SYNTHETIC-TO-REAL DOMAIN SHIFT IN MARITIME SEARCH AND RESCUE VIDEO SCENES: DECOMPOSITION, MEASUREMENT AND ADAPTATION STRATEGIES

The paper addresses domain shift as the dominant factor limiting the practical applicability of neural detectors in maritime search and rescue (SAR) scenes. A decomposition of the shift into four components — appearance, geometric, prior and conditional — is proposed, and a domain degradation coefficient is introduced as a unified measure. Based on published data it is shown that the shift is class-dependent: for the “person in water” class the degradation reaches 0.98, whereas for “small watercraft” it is 0.65, which suggests — as a hypothesis still to be tested — that the geometric component dominates for the priority class. Adaptation strategies are systematised accordingly.