A Lightweight Coarse-to-Fine Framework for UAV-to-Satellite Image Retrieval with Learned Reranking and Controllable Blender Simulation

UAV-to-satellite cross-view geo-localization must retain the correct location in a sufficiently broad global shortlist while limiting the online cost of candidate-level geometric verification. Existing work has largely emphasized stronger cross-view representations or multi-frame/trajectory reasoning; this paper instead studies how to allocate a fixed local-verification budget for a single UAV query. We propose a budget-aware coarse-to-fine framework in which MobileNetV3-Small retrieves a Top-100 gallery pool, spatially balanced SuperPoint features provide lightweight candidate-level evidence for either a non-parametric soft-overlap reranker or a learned 29-dimensional MLP reranker, and SuperPoint–LightGlue with RANSAC verifies only the selected Top-10. The reranking module is designed as a matcher-independent candidate allocation component, allowing the same principle to potentially extend to different local verification pipelines. The reranker is used twice: first to allocate the verification budget and then as a candidate prior fused with local geometric evidence for final ranking. On University-1652 selected-view queries, the MLP increases candidate Selection@1 from 53.92% to 61.39%; under the same Top-10 LightGlue budget, final Top-1 accuracy improves from 77.51% without reranking to 80.08% with soft-overlap and 80.13% with MLP-based fusion. On mid-angle queries, MLP fusion reaches 70.36%, compared with the 71.27% high-budget reference obtained by verifying all 100 candidates. A controlled feature ablation shows that the main gain of the 29-dimensional representation comes from combining the CNN retrieval prior with overlap/MNN statistics, while spatial-distribution cues provide a smaller complementary contribution. Separate component profiling supports the lightweight base cascade: MobileNetV3-Small and the sparse SuperPoint–LightGlue verification path have substantially lower model and memory/latency footprints than the ConvNeXt/LoFTR references. These results show that reranking improves candidate quality and recovers part of the accuracy lost by aggressive candidate truncation without increasing the number of locally verified candidates; the computational-efficiency evidence is attributed to the lightweight base cascade rather than to the MLP itself. A Blender module additionally provides a controllable simulation environment for examining heading and perspective variations beyond the real-data operating range.