서브메뉴
검색
Minimal 3D Priors for Sparse View Reconstruction
Minimal 3D Priors for Sparse View Reconstruction
상세정보
- 자료유형
- 학위논문 서양
- 최종처리일시
- 20260202105212
- ISBN
- 9798291564776
- DDC
- 004
- 저자명
- Rockwell, Chris.
- 서명/저자
- Minimal 3D Priors for Sparse View Reconstruction
- 발행사항
- [Sl] : University of Michigan, 2025
- 발행사항
- Ann Arbor : ProQuest Dissertations & Theses, 2025
- 형태사항
- 173 p
- 주기사항
- Source: Dissertations Abstracts International, Volume: 87-02, Section: B.
- 주기사항
- Advisor: Fouhey, David F.;Johnson, Justin.
- 학위논문주기
- Thesis (Ph.D.)--University of Michigan, 2025.
- 초록/해제
- 요약Sparse view reconstruction is a fundamental problem in computer vision with applications in extended reality, generative AI and robotics. Unfortunately, while learned methods have found great success in 2D and more recently dense view 3D, obtaining 3D from sparse views is an under-constrained problem challenging state-of-the-art models. The overarching idea of this dissertation is that we need only to introduce minimal 3D priors to these learned methods to significantly improve sparse view performance. We focus on the central tasks of novel view synthesis and relative camera pose estimation, proposing state-of-the-art published frameworks containing thorough comparison with prior methods and ablations, confirming the effectiveness of the proposed 3D priors. "PixelSynth" approaches the task of novel view synthesis across large view changes given a single image. This is a challenging task requiring the model to imagine new content, and produce it in a manner consistent with other generated views. While prior work is not up to this task, our contribution of 3D-Consistent Scene Extrapolation achieves high-quality and consistent extrapolations by incorporating a 3D representation into a generative model. "The 8-Point ViT" estimates relative camera pose, including scale, from a pair of wide-baseline images. We show adding minimal 3D Priors can enable a ViT block to compute entries used in the classical 8-Point Algorithm, leading to large gains over prior work, particularly in the case of limited data. We refer to this as Geometry-Guided Relative Pose Estimation. We further improve pose performance in "FAR." It uses the key insight that learned methods like "The 8-Point ViT" tend to be relatively robust, while classical correspondence solvers tend to be relatively precise. "FAR" utilizes the best of both methods in an adaptable design, producing "Flexible, Accurate and Robust" pose estimation. Finally, we apply a similar combination of 3D solvers with learned methods to the case of dense views. Armed with more views, we show state-of-the-art results in the highly challenging case of dynamic Internet video, and collect the largest publicly available diverse, dynamic camera pose dataset "DynPose-100K".
- 일반주제명
- Computer science
- 일반주제명
- Computer engineering
- 키워드
- 3D priors
- 키워드
- Camera pose
- 기타저자
- University of Michigan Computer Science & Engineering
- 기본자료저록
- Dissertations Abstracts International. 87-02B.
- 전자적 위치 및 접속
- 로그인 후 원문을 볼 수 있습니다.
MARC
008260126s2025 us c eng d■001000017359779
■00520260202105212
■006m o d
■007cr#unu||||||||
■020 ▼a9798291564776
■035 ▼a(MiAaPQ)AAI32271721
■035 ▼a(MiAaPQ)umichrackham006253
■040 ▼aMiAaPQ▼cMiAaPQ
■0820 ▼a004
■1001 ▼aRockwell, Chris.
■24510▼aMinimal 3D Priors for Sparse View Reconstruction
■260 ▼a[Sl]▼bUniversity of Michigan▼c2025
■260 1▼aAnn Arbor▼bProQuest Dissertations & Theses▼c2025
■300 ▼a173 p
■500 ▼aSource: Dissertations Abstracts International, Volume: 87-02, Section: B.
■500 ▼aAdvisor: Fouhey, David F.;Johnson, Justin.
■5021 ▼aThesis (Ph.D.)--University of Michigan, 2025.
■520 ▼aSparse view reconstruction is a fundamental problem in computer vision with applications in extended reality, generative AI and robotics. Unfortunately, while learned methods have found great success in 2D and more recently dense view 3D, obtaining 3D from sparse views is an under-constrained problem challenging state-of-the-art models. The overarching idea of this dissertation is that we need only to introduce minimal 3D priors to these learned methods to significantly improve sparse view performance. We focus on the central tasks of novel view synthesis and relative camera pose estimation, proposing state-of-the-art published frameworks containing thorough comparison with prior methods and ablations, confirming the effectiveness of the proposed 3D priors. "PixelSynth" approaches the task of novel view synthesis across large view changes given a single image. This is a challenging task requiring the model to imagine new content, and produce it in a manner consistent with other generated views. While prior work is not up to this task, our contribution of 3D-Consistent Scene Extrapolation achieves high-quality and consistent extrapolations by incorporating a 3D representation into a generative model. "The 8-Point ViT" estimates relative camera pose, including scale, from a pair of wide-baseline images. We show adding minimal 3D Priors can enable a ViT block to compute entries used in the classical 8-Point Algorithm, leading to large gains over prior work, particularly in the case of limited data. We refer to this as Geometry-Guided Relative Pose Estimation. We further improve pose performance in "FAR." It uses the key insight that learned methods like "The 8-Point ViT" tend to be relatively robust, while classical correspondence solvers tend to be relatively precise. "FAR" utilizes the best of both methods in an adaptable design, producing "Flexible, Accurate and Robust" pose estimation. Finally, we apply a similar combination of 3D solvers with learned methods to the case of dense views. Armed with more views, we show state-of-the-art results in the highly challenging case of dynamic Internet video, and collect the largest publicly available diverse, dynamic camera pose dataset "DynPose-100K".
■590 ▼aSchool code: 0127.
■650 4▼aComputer science
■650 4▼aComputer engineering
■653 ▼a3D computer vision
■653 ▼aSparse view reconstruction
■653 ▼a3D priors
■653 ▼aCamera pose
■690 ▼a0984
■690 ▼a0464
■690 ▼a0800
■71020▼aUniversity of Michigan▼bComputer Science & Engineering.
■7730 ▼tDissertations Abstracts International▼g87-02B.
■790 ▼a0127
■791 ▼aPh.D.
■792 ▼a2025
■793 ▼aEnglish
■85640▼uhttp://www.riss.kr/pdu/ddodLink.do?id=T17359779▼nKERIS▼z이 자료의 원문은 한국교육학술정보원에서 제공합니다.


