Source-linked AI summary
GeoStore: Finding Small Storefronts in Large Scenes -- A Fine-Grained POI Localization Benchmark with Global-to-Local Asymmetric Matching
Lu Han, Xiting Sun, Hao Wang, Zhiqiang Cao, Ruihuan Du, Ziquan Zeng, Chunlong Lv
TL;DR
POI localization matches close-up storefront queries to wide vehicle-mounted street views, a fine-grained open-set setting that conventional symmetric VPR benchmarks do not capture. The paper introduces GeoStore and GLAM, whose asymmetric local pathway complements global retrieval; GLAM achieves the best accuracy on GeoStore at lower storage and matching cost than dense local re-ranking.
Problem
POI localization must match close-up user photos against wide street-view images despite scale asymmetry, fine-grained open-set matching, and capture-domain heterogeneity.
Method
The paper introduces GeoStore and GLAM, which combines a retrieval-anchoring global descriptor with compact reference region tokens matched to a single query probe and reused for lightweight re-ranking.
Results
GLAM achieves the best accuracy on GeoStore, surpassing global and two-stage VPR baselines while using a fraction of the storage and matching cost of dense local re-ranking.
Takeaways & Limitations
GeoStore establishes an open-set benchmark for asymmetric POI localization, while GLAM shows that global retrieval can be strengthened by index-friendly local evidence.
Abstract
from arXiv · showhide
Point-of-interest (POI) localization -- matching a user's close-up storefront photograph against large-scale geo-tagged street-view imagery -- underpins map construction, POI verification, and location-based services. Its closest existing paradigm, visual place recognition (VPR), assumes symmetric, whole-image matching of the same scene at a comparable scale; POI localization instead must match a close-up query, in which the target fills the frame, against wide references in which the same POI occupies only a small, off-center region among visually similar shops, under a substantial capture-domain gap. We introduce GeoStore, to our knowledge the first benchmark dedicated to this asymmetric, fine-grained, open-set formulation, and show that global-descriptor methods tuned for symmetric VPR are systematically limited on it, since a single global vector dilutes the small target. We further propose GLAM (Global-to-Local Asymmetric Matching), which couples a retrieval-anchoring global descriptor with an asymmetric local pathway: each reference is kept as a compact set of pooled region tokens and matched against a single query probe through a learnable soft late interaction; at inference, the same tokens enable a lightweight mutual-nearest-neighbor re-ranking. GLAM surpasses strong global and two-stage baselines on Recall@1/5/10 and mAP, with ~5x smaller re-ranking features and ~two orders of magnitude lower per-pair matching cost than prior local re-ranking. The benchmark and code will be publicly released.
1. INTRODUCTION
POI localization matches close-up user photos to wide vehicle-mounted street views, creating asymmetric, fine-grained, open-set conditions that conventional symmetric VPR benchmarks and global descriptors do not adequately address. GeoStore formalizes this setting, while GLAM combines global retrieval anchoring with asymmetric local matching and outperforms strong baselines.
- POI localization supports map construction, POI verification, merchant annotation, and location-based services by matching user-uploaded close-ups to geo-tagged vehicle imagery.
- Unlike symmetric VPR, POI localization requires instance-level matching across different scales, fields of view, scene extents, and capture domains.
- GeoStore is an open-set benchmark with close-up user queries and wide-field vehicle references containing complex backgrounds.
- Global descriptors are systematically limited because summarizing the entire reference image dilutes the small target storefront.
- GLAM couples a retrieval-anchoring global descriptor with compact reference region tokens, a single query probe, learnable soft late interaction, and mutual-nearest-neighbor re-ranking.
- GLAM surpasses strong global and two-stage VPR methods on Recall@1/5/10 at lower storage and matching cost than prior local re-ranking.
2. GEOSTORE BENCHMARK
GeoStore is constructed from independent production streams of user-uploaded close-ups and continuously collected vehicle-mounted street views. Its curated open-set benchmark spans disjoint places and images, with test queries ranked against a shared reference database.
- GeoStore jointly captures heterogeneous asymmetric sources, a real capture-domain gap, fine-grained open-set recognition, and realistic weather and day–night conditions.
- Queries are handheld close-range photographs of storefronts, signboards, or entrances, while references are wide-field dashcam frames collected continuously on urban roads.
- The benchmark contains 1,215 places and 11,133 reference images, split into 972 training and 243 test places disjoint at both place and image levels.
- Each of 170 valid test queries is ranked against the union of references from all test places, using Recall@1/5/10 and mAP.
