Calibrating Retrieval Geometry: Reliability-Guided Training-Free Aggregation for Visual Place Recognition

Xin Li; Zhimin Mao; Shang Wang; Siyuan Duan; Geng Zhang

cs.CVSep 22, 2026

Calibrating Retrieval Geometry: Reliability-Guided Training-Free Aggregation for Visual Place Recognition

Authors: Xin Li, Zhimin Mao, Shang Wang, Siyuan Duan, Geng Zhang

Organizations: Key Laboratory of Spectral Imaging Technology CAS, Xi’an Institute of Optics and Precision Mechanics, Chinese Academy of Sciences, Xi’an, 710119, China · University of Chinese Academy of Sciences, Beijing, 100049, China

Abstract

Frozen visual foundation models provide transferable features for visual place recognition, but fixed aggregation can suppress useful distinctions in new environments. We introduce TFA, a reliability-guided, training-free aggregation method requiring neither place labels nor task-specific weight updates. Our key observation is that reproducible retrieval need not be discriminative: independent codebooks can consistently retrieve a few database hubs. TFA combines cross-codebook agreement, retrieval coverage, and spectral statistics to control residual assignment, spectral shaping, and global-feature fusion. Its spectral kernel exactly recovers original descriptor similarity at zero intervention. Database-only TFA fixes its rules before accessing queries; TFA-C64 uses 64 disjoint unlabeled target images to calibrate retrieval for subsequent queries. Across 20 ground protocols with a fixed DINOv2-B backbone and matched resolution, database-only TFA improves Recall@1 over AnyLoc by 17.39 percentage points on MSLS-val and 9.55 on SPED. C64 mitigates failures of database-only calibration in driving environments. Across eight aerial/cross-view protocols, TFA achieves the highest Recall@1 among compared training-free heads in 14 of 16 DINOv2/DINOv3 backbone-protocol combinations. In a separate native-system comparison, DINOv2-G-based TFA-C64 reaches 91.46% Recall@1 on Pitts30k and 76.29% on VPAIR, outperforming the displayed training-free comparators on all five benchmarks. These results show that reliability-guided aggregation can recover additional retrieval capability from frozen representations, providing a practical baseline for new environments with scarce place supervision.

Figures & tables

Figure 1: TFA overview. DB-fitted geometry and reliability evidence determine a fixed aggregation head. C64 adds 64 disjoint target images for calibration. Dashed arrows: fixed controls; solid: data. Feature diagrams are schematic; retrieval shows DB-only SPED q55, seed 0.

	Supervised References		Training Free Systems
Dataset	SALAD	SelaVPR++	AnyLoc	TF-VPR	RIA d	TFA	TFA-C64
Urban / ground
Pitts30k	92.37 r	93.21 r	82.12	84.93	81.84	86.01	84.33 $\pm$ 1.59
Pitts250k	95.08 r	95.95 r	82.97	82.96	84.73	87.01	83.70 $\pm$ 1.63
MSLS-val	92.03 r	93.78 r	52.70	41.49	44.41	70.09	71.16 $\pm$ 2.55
St-Lucia	100.00 r	99.93 r	92.31	88.11	85.31	97.61	95.72 $\pm$ 2.87

Table 1: Ground retrieval, R@1 (%). Frozen DINOv2-B/14,

322^{2}

. TFA: DB-only; C64: 64-query calibration (mean

\pm

SD). Protocol: Sec. 4.1 ; dataset sources: Table 5 .

UAV-VisLoc
	Trained Reference			Training Free Systems
	Sample4Geo			AnyLoc			TF-VPR			RIA d			TFA
Backbone	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10	R@1	R@5	R@10
DINOv2-B	10.44 r	21.33	27.16	12.98	27.61	34.89	6.02	14.96	20.20	13.16	26.27	32.64	15.63	34.02	41.59
DINOv3-B	10.44 r	21.33	27.16	14.38	30.31	38.25	10.52	23.34	29.54	12.76	23.49	29.29	19.94	37.80	45.17
VPAIR-full

Table 2: Aerial / cross-view retrieval, R@1/5/10 (%). TFA: DB-only. Bold/underline rank training-free heads within each backbone. Calibration variants: Table 9 ; dataset sources: Table 5 .

Figure 2: Selected Top-1 retrievals across five scene families. DINOv2-B, seed 0, DB-only TFA; green/red denote correct/incorrect matches. Each selected query is recovered by TFA but missed by all three baselines. RIA uses our reconstruction. Aggregate results: Tables 1 and 2 .

Supervised References
System	Backbone/interface	Pitts30k	SPED	St-Lucia	SVOX-5	VPAIR-full
BoQ p	native	93.7	92.5	100	98.38	29.3
SALAD p	DINOv2-B	92.4	92.1	100	97.66	22.1
SelaVPR++	DINOv2-L (perf.)	94.4 p	92.75 r	100.00 r	98.54 r	42.39 r
Training Free Systems
TF-VPR p	DINOv2-B	84.3	77.6	90.8	58.54	67.5

Table 3: Native-system comparison, R@1 (%). TFA uses the dual-vocabulary DINOv2-G L31 value configuration at

640\times 480

. TFA-C

N

N

disjoint calibration images (mean

\pm

SD); TFA-FullQ (Cross-fit): two-fold query cross-fitting.

