PRIVER: Prompt-Stable Reciprocal Inter-Scale Visual Evidence Reranking for High-Resolution Remote-Sensing Imagery
Abstract
High-resolution remote-sensing images contain dense object layouts and fine spatial details, but their spatial extent makes direct vision–language interpretation difficult. Patch retrieval offers a practical alternative by returning local regions as candidate evidence for downstream interpretation. However, independent semantic scores can leave relevant patches at isolated ranks or return lists with substantial irrelevant content. We propose Prompt-Stable Reciprocal Inter-scale Visual Evidence Reranking (PRIVER), a post-retrieval method for ranking patches within one image using controlled object-category queries. It aggregates four predefined prompt templates and combines semantic scores with geometric support, principally from a degree-normalized inter-scale graph. No encoder weights are updated, and retrieval and reranking do not access object annotations. Parameters are selected on annotated DOTA-v1.5 development data and held fixed. With relevance defined by annotated object coverage, Precision@10 on 458 DOTA-v1.5 internal-test images increases from 0.567 to 0.625 (absolute gain 0.058; 95% source-image cluster bootstrap interval [0.052, 0.065]). Normalized discounted cumulative gain at rank 10 (nDCG@10) increases from 0.647 to 0.730 and distinct-object coverage at rank 10 from 0.773 to 0.873. On a disjoint 526-image SODA-A subset, the corresponding metrics increase from 0.414 to 0.502, 0.445 to 0.550, and 0.467 to 0.681. Coverage-based gains persist with GeoRSCLIP and RemoteCLIP and under altered tiling policies on fixed image subsets. Inter-scale support is the principal source of improvement. These results support training-free, coverage-defined patch prioritization for within-image visual evidence retrieval.