GIPACount: Global interaction and prototype-guided aggregation network for exemplar-conditioned few-shot object counting.

Few-shot object counting (FSC) aims to estimate counts for previously unseen (novel) categories given only a few exemplars, leveraging knowledge learned on seen (base) categories for generalization. However, most existing methods rely on unidirectional matching and underuse scene context, limiting the effective use of exemplar priors as reliable guidance. To address these limitations, this paper proposes a global interaction and prototype-guided aggregation counting network (GIPACount), which formulates FSC as a two-stage global-to-local process. First, a global interaction mechanism refines exemplar features and injects global context, yielding exemplar-conditioned global prototype tokens. Next, these prototype tokens generate dynamic filters to aggregate spatial-channel evidence over query image features, producing refined similarity responses for density decoding. In addition, a category-contrastive supervision strategy regularizes training to reduce bias-driven shortcuts. Experiments on FSC-147 and cross-dataset transfer to CARPK, PUCPR+, and ShanghaiTech show that GIPACount effectively transfers counting knowledge from base to novel categories without extra labeled data or retraining, while delivering exemplar-faithful, interpretable density predictions and robust cross-dataset generalization.