Leveraging Neural Radiance Field for Spare Keypoints Scene Coordinate Regressor

Abstract

Spare Keypoint Scene Coordinate Regressor (SKSCR) represents a learning-based approach to visual localization, aiming to overcome challenges posed by structural methods, including concerns related to storage, speed, and privacy. While SKSCR demonstrates competitive performance with structural methods across various benchmarks, its efficacy diminishes in scenarios with limited training data, inherent to deep learning models. In this study, we explore the integration of Neural Radiance Fields for synthesizing keypoint descriptors, providing a potential enhancement to SKSCR’s performance in situations where data availability is scarce.