York-Tank: A Degradation-Aware and Difficulty-Stratified Stereo Dataset for Underwater Visual SLAM

Abstract

Coupled visual degradations in underwater recordings make it difficult to relate simultaneous localization and mapping (SLAM) performance changes to specific factors and their severity. We present York-Tank, which is a controlled stereo dataset combining dominant-factor sequence design with within-factor severity stratification. Its 16 sequences contain 18,948 synchronized stereo image pairs covering low illumination, non-uniform illumination, fast motion, texture deficiency, and dynamic interference. The dataset includes baseline sequences, easy and hard conditions within each degradation category, and hybrid sequences combining multiple degradations. Preset acquisition settings and post-acquisition image and motion indicators characterize the recorded degradation severity. York-Tank-GT provides six-degree-of-freedom reference trajectories using a single overhead camera observing an above-water AprilTag array rigidly connected to the robot without directly relying on degraded underwater stereo images. Evaluations of ORB-SLAM3 and SVO Pro reveal system-dependent responses to the same degradation conditions. Controlled front-end comparisons within ORB-SLAM3 show that the effects of contrast-limited adaptive histogram equalization and dynamic-region feature rejection vary with degradation type and severity. York-Tank supports the targeted evaluation of degradation sensitivity and front-end interventions, complementing field-collected underwater datasets.