguishing them across different individuals, where the representation can be human appearance, specific walking patterns, and body shape. Each representation can be understood as a specific modality and has its own strengths and weaknesses, while different modalities can sometimes complement each other. Therefore, combining two or more modalities introduces a more robust system for person re-identification.In this thesis, we cover different combinations of biometric representations for whole-body person re-identification, including appearance, gait, and body shape. Appearance is one of the most widely used biometric signals as it provides abundant information. Gait, represented as skeleton or binary silhouette sequences, captures the walking patterns of a person. Different from other representations, 3-D shape complements the body information with external human body shape prior and enhances the appearance captured in the 2-D images. Body shape also provides a strong prior of the person and helps complete the body shape to deal with occlusions. We discuss the combination of different representations and biometric signals that leverage their strengths, along with a system using the three signals for person re-identification in the wild. As the current body shapes used for person re-identification are usually not accurate enough to provide a distinguishable signal, we further discuss the improvement of the representations and how they can be applied for downstream vision tasks, such as person identification.We begin with three works that explicitly extract and combine different modalities for re-identification, including two gait representations (silhouettes and skeletons), two different shape-related modalities (gait and 3-D body shape), and the additional use of appearance along with the two shape-related modalities. Although 3-D body shape offers invaluable external shape-related information that 2-D images lack, existing body shape representations often fall short in accuracy or demand extensive image data, which is unavailable for re-identification tasks. Following this, we explore the potential of using more accurate body shape to further improve the model and introduce two other methods for more accurate 3-D shape representation and reconstruction: Implicit Functions (IF) and Neural Radiance Fields (NeRF). Since a fine-grained representation is needed for downstream vision tasks, we discuss how to include more semantic representation to assist the training of the 3-D reconstruction model and how it can aid with a limited number of input views. Lastly, with the fine-grained representation, we discuss using them for body shape representation to enhance appearance for person re-identification. We conclude the thesis with potential future work for further improvements.

Subject Added Entry-Topical Term  

Computer science.

Subject Added Entry-Topical Term  

Computer engineering.

Index Term-Uncontrolled  

Biometrics

Index Term-Uncontrolled  

Body shape reconstruction

Index Term-Uncontrolled  

Neural rendering

Index Term-Uncontrolled  

Person re-identification

Index Term-Uncontrolled  

Implicit Functions

Added Entry-Corporate Name  

University of Southern California Computer Science

Host Item Entry  

Dissertations Abstracts International. 85-11B.

Electronic Location and Access  

로그인을 한후 보실 수 있는 자료입니다.

Control Number  

joongbu:656085