Support-Derived Multi-Prototypes for Hierarchical Few-Shot Open-Set Insect Classification

Fine-grained insect classification becomes particularly challenging when new species must be recognized from only a few labeled examples while test queries may also belong to species that are not represented in the current support set. This setting combines few-shot generalization with open-set recognition and is additionally structured by biological taxonomy, where species belong to genera and families. We study this problem under class-disjoint episodic learning, where species used during validation and testing are not used as episode classes during meta-training. Species form the episode classes, while genus and family labels provide additional taxonomic supervision through rank-specific prototypes. In this work, we investigate whether representing each episode species by more than one support-derived prototype can better capture intra-class variation in the few available support images, and whether this effect differs between Euclidean and hyperbolic geometry. Rather than introducing persistent learnable class centres, the sub-prototypes are reconstructed independently from the support set of every episode. Support samples are partitioned using a deterministic anchor-based procedure: a central medoid-like sample is selected first, additional anchors are chosen using farthest-first traversal, and the remaining support samples are assigned to their nearest anchor. A prototype is then constructed from each resulting support group, and query-to-species distance is obtained through a soft-min reduction over the sub-prototype distances. Genus and family classes retain a single support-derived prototype. We evaluate the approach on a curated subset of Insect-1M containing 235,339 images from 9,196 species using frozen DINOv2 representations. In 5-way 5-shot evaluation, adding a second species prototype improves the hierarchical Euclidean model from 0.9292 to 0.9326 AUROC, from 0.9065 to 0.9116 OSCR, and from 0.8489 to 0.8599 open-set accuracy, while FPR95 decreases from 0.3470 to 0.3371. In the hierarchical Lorentz model, accuracy improves from 0.9679 to 0.9707, AUROC from 0.9502 to 0.9524, OSCR from 0.9279 to 0.9329, and open-set accuracy from 0.8782 to 0.8855, while FPR95 decreases from 0.2746 to 0.2666. The hyperbolic two-prototype model also outperforms its Euclidean counterpart across the reported classification and open-set metrics. Multi-prototype learning itself is not new. Our contribution is the use of support-derived multi-prototypes inside each few-shot episode in a hierarchical open-set setting with unseen species, together with a controlled comparison under Euclidean and hyperbolic geometry. The results show consistent but modest gains from additional local representatives and stronger overall open-set performance with the Lorentz formulation.