AcrossPlace: Do Vision--Language Models Recognize Places Beyond Their Source Views? A Comprehensive Benchmark
Abstract
Language-guided place recognition requires recovering geographic identity across visual observations beyond the image from which a description originates. However, aggregate retrieval scores often entangle observation transfer, gallery competition, and linguistic evidence, obscuring the underlying recognition capability. We introduce AcrossPlace, a provenance-aware benchmark establishing a systematic evaluation standard for bidirectional language--vision place recognition. Built upon GSV-Cities, AcrossPlace comprises 56,726 places across 23 cities, 156,800 automatically generated, rule-filtered descriptions, and 113,404 target images. Each description retains its source-image provenance, while retrieval operates on disjoint target observations. Text-to-image (T2I) and image-to-text (I2T) retrieval share a unified place-identity relevance criterion supporting multiple valid matches. Two geographic protocols evaluate unseen places within represented cities (P0) and generalization to held-out cities (P1). Paired observation comparisons, gallery-controlled evaluations, fixed-query scaling, and blinded annotation auditing systematically characterize retrieval performance. Across three training seeds, place-supervised SigLIP2 T2I Recall@1 decreases from 19.69% on source images to 12.31% on independent targets under P0. For fixed P1 queries, expanding the gallery from 500 to 8,425 places reduces Recall@1 from 14.67% to 2.87%. A 240-description audit further reveals discrepancies between automated acceptance and factual evidence. These findings establish AcrossPlace as a reproducible evaluation standard for assessing transferable place identity across visual observations and geographic environments.
Figures & tables
| Benchmark | Scale (unit) | T2I | I2T | Prov. | Paired Obs. | Geo. Ctrl. | Text Audit | Eval |
|---|---|---|---|---|---|---|---|---|
| Street360Loc [ 16 ] | 7k locations | Direct | ||||||
| UniLoc [ 23 ] | 29,438 triplets | Direct | ||||||
| LaVPR [ 4 ] | 651,865 pairs | Direct | ||||||
| DlgQuest-Cities [ 18 ] | 106,880 images | – | Direct | |||||
| AcrossPlace | 56,726 places | Direct/Diagnostic |
| Protocol | Split | Places | Texts | Targets |
|---|---|---|---|---|
| P0 | Train | 54,009 | 149,126 | 107,970 |
| P0 | Validation | 1,355 | 3,876 | 2,710 |
| P0 | Test | 1,362 | 3,798 | 2,724 |
| P1 | Train | 39,706 | 109,809 | 79,390 |
| P1 | Validation | 8,511 | 23,622 | 17,000 |
| P1 | Test | 8,509 | 23,369 | 17,014 |
| Reference | Scoring | T2I | I2T | ||||
|---|---|---|---|---|---|---|---|
| R@1 | R@5 | R@10 | R@1 | R@5 | R@10 | ||
| CLIP B/32 (zero-shot) [ 12 ] | Coarse | 1.76 | 5.87 | 9.35 | 2.68 | 8.48 | 12.74 |
| CLIP B/16 (zero-shot) [ 12 ] | Coarse | 2.13 | 5.61 | 8.77 | 3.63 | 9.58 | 13.62 |
| BLIP (zero-shot) [ 5 ] | Coarse | 4.08 | 10.01 | 13.72 | 4.41 | 11.01 | 16.04 |
| SigLIP2 (zero-shot) [ 20 ] | Coarse | 4.16 | 9.06 | 12.59 | 5.29 | 11.34 | 15.93 |
| SigLIP2 (zero-shot) | + ITM | 7.61 | 16.06 | 20.30 | 10.02 | 19.42 | 24.01 |
| Reference | Scoring | T2I | I2T | ||||
|---|---|---|---|---|---|---|---|
| R@1 | R@5 | R@10 | R@1 | R@5 | R@10 | ||
| CLIP B/32 | Coarse | 0.57 | 1.70 | 2.66 | 0.81 | 2.42 | 3.79 |
| CLIP B/16 | Coarse | 0.42 | 1.51 | 2.46 | 0.92 | 2.66 | 4.08 |
| BLIP | Coarse | 1.14 | 3.30 | 4.97 | 1.10 | 3.27 | 4.91 |
| SigLIP2 | Coarse | 1.20 | 3.54 | 5.23 | 1.42 | 3.72 | 5.30 |
| SigLIP2 | + ITM | 2.75 | 6.35 | 8.60 | 2.66 | 6.49 | 8.80 |
| Protocol | Reference | T2I R@1 | I2T R@1 | ||||
|---|---|---|---|---|---|---|---|
| Source | Target | [95% CI] | Source | Target | [95% CI] | ||
| P0 | BLIP (pretrained) | 7.78 | 3.48 | -4.30 [-5.78, -2.82] | 6.89 | 3.63 | -3.26 [-4.52, -2.00] |
| P0 | SigLIP2 (pretrained) | 5.71 | 4.89 | -0.82 [-2.15, +0.52] | 6.08 | 4.60 | -1.48 [-2.82, -0.22] |
| P0 | SigLIP2 (place-supervised) | 19.69 | 12.31 | -7.39 [-9.24, -5.49] | 18.73 | 10.30 | -8.43 [-10.16, -6.67] |
| P1 | BLIP (pretrained) | 2.85 | 1.12 | -1.73 [-2.12, -1.35] | 2.49 | 1.04 | -1.45 [-1.82, -1.09] |
| P1 | SigLIP2 (pretrained) | 2.99 | 1.27 | -1.72 [-2.10, -1.36] | 2.56 | 1.13 | -1.44 [-1.76, -1.10] |
| Reference | P0 | P1 | ||
|---|---|---|---|---|
| T2I | I2T | T2I | I2T | |
| Pretrained BLIP | 3.88 | 4.00 | 4.12 | 3.40 |
| Pretrained SigLIP2 | 5.36 | 5.08 | 3.64 | 3.14 |
| Supervised SigLIP2 | 13.58 | 11.69 | 10.61 | 9.37 |
| Axis | Judgment | All | Valid | Uncertain |
|---|---|---|---|---|
| Source correctness | Supported | 166 | 81 | 85 |
| Partial | 8 | 2 | 6 | |
| Contradicted | 23 | 14 | 9 | |
| Unknown | 43 | 23 | 20 | |
| Target visibility | Visible | 148 | 75 | 73 |
| Partial | 9 | 2 | 7 |
VLM Evaluation
Vision-language model evaluation measures multimodal systems’ perception, reasoning, grounding, robustness, and reliability across controlled and real-world tasks.
See all 1,042 papers