VPR-Space: A Benchmark for Visual Place Recognition in Space Analog Domains

Richard Fechner; Peer Neubert

cs.ROOct 26, 2026

VPR-Space: A Benchmark for Visual Place Recognition in Space Analog Domains

Authors: Richard Fechner, Peer Neubert

Organizations: Institute for Computational Visualistics, University of Koblenz

Abstract

Visual Place Recognition (VPR) is an important means for loop closure detection in visual Simultaneous Localization and Mapping (SLAM). While state-of-the-art VPR and SLAM approaches excel in urban and feature-rich natural environments, the vast and feature-poor environments of extraterrestrial surfaces like on the Moon or Mars pose severe challenges for existing approaches. This paper presents a novel benchmark, VPR-Space, to evaluate VPR capabilities on extraterrestrial surfaces. There are three contributions: (1) We collected sixteen VPR database-query datasets based on existing data from Space Analog Domains (SAD), terrestrial environments that resemble properties of the Moon or Mars. The collection includes a range of challenges regarding environment, viewpoint change, and illumination conditions. (2) To address the ill-posed nature of VPR, we propose a novel approach