Re-Ranking Methods for Instance-Level Image Retrieval: A Review

Image retrieval re-ranking improves an initial retrieval result by incorporating local image features, neighborhood relationships, or learned similarity functions. This selective narrative review examines re-ranking methods for instance and landmark image retrieval, covering geometric verification, query expansion, neighborhood encoding, graph diffusion, and learned pairwise and listwise models. Methods are classified by their inference evidence, contextual access, and ability to modify the candidate set. Mathematical formulations and their assumptions clarify the relationships among query refinement, reciprocal encoding, graph propagation, and attention. Deterministic numerical examples illustrate query drift and propagation through an erroneous graph edge. Published benchmark results are synthesized through within-study comparisons, with explicit controls for first-stage representations, local evidence, candidate depth, and evaluation protocols. Reported computation and storage are related to the corresponding performance gains. The selected evidence indicates that contextual refinement benefits from reliable neighborhoods, while local verification supplies complementary information under clutter and domain shift. The review concludes with deployment trade-offs and research directions in adaptive computation, uncertainty-aware contextual modeling, compressed local representations, dynamic galleries, and multimodal re-ranking.