3. METHOD: GLAM
GLAM resolves asymmetric POI matching by anchoring retrieval globally while matching a query probe against compact reference region tokens locally. The fused model is trained to make those tokens independently discriminative and reuses them for lightweight re-ranking.
- 3. METHOD: GLAM: GLAM combines a retrieval-anchoring global descriptor with an asymmetric local pathway and mutual-nearest-neighbor re-ranking.The local pathway matches compact reference tokens against a single query probe, while the same tokens support inference-time re-ranking.
- 3. METHOD: GLAM: A shared DINOv2 backbone produces an 8448-d SALAD descriptor for global cosine retrieval.The global branch preserves pretrained similarity structure and provides sensible retrieval from the first training epoch.
- 3. METHOD: GLAM: The local pathway converts each reference into 361 pooled region tokens and each query into one attention-pooled probe.A 5×5 average-pooling operator with stride 2 and a 1×1 projection to D=256 provides the region representation; the asymmetric read-out searches the reference, not the query.
- 3. METHOD: GLAM: Soft top-k MaxSim aggregates the eight strongest probe-to-region similarities using a learnable inverse temperature.This soft selection is more stable than a hard maximum over hundreds of regions, while k=1 reduces to hard maximum behavior.
- 3. METHOD: GLAM: The final similarity fuses global and local scores with a learnable λ initialized small so global retrieval dominates early training.As region tokens become discriminative, λ can increase during training.
- 3. METHOD: GLAM: An auxiliary local-only InfoNCE term with α=0.5 makes region tokens independently discriminative for later re-ranking.Without this auxiliary term, gradients from the stronger global branch leave the local tokens nearly inert.
4. EXPERIMENTS
Experiments on GeoStore evaluate methods under a unified asymmetric query-to-reference protocol. GLAM combines compact local tokens with soft matching and achieves stronger accuracy and efficiency than competing approaches, while qualitative evidence indicates implicit storefront localization.
- 4.2. Main Results: GLAM’s fused stage-1 leads all baselines on R@5, R@10, and mAP, showing that its asymmetric local pathway recovers targets diluted by global descriptors.Under the same mutual-nearest-neighbor re-ranking, GLAM reaches 25.3 R@1 versus FoL’s 20.0 and gains +7.1 versus +5.9 R@1.
- 4.2. Main Results: 25.3 R@1 is achieved by GLAM after mutual-nearest-neighbor re-ranking, compared with 20.0 for FoL.GLAM also gains more from re-ranking: +7.1 versus +5.9 R@1.
- 4.2. Main Results: Raising SALAD’s resolution from 224 to 518 produces no gain, indicating that symmetric global modeling does not exploit high-resolution references.This comparison is reported alongside the fused-stage results in Table 1.
- 4.3. Ablations: Average pooling generalizes best for the region operator, whereas learnable region convolutions overfit place-specific patterns and consistently underperform.The ablation evaluates the local branch using R@1.
- 4.3. Ablations: The soft top-k MaxSim uses k=8 because k=1 is sensitive to spurious matches and k=16 dilutes the target with background tokens.The hard maximum preserves R@1 but substantially degrades R@5, R@10, and mAP.
- 4.4. Efficiency: GLAM stores 5× less local data per reference and makes per-pair matching about two orders of magnitude cheaper than FoL’s dense re-ranking.The efficiency comparison is reported together with higher accuracy in Table 1.
- 4.5. Qualitative Analysis: Soft top-k MaxSim selects spatially contiguous tokens around queried storefronts despite retrieval-only supervision, providing evidence of implicit localization.The qualitative analysis concerns strong reflections, vehicle occlusion, day–night gaps, and dense street clutter.
5. CONCLUSION
The paper formulates POI localization as asymmetric retrieval and introduces GeoStore and GLAM for matching close-up storefront queries to wide vehicle-mounted street views. GLAM achieves the best reported accuracy with lower storage and matching costs than dense local re-ranking, while future work targets larger benchmarks and richer asymmetric interactions.
- 5. CONCLUSION: GeoStore is an open-set benchmark for matching close-up storefront queries against wide vehicle-mounted street views.The conclusion frames POI localization as an asymmetric retrieval problem.
- 5. CONCLUSION: GLAM couples a global anchor with an asymmetric local pathway whose region tokens also support lightweight mutual-nearest-neighbor re-ranking.The same design is presented as the proposed alternative to dense local matching.
- 5. CONCLUSION: GLAM achieves the best accuracy at a fraction of the storage and matching cost of dense local re-ranking.The conclusion states the result at the paper level without specifying individual metric values.
- 5. CONCLUSION: Future work includes scaling GeoStore and exploring richer asymmetric interactions that remain index-friendly.These directions are explicitly identified in the conclusion.