Fan's survey paper accepted in Computer Vision and Image Understanding

Camera relocalization aims to estimate the 6-DoF pose of a query image relative to a previously mapped scene. Recent advances in neural radiance fields (NeRFs) and 3D Gaussian Splatting (3DGS) have substantially expanded the range of scene representations available for relocalization, making representation type alone insufficient to characterize how camera poses are inferred. This survey presents a unified framework that considers two complementary dimensions: scene representation and pose-inference mechanism. We distinguish explicit, learned, and hybrid scene representations, while categorizing relocalization methods primarily by how the final pose is inferred: direct learned pose inference, geometric correspondence-based inference, and differentiable dense alignment. This taxonomy further clarifies the role of rendering in modern neural scene representations, where rendered observations can either establish correspondences for geometric pose estimation or provide residuals for direct optimization of the camera pose. We review classical landmark-based localization alongside recent NeRF- and 3DGS-based approaches, analyzing their requirements, trade-offs, and failure modes in terms of initialization, map fidelity, appearance variation, robustness, and computational cost. Beyond the core taxonomy, we examine cross-cutting factors and deployment considerations that influence practical performance. By relating these factors to the underlying scene representations and pose-inference mechanisms, this survey provides a principled basis for comparing existing methods and practical guidance for designing and selecting relocalization systems for robotics, augmented reality, and autonomous applications.

2026

  1. Understanding RGB camera relocalization through scene representation and final pose-inference mechanisms: A survey
    Fan Wu, Amine Kacete, Jérôme Royan, and Guillaume Moreau
    Computer Vision and Image Understanding, 2026