서브메뉴
검색
Human-Centric Perception With Limited Supervision: Improving Generalizability in the Wild.
Human-Centric Perception With Limited Supervision: Improving Generalizability in the Wild.
상세정보
- 자료유형
- 학위논문
- Control Number
- 0017163731
- International Standard Book Number
- 9798342114684
- Dewey Decimal Classification Number
- 574
- Main Entry-Personal Name
- Weng, Zhenzhen.
- Publication, Distribution, etc. (Imprint
- [S.l.] : Stanford University., 2024
- Publication, Distribution, etc. (Imprint
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- Physical Description
- 214 p.
- General Note
- Source: Dissertations Abstracts International, Volume: 86-05, Section: B.
- General Note
- Advisor: Yeung, Serena.
- Dissertation Note
- Thesis (Ph.D.)--Stanford University, 2024.
- Summary, Etc.
- 요약Human-centric perception has numerous valuable applications in self-driving vehicles, AI-assisted healthcare, and content creation, where understanding human poses and behaviors enhances planning, intervention strategies, and visual e↵ects. However, challenges such as the high cost of data capture and annotation limit the availability of training data for 3D applications. This thesis addresses these challenges through various methods leveraging priors and pre-trained models to improve perception models with limited supervision.The first part of the thesis explores leveraging priors to enhance perception models. It introduces a method using common room layout knowledge to refine indoor human and scene reconstructions and a method that learns human body priors from unannotated LiDAR sensor data, reducing the need for annotated training data. This is particularly beneficial in self-driving, where large sensor data captures are common, but 3D annotation is complex.The second part presents a study that bootstraps synthetic data for training perception models using human pose priors. Incorporating diverse synthetic data results in more robust models that perform well in challenging scenarios. This method, applied to caregiver-child interaction videos, extracts human behavior aspects to facilitate behavioral and developmental research, enhancing perception models and providing insights into developmental processes.The third part discusses using pre-trained generative models to create realistic synthetic human training data. It introduces methods employing pre-trained generative and di↵usion models to enhance perception models, leveraging powerful 2D models to reduce dependence on high-quality 3D training data.Overall, this thesis demonstrates that data-driven methods for understanding human form and behavior significantly improve our ability to perceive and generate realistic human representations with high fidelity and accuracy.
- Subject Added Entry-Topical Term
- Adaptation.
- Subject Added Entry-Topical Term
- Caregivers.
- Subject Added Entry-Topical Term
- Geometry.
- Subject Added Entry-Topical Term
- Computer engineering.
- Subject Added Entry-Topical Term
- Applied mathematics.
- Added Entry-Corporate Name
- Stanford University.
- Host Item Entry
- Dissertations Abstracts International. 86-05B.
- Electronic Location and Access
- 로그인을 한후 보실 수 있는 자료입니다.
- Control Number
- joongbu:657566
MARC
008250224s2024 us ||||||||||||||c||eng d■001000017163731
■00520250211152746
■006m o d
■007cr#unu||||||||
■020 ▼a9798342114684
■035 ▼a(MiAaPQ)AAI31520294
■035 ▼a(MiAaPQ)Stanfordnk873jd2198
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a574
■1001 ▼aWeng, Zhenzhen.
■24510▼aHuman-Centric Perception With Limited Supervision: Improving Generalizability in the Wild.
■260 ▼a[S.l.]▼bStanford University. ▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a214 p.
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-05, Section: B.
■500 ▼aAdvisor: Yeung, Serena.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼aHuman-centric perception has numerous valuable applications in self-driving vehicles, AI-assisted healthcare, and content creation, where understanding human poses and behaviors enhances planning, intervention strategies, and visual e↵ects. However, challenges such as the high cost of data capture and annotation limit the availability of training data for 3D applications. This thesis addresses these challenges through various methods leveraging priors and pre-trained models to improve perception models with limited supervision.The first part of the thesis explores leveraging priors to enhance perception models. It introduces a method using common room layout knowledge to refine indoor human and scene reconstructions and a method that learns human body priors from unannotated LiDAR sensor data, reducing the need for annotated training data. This is particularly beneficial in self-driving, where large sensor data captures are common, but 3D annotation is complex.The second part presents a study that bootstraps synthetic data for training perception models using human pose priors. Incorporating diverse synthetic data results in more robust models that perform well in challenging scenarios. This method, applied to caregiver-child interaction videos, extracts human behavior aspects to facilitate behavioral and developmental research, enhancing perception models and providing insights into developmental processes.The third part discusses using pre-trained generative models to create realistic synthetic human training data. It introduces methods employing pre-trained generative and di↵usion models to enhance perception models, leveraging powerful 2D models to reduce dependence on high-quality 3D training data.Overall, this thesis demonstrates that data-driven methods for understanding human form and behavior significantly improve our ability to perceive and generate realistic human representations with high fidelity and accuracy.
■590 ▼aSchool code: 0212.
■650 4▼aAdaptation.
■650 4▼aCaregivers.
■650 4▼aGeometry.
■650 4▼aComputer engineering.
■650 4▼aApplied mathematics.
■690 ▼a0464
■690 ▼a0364
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-05B.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17163731▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.
