Back to search

Cornell University

Pushing the Boundaries of 3D Spatial Understanding

Abstract

dc:description.abstract

Understanding the 3D world from data collected by sensors such as cameras and LiDAR is a fundamental problem in computer vision, with applications in robotics, augmented and virtual reality, and autonomous systems. Although current algorithms can produce high-quality 3D reconstructions, they often struggle under real-world conditions, such as when data is sparse, visual overlap is limited, or scenes contain strong ambiguities like repetition and symmetry. This thesis explores how learned priors can help overcome these challenges and improve 3D spatial understanding. I begin by addressing the problem of shape generation and reconstruction from sparse point clouds, proposing a method that learns shape priors through modeling gradient fields over 3D point cloud distributions. Next, I tackle the challenge of extreme camera pose estimation between image pairs with little or no overlap, using dense correlation volumes to extract semantic and geometric cues. Building on this, I further improve extreme pose estimation by leveraging visual world priors from generative video models, which hallucinate plausible intermediate frames to provide useful context. Finally, I address visual ambiguity in structure-from-motion by using 3D priors from feature-matching models to disambiguate visually similar but incorrect matches, what we call “doppelgangers”, in symmetric scenes like the Arc de Triomphe.

Degree

thesis:*
Name thesis:degree_name
Ph. D., Computer Science
Level thesis:degree_level
Doctor of Philosophy
Discipline thesis:degree_discipline
Computer Science
Grantor
Cornell University
Year dc:date.issued
2025

Author and committee

dc:creator, dc:contributor.*
Author dc:creator
  • Cai, Ruojin
Chair dc:contributor.chair
  • Snavely, Keith
Committee members dc:contributor.committeemember
  • Hariharan, Bharath
  • Marschner, Stephen
  • Zhang, Cheng

Subjects

dc:subject × 4

Rights

dc:rights
Statement dc:rights
  • Attribution 4.0 International
Language dc:language.iso
en

Identifiers

dc:identifier.*
Dc Identifier Bibid
17422536
Dc Identifier Other
ProQuest Submission ID: 15215
ProQuest Publication ID: 32170536

Chain of custody

source
Harvested from
Cornell University
Base URL
ecommons.cornell.edu/server/oai/request
Last updated
2026-11-19
Source record
OAI-PMH GetRecord
citation

Cai, Ruojin. Pushing the Boundaries of 3D Spatial Understanding. Doctor of Philosophy thesis, Cornell University, 2025. https://hdl.handle.net/1813/120906