Learning-based essential matrix estimation for visual localization

Abstract

Visual localization is defined as finding the camera pose from two-dimensional images, which is a core technique in many computer vision tasks, including robot navigation, autonomous driving, augmented/mixed/virtual reality, mapping, etc. In this study, we address the pose estimation problem from a single-color image using a neural network. We propose a coarse-to-fine approach based on a deep learning framework, which consists of two steps: direct regression-based coarse pose estimation that obtains a pose by finding a pose-based similar image retrieval and Siamese network-based essential matrix estimation to obtain a refined pose. Experimental results using the 7-scenes, Cambridge, and RobotCar datasets demonstrate that the proposed method performs better than the existing methods in terms of accuracy and stability.