Reliable online updating of class geometry

Develop reliable methods to learn or update class geometry as new categories arrive, particularly when visual and semantic class geometries disagree.

Background

Class-geometry supervision relies on target dissimilarity matrices derived from visual morphology or semantic descriptions to organize prototype spaces for open-world detection. The paper shows that visual geometry is generally more reliable for few-shot biomedical detection, while text-derived and random geometries can produce different performance patterns. As new categories are incrementally introduced, the target geometry must be expanded or updated to encode relationships between new and existing classes. The paper identifies the unresolved challenge of doing this reliably when visual and semantic sources provide conflicting relational information.