HOSS-Bench: Open-Set Cross-Modal Vessel Re-Identification Benchmark

Vessel re-identification aims to match a query image of a vessel against a gallery of candidate images to determine if the same vessel appears. Cross-modal RGB-SAR vessel re-identification is particularly challenging due to large appearance differences across sensing modalities. The Hybrid Optical and SAR Ship Re-Identification Dataset (HOSS) was recently released alongside a baseline method. In this work, we extend the evaluation protocol of HOSS into a realistic benchmark by (a) proposing an improved training/validation/test split; (b) introducing open-set evaluation, where the query image may depict an unknown vessel; (c) measuring re-identification performance in the most challenging cross-modal setting where the query vessel appears in the gallery only in a different modality, while the gallery contains multiple similar vessels in the same modality as the query. We formulate vessel re-identification as an open-set recognition problem and use open-set accuracy as the primary metric, alongside standard closed-set ReID metrics. We re-train the TransOSS model according to the new protocol and compare it with a DINOv3-based baseline, highlighting that current performance remains insufficient for realistic online open-set vessel clustering applications.