(a) Fixed transformations have map-dependent effects
	Spectral controls (TFA-FullQ assignment)				Complete heads
Map	Original	Partial	Full	Adaptive	TFA	TFA-C64	TFA-FullQ
	cosine	whitening	whitening	spectrum
Pitts30k	83.97	86.00	85.00	85.64	86.01	84.33 $\pm$ 1.59	86.00
SPED	76.00	83.31	85.34	85.01	85.12	83.79 $\pm$ 0.63	85.01
AmsterTime	45.71	52.69	0.76	48.87	44.84	52.00 $\pm$ 1.76	52.72

Table 4: Spectral controls and complete heads, R@1 (%). Controls fix the TFA-FullQ assignment; adaptive spectrum is

S_{A}

before selection and CLS. C64: mean

\pm

SD. Bold/underline: best/second-best.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset / protocols	Original source and benchmark attribution
Pitts30k / Pitts250k	Pittsburgh imagery ( Torii et al., 2013 ) ; retrieval splits ( Arandjelović et al., 2016 ) .
MSLS-val	Mapillary Street-Level Sequences ( Warburg et al., 2020 ) .
St-Lucia	Milford & Wyeth (2008) ; curated retrieval release ( Berton et al., 2022 ) .
Eynsham	Cummins & Newman (2009) ; curated retrieval release ( Berton et al., 2022 ) .
GSV-Brussels	Brussels subset of GSV-Cities ( Ali-bey et al., 2022 ) .
Essex3in1	Zaffar et al. (2021a) ; VPR-Bench release ( Zaffar et al., 2021b ) .

Appendix

Table 5: Dataset sources. Multiple maps, conditions, heights, and retrieval directions share their parent dataset reference.

System	Setup (s)	Retrieval (s)	Amortized (ms)
TFA	78.38 $\pm$ 1.78	7.01 $\pm$ 0.17	0.346 $\pm$ 0.008
TFA-C64	138.52 $\pm$ 2.83	27.22 $\pm$ 1.02	1.344 $\pm$ 0.050

Appendix

Table 6: Pitts30k deployment costs on H100, mean

\pm

sample SD over three repetitions. Retrieval covers 6,752 queries and three codebook replicas. Amortized time is per query per replica, rather than batch-one latency.

Dataset	Backbone	AnyLoc	TF-VPR	RIA d	TFA	TFA-C64	TFA-FullQ
MSLS-val	DINOv2-B	52.70	41.49	44.41	70.09	71.16 $\pm$ 2.55	75.86
MSLS-val	DINOv3-B	47.39	40.54	39.77	72.21	63.17 $\pm$ 5.24	71.53
SPED	DINOv2-B	75.57	73.48	69.85	85.12	83.79 $\pm$ 0.63	85.01
SPED	DINOv3-B	70.95	76.94	69.85	87.92	85.75 $\pm$ 0.84	86.55
Nordland	DINOv2-B	29.50	29.46	21.03	39.60	30.23 $\pm$ 3.27	35.76
Nordland	DINOv3-B	25.15	30.81	19.07	40.17	31.51 $\pm$ 4.90	30.98

Appendix

Table 7: Ground-backbone comparison, R@1 (%). TFA uses DB-only calibration; C64 uses disjoint calibration (mean

\pm

SD). TFA-FullQ is the transductive reference.

Dataset	TFA-C64	TFA-FullQ	Dataset	TFA-C64	TFA-FullQ
Pitts30k	84.33 $\pm$ 1.59	86.00	CrossSeason	99.97 $\pm$ 0.08	100.00
Pitts250k	83.70 $\pm$ 1.63	87.51	4S Business Campus	95.55 $\pm$ 0.65	94.64
MSLS-val	71.16 $\pm$ 2.55	75.86	4S City Loop	82.88 $\pm$ 0.77	83.94
St-Lucia	95.72 $\pm$ 2.87	97.95	4S Countryside	36.76 $\pm$ 5.75	41.85
Eynsham	85.19 $\pm$ 3.92	88.45	4S Office Loop	90.20 $\pm$ 0.75	90.54
GSV-Brussels	84.75 $\pm$ 0.90	84.36	4S Old Town	70.20 $\pm$ 1.87	70.71

Appendix

Table 8: Ground query-calibrated variants, R@1 (%). DINOv2-B/14,

322^{2}

. C64: mean

\pm

SD; DB-only results appear in Table 1 .

