Reducing Uncertainty in SAR Ship Zero-Shot Instance Segmentation via Neural Schrödinger Bridge

Zero-shot instance segmentation using vision foundation models is attractive for Synthetic Aperture Radar (SAR) imagery; however, the modality and distribution shift between SAR and the RGB domain used for pre-training due to differences in sensor physics and image statistics often yields low-fidelity masks. I address this by translating SAR to pseudo-optical with an unpaired Neural Schrödinger Bridge (UNSB), then segmenting with the Segment Anything Model (SAM) augmented by an Uncertainty-aware head (USAM). USAM exposes predictive reliability via SAM’s predictedIoU head and decomposes errors into three actionable components: ∆XP (prompt/selection), ∆A (task), and ∆Θ (model), which localize where further gains are feasible. I ablate the number of function evaluations (NFE) in UNSB and observe that across four of the five benchmarks (SRSDD, HRSID, LS-SSDD, and SSDD), translation reduces both ∆XP and ∆A : ∆XP declines by about 10–22%, and ∆A falls by roughly 3–22%