Camera Relocalization through Distribution Modeling

QC 20251117

Abstract

Relocalization is a key component of robot navigation: in order to move successfully within an environment, a robot must know its location in relation to that environment. Cameras are inexpensive sensors that enable relocalization by comparing visual observations with a model of the scene. To this end, camera relocalization, which also finds applications in augmented reality, has long been a topic of research, leading to elaborately designed pipelines for accurate camera pose estimation. Recently, a paradigm shift has seen explicit models of the scene replaced by implicit ones, where the scene is encoded in the weights of neural networks. This shift simplifies relocalization pipelines but leaves open a fundamental challenge: scenes with repetitive structures often produce ambiguous observations, meaning that the same visual input can correspond to multiple distinct camera poses. This thesis addresses this challenge, with a particular focus on implicit relocalization methods. It critically examines the assumptions underlying existing paradigms such as Absolute Pose Regression (APR) and Scene Coordinate Regression (SCR) about the uniqueness of appearances. As its central contribution, the thesis proposes to model the full distribution of possible solutions, which can be arbitrarily shaped, rather than attempting to recover a single best estimate. To this end, it proposes to leverage Conditional Variational Autoencoders (C-VAEs) as generative models capable of representing both distributions over poses and distributions over points. Furthermore, likelihood estimation within this framework provides a principled means of attaching confidence measures to predictions. These contributions, together with the suggested applications and directions for future work, lay a foundation for simplifying relocalization pipelines by more effectively handling ambiguities in observations.

urn.kb.se/resolve?urn=urn:nbn:se:kth:diva-372920