		TFA-C64			TFA-FullQ
Dataset	Backbone	R@1	R@5	R@10	R@1	R@5	R@10
UAV-VisLoc	D2-B	15.05 $\pm$ 2.34	31.54 $\pm$ 3.89	38.92 $\pm$ 4.00	18.50	37.46	44.94
	D3-B	18.33 $\pm$ 2.22	36.43 $\pm$ 3.55	43.90 $\pm$ 3.59	20.61	40.42	47.96
VPAIR-full	D2-B	38.52 $\pm$ 2.93	57.56 $\pm$ 3.48	65.71 $\pm$ 3.53	39.71	60.05	68.63
	D3-B	38.71 $\pm$ 1.73	57.42 $\pm$ 1.99	65.69 $\pm$ 1.98	40.32	60.52	69.30
SUES-200	D2-B	58.65 $\pm$ 2.00	82.58 $\pm$ 1.26	89.72 $\pm$ 0.87	60.89	83.77	90.45

Appendix

Table 9: Aerial query-calibrated variants, R@1/5/10 (%). C64: disjoint 64-query calibration (mean

\pm

SD); TFA-FullQ: full-query calibration. DB-only results appear in Table 2 .

Figure 3: Disjoint calibration-size experiment, R@1. Blue: DINOv2-G dual-vocabulary TFA-C

N

system, mean

\pm

SD over ten splits. The separated TFA-FullQ (Cross-fit) point uses opposite-fold query statistics (Table 3 ); its horizontal position is schematic. All points use the same backbone and dual-vocabulary configuration; calibration and evaluation sets follow the stated protocols.

Figure 4: Additional Top-1 retrievals where DB-only TFA succeeds and at least one baseline fails (DINOv2-B, seed 0). Green/red indicate correctness under the dataset positives; RIA uses our same-feature reconstruction. Two query-index-tertile examples per dataset exclude the main-figure queries. Images retain their full field of view.

Figure 5: Complementary failures: DB-only TFA misses a match recovered by at least one baseline. The protocol and sampling rule follow Figure 4 , with two examples per dataset. Correctness follows the dataset positives, not visual similarity alone.

Visual Place Recognition

Visual Place Recognition matches camera observations to mapped places despite changes in viewpoint, lighting, weather, or time.

See all 84 papers

TRAIL: Trajectory-Aware Visual Place Recognition against Unordered Databases
Sep 7, 2026Dominik A. Kloepfer, Patrick WenzelVisual Place Recognition
Visual Geometry Group (VGG), University of Oxford · Helsing
Are Visual Place Recognition Models Recognizing Places or Conditions? Distractor-Augmented Evaluation and Condition Suppression
Aug 7, 2026Beomsu Kim, Minwoo Jung, Giseop KimVisual Place Recognition Retrieval Robustness
Department of Robotics and Mechatronics Engineering, DGIST, Daegu, Republic of Korea · Department of Mechanical Engineering, Seoul National University, Seoul, Republic of Korea
D $^{2}$ -VPR: A Parameter-efficient Visual-foundation-model-based Visual Place Recognition Method via Knowledge Distillation and Deformable Aggregation
Nov 16, 2025Zheyuan Zhang, Jiwei Zhang, Boyu Zhou +2Visual Place Recognition Vision Foundation Model Adaptation
School of Computer Science, Beijing University of Posts and Telecommunications · Key Laboratory of Interactive Technology and Experience System, Ministry of Culture and Tourism, Beijing University of Posts and Telecommunications

Sep 7, 2026J/K move · Enter open · S save

TRAIL: Trajectory-Aware Visual Place Recognition against Unordered Databases

Dominik A. Kloepfer, Patrick Wenzel

Visual Geometry Group (VGG), University of Oxford · Helsing

Visual Place Recognition

Open paper AskPDF arXiv

Modern Visual Place Recognition (VPR) methods excel on standard benchmarks yet remain brittle in feature-poor environments. By treating each query image in isolation, they discard the sequential context in any real trajectory. We formalize a task that exploits this context: given a query sequence, localize the final image against an unordered reference database -- which, unlike sequence-to-sequence methods, requires no sequential structure in the database. We propose TRAIL (TRajectory-Aware Image Localization), a principled framework based on Conditional Random Fields (CRF) that combines learned functions for visual similarity and for camera-motion consistency, refining a distribution over candidate references as each query arrives. A lightweight post-processing layer atop any pre-trained VPR backbone, TRAIL improves a state-of-the-art baseline by up to 8.3 percentage points on our primary benchmark, transfers to unseen datasets without retraining, and delivers its largest gains where visual cues are scarce.

Abstract

Visual Place Recognition

Explore similar work

TRAIL: Trajectory-Aware Visual Place Recognition against Unordered Databases

Are Visual Place Recognition Models Recognizing Places or Conditions? Distractor-Augmented Evaluation and Condition Suppression

D2^{2}2-VPR: A Parameter-efficient Visual-foundation-model-based Visual Place Recognition Method via Knowledge Distillation and Deformable Aggregation

D $^{2}$ -VPR: A Parameter-efficient Visual-foundation-model-based Visual Place Recognition Method via Knowledge Distillation and Deformable Aggregation