Contour-Guided Feature Selection for Visual Relocalization

Abstract

Scene coordinate regression (SCR) offers an efficient alternative to computation-heavy feature-matching methods for visual relocalization but often suffers from incorrect geometric associations in complex environments. This paper proposes a novel contour-guided feature selection framework to enhance SCR robustness by integrating point and line features. We introduce two key mechanisms: a Feature Contribution Estimation (FCE) module that dynamically reweights features to suppress noise, and a Contour Guidance (CG) module that leverages edge maps to prioritize geometrically significant structures during training. Extensive experiments on the 7-Scenes and Cambridge Landmarks datasets demonstrate that our method outperforms state-of-the-art learning-based baselines, achieving an average accuracy of 80.6% on the 7-Scenes dataset. This approach encourages the model to learn more stable and semantically meaningful features, ultimately enhancing localization performance.