Stress-Testing Attention-Based Feature Matchers Under Progressive Image Degradation

Abstract

Robust 6DoF camera relocalization is critical for autonomous robotics, yet the performance of modern local feature matchers under real-world sensor degradation remains poorly quantified. This paper presents a systematic empirical evaluation of visual localization pipelines under progressive JPEG compression, additive Gaussian noise, and motion blur. By decomposing global runtime, we expose two critical anomalies: the Sinkhorn Computational Collapse in dense transformer matchers (SuperGlue), where normalization latency increases exponentially up to 3354.9 ms, and the RANSAC Paradox in context-free matching (NN-mutual), which pollutes spatial solvers with over 87% outliers, extending verification time up to 2099.6 ms. Additionally, we discover a JPEG Low-Pass Regularization Effect that generates localized accuracy peaks. Pareto frontier analysis demonstrates that LightGlue eliminates these bottlenecks, maintaining a predictable execution profile (1013.6–1273.2 ms) and stable 10–15 cm precision, providing a foundational paradigm for resource-constrained embedded systems.