서브메뉴
검색
Challenges in 3D Machine Learning: Data Scarcity and Structure-Awareness
Challenges in 3D Machine Learning: Data Scarcity and Structure-Awareness
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20250211153059
- ISBN
- 9798346390992
- DDC
- 100
- 서명/저자
- Challenges in 3D Machine Learning: Data Scarcity and Structure-Awareness
- 발행사항
- [Sl] : Stanford University, 2024
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2024
- 형태사항
- 144 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 86-05, Section: A.
- 주기사항
- Advisor: Guibas, Leonidas.
- 학위논문주기
- Thesis (Ph.D.)--Stanford University, 2024.
- 초록/해제
- 요약3D content creation is an important and increasingly popular task for different applications such as graphic design, gaming and animation, virtual reality experiences, manufacturing, and robotics. This growing popularity creates a strong demand for tools that facilitate the creation and synthesis of new 3D content. Traditionally, designers would need to manually design and create each new model, which is a difficult and labor intensive task, requiring expert knowledge of 3D modeling or computer-aided design (CAD) tools. With the advent of 3D machine learning, there is an opportunity to make the 3D creation process more accessible -- but there are challenges on the way. For example, difficulties arise when 3D data repositories of sufficient quality and size are not available, making approaches that require 3D supervision implausible. Another difficulty occurs when structure-aware, fine-grained control or further manipulation and editing is desired for the output, a task for which coarse homogeneous 3D representations are not ideal. In this dissertation, we address two important challenges in 3D machine learning, namely data scarcity and structure-awareness, thus taking promising steps towards automating more parts of the 3D content creation process.First, we introduce a novel approach for reconstructing a 3D scene when no 3D data is available for direct supervision at training time and instead only a sparse set of image captures are given. We show that by introducing some advances in neural rendering, we are able to coarsely reconstruct a 3D scene from only few 2D input views by modeling ambiguity-aware depth estimates. Moreover, we additionally introduce a new image formation model that allows better neural-rendering-based geometric reconstruction. Second, we address the challenge of creating 3D content that is structure-aware, allowing for the creation of 3D assets with a fine level of geometric detail as well as enabling further user manipulation and editing. Concretely, we propose an approach that leverages existing artist-generated databases (meshes) to create controllable deformations of existing 3D data by conditioning on target images or scans. We also study learning representations of shapes that are more structured, e.g., CAD models, and are directly editable, thus improving user control. Depending on the task, use-case and user preference, these techniques offer different advantages towards making 3D content creation more universally accessible.
- 일반주제명
- Ambiguity
- 일반주제명
- Deformation
- 일반주제명
- Deadlines
- 일반주제명
- Semantics
- 일반주제명
- Logic
- 일반주제명
- Engineering
- 기타저자
- Stanford University.
- 기본자료저록
- Dissertations Abstracts International. 86-05A.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008250123s2024 us c eng d■001000017164890
■00520250211153059
■006m o d
■007cr#unu||||||||
■020 ▼a9798346390992
■035 ▼a(MiAaPQ)AAI31652055
■035 ▼a(MiAaPQ)Stanfordpr895qw3725
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a100
■1001 ▼aUy, Mikaela Angelina.
■24510▼aChallenges in 3D Machine Learning: Data Scarcity and Structure-Awareness
■260 ▼a[Sl]▼bStanford University▼c2024
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2024
■300 ▼a144 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 86-05, Section: A.
■500 ▼aAdvisor: Guibas, Leonidas.
■5021 ▼aThesis (Ph.D.)--Stanford University, 2024.
■520 ▼a3D content creation is an important and increasingly popular task for different applications such as graphic design, gaming and animation, virtual reality experiences, manufacturing, and robotics. This growing popularity creates a strong demand for tools that facilitate the creation and synthesis of new 3D content. Traditionally, designers would need to manually design and create each new model, which is a difficult and labor intensive task, requiring expert knowledge of 3D modeling or computer-aided design (CAD) tools. With the advent of 3D machine learning, there is an opportunity to make the 3D creation process more accessible -- but there are challenges on the way. For example, difficulties arise when 3D data repositories of sufficient quality and size are not available, making approaches that require 3D supervision implausible. Another difficulty occurs when structure-aware, fine-grained control or further manipulation and editing is desired for the output, a task for which coarse homogeneous 3D representations are not ideal. In this dissertation, we address two important challenges in 3D machine learning, namely data scarcity and structure-awareness, thus taking promising steps towards automating more parts of the 3D content creation process.First, we introduce a novel approach for reconstructing a 3D scene when no 3D data is available for direct supervision at training time and instead only a sparse set of image captures are given. We show that by introducing some advances in neural rendering, we are able to coarsely reconstruct a 3D scene from only few 2D input views by modeling ambiguity-aware depth estimates. Moreover, we additionally introduce a new image formation model that allows better neural-rendering-based geometric reconstruction. Second, we address the challenge of creating 3D content that is structure-aware, allowing for the creation of 3D assets with a fine level of geometric detail as well as enabling further user manipulation and editing. Concretely, we propose an approach that leverages existing artist-generated databases (meshes) to create controllable deformations of existing 3D data by conditioning on target images or scans. We also study learning representations of shapes that are more structured, e.g., CAD models, and are directly editable, thus improving user control. Depending on the task, use-case and user preference, these techniques offer different advantages towards making 3D content creation more universally accessible.
■590 ▼aSchool code: 0212.
■650 4▼aAmbiguity
■650 4▼aDeformation
■650 4▼aDeadlines
■650 4▼aSemantics
■650 4▼aLogic
■650 4▼aEngineering
■690 ▼a0800
■690 ▼a0395
■690 ▼a0537
■71020▼aStanford University.
■7730 ▼tDissertations Abstracts International▼g86-05A.
■790 ▼a0212
■791 ▼aPh.D.
■792 ▼a2024
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17164890▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


