Recognizing a vessel's type from an image is useful for maritime surveillance and sanctions monitoring. But a classifier that appears accurate may be taking a shortcut: instead of focusing on the distinctive superstructure, it may rely on the generic hull, which is less informative for distinguishing vessel types.
The new method, called HullProbe, is designed to measure this kind of reliance. It uses an interventional approach to quantify how much a classifier depends on different parts of a vessel image, specifically separating the hull from the superstructure.
By exposing when a model is leaning on generic hull features, HullProbe could help developers build more robust classifiers for maritime applications. The paper's abstract suggests this reliance is a concern, though the full findings are not detailed in the available text.