Deep-direct Visual Localization using learned feature optimization

Convergence of feature-metric optimization

Visual Localization plays important role in development of autonomous systems such as cars and drones. Traditional localization pipeline works in multiple phase: (a) retrieving nearest images from reference database (b) detecting and matching features, and (c) estimating pose from the matched features. In this project, we try to combine (b) and (c) and design algorithm for direct camera pose estimation for a query image given a (query, reference) pair. To design a direct localization algorithm, we use learned CNN features and optimize the pose for a query image to minimize the feature error between a query and a reference image pair. We presented initial results of this work at CVPR 2020 workshop VisLocOdomMap

Ajay Unagar
Ajay Unagar
MSc Student

My research interests include 3D Vision, Reinforcement Learning, and Robotics.