RSAND: A Fine-Grained Dataset and Benchmark for AtoN Detection in River–Sea Intermodal and Complex Estuarine Environments

Abstract

Robust visual perception of Aids to Navigation (AtoN) is essential for Maritime Autonomous Surface Ships (MASS) operating in restricted navigational waters, where estuarine clutter, fog, glare, and dense traffic can severely degrade detection reliability. Existing maritime vision datasets largely emphasize open-sea targets or coarse AtoN categories, leaving a granularity gap for IALA-compliant fine-grained understanding in river–sea transition and port-approach channels. The River–Sea AtoN Navigation Dataset (RSAND) is introduced, a large-scale benchmark collected along the Yangtze River Deepwater Channel from inland corridors to open estuarine waters. RSAND contains 39,926 images with expert-verified bounding-box annotations and a hierarchical taxonomy that jointly captures AtoN location, shape, and functional semantics across 29 categories. To support both realistic long-tailed evaluation and standardized model comparison, two protocols are provided: RSAND-Full (29 categories) and RSAND-Balanced (10 critical categories). All quantitative benchmarking results in this paper are reported on RSAND-Balanced, while RSAND-Full is released for future large-scale long-tailed robustness studies. Benchmarking experiments on 14 state-of-the-art detectors demonstrate that YOLOv12x achieves superior performance with an mAP50-95 of 80.7%, significantly outperforming previous baselines. However, the analysis reveals persistent challenges in detecting small, distant targets and distinguishing visually similar functional markers. RSAND and the accompanying evaluation toolkit are released to facilitate reproducible research toward safer and smarter marine and coastal navigation.

Bibliographic Information