Learning from local features to global scene coordinates for camera pose estimation

Abstract

This study presents a simple neural-based mapping method for 6DoF camera pose estimation. We use a neural-representing-based method to query descriptor scene coordinates as an alternative to directly using an expensive 3D model. The final goal is a fast and lightweight solution for estimating the camera pose. This approach is designed to be highly efficient and scalable, making it suitable for various consumer-grade devices that may have limited storage and communication bandwidth.