RIM: A Retrieval-In-Matching Framework for Cross-Domain Global Visual Localization of UAVs

It proposes the RIM framework, unifying retrieval and matching within a single query-side DINOv2-B forward by freezing the adapted retriever and distilling a local-descriptor decoder from its token field and a shallow VGG19 detail stream. Relative to separate retrieval-then-matching pipelines that need an additional foundation-model backbone, this design preserves the retrieval descriptors by construction and eliminates a second foundation-model backbone. The abstract states the architectural design and reports comparison against ten retrieval baselines, but the loaded text does not include the detailed ablation tables or implementation specifics.

It renders UAV-viewpoint references from Google 3D Tiles across locations, altitudes, and orientations, adapts SALAD in two stages with pose-near positives and geographically distant hard negatives, and re-ranks Top-K candidates by local geometric consistency. For cross-domain appearance and viewpoint shifts arising from differences in acquisition time and imaging platform between UAV and reference imagery, it offers a combined data-construction and training strategy. The abstract describes the full pipeline but does not give the rendering scale or the number of positive and negative samples.

Zero-shot evaluation on EPFL Urbanscape and Chang'an Park, both geographically disjoint from the training data, improves Recall@1 over SALAD by 8.55/13.77 and 4.45/8.94 percentage points under the full 3D distance metric at 25/50 m. It reports cross-domain generalization under a geographically disjoint zero-shot setting and quantifies the gain relative to SALAD. Two datasets with explicit distance thresholds and percentage-point gains, though the loaded text provides no sample sizes, repetitions, or confidence intervals.

At Top-K=5 the online query path through retrieval, candidate matching, and robust geometric verification takes 90.8 ms, 1.2 times faster than the strongest separate sparse-matching baseline and over 30 times faster than RoMa, with comparable re-ranking accuracy. It reports end-to-end online latency while maintaining re-ranking accuracy, pointing toward a practical localization pipeline under unreliable satellite navigation. Concrete millisecond-level latency and relative speedup factors are given, but the measurement hardware, batch settings, and statistical treatment are not described.

The result targets global visual localization scenarios that use remote-sensing reference maps and can render UAV-viewpoint references from Google 3D Tiles, applicable to 6-DoF pose estimation when satellite navigation is unreliable; its zero-shot conclusions rest on the EPFL Urbanscape and Chang'an Park datasets, both geographically disjoint from the training data, and can next be examined under this setting across more regions, more reference data sources, and different flight conditions.

The loaded text is the arXiv abstract and metadata page and does not include the main-body figures, ablation experiments, sample sizes, or statistical details, so the scale of rendered references, the specific parameters of positive and negative construction, the hardware and batch settings of the latency measurement, and the stability of the gains across distance thresholds still need confirmation in the full text; how strongly the method depends on the coverage and quality of Google 3D Tiles is also an open question worth watching.