Neural Feature Filtering for Faster Structure-from-Motion Localisation
Alexandros Rotsidis (University of Bath),* Yuxin Wang (École polytechnique fédérale de Lausanne), Yiorgos Chrysanthou (CYENS Centre of Excellence), Christian Richardt (Meta)
The 34th British Machine Vision Conference
Abstract
Estimating a camera’s pose in an offline map, i.e. camera localisation, is an important task for mobile applications such as augmented reality, self-driving cars and robotics. Many camera localisation pipelines comprise stages for feature detection, matching, outlier filtering, and solving for the camera pose. The bottleneck in localisation pipelines is typically feature matching, which becomes increasingly slower as more features are considered. This work focuses on improving feature matching speed. Specifically, we propose a neural filtering stage that reduces the number of features, drastically reducing feature matching time, with minimal loss in accuracy. This is achieved by training a scene-specific neural network to estimate how reliable (or matchable) each detected feature descriptor is. This allows us to efficiently select only the top most matchable keypoints for the remaining pose estimation pipeline. Our method is applicable to any existing structure-from-motion data. We evaluated our method on large indoor and out-door datasets, and compare to two related methods that address the same problem. We release code of our proposed method.Video
Citation
@inproceedings{Rotsidis_2023_BMVC,
author = {Alexandros Rotsidis and Yuxin Wang and Yiorgos Chrysanthou and Christian Richardt},
title = {Neural Feature Filtering for Faster Structure-from-Motion Localisation},
booktitle = {34th British Machine Vision Conference 2023, {BMVC} 2023, Aberdeen, UK, November 20-24, 2023},
publisher = {BMVA},
year = {2023},
url = {https://papers.bmvc2023.org/0378.pdf}
}
Copyright © 2023
The British Machine Vision Conference is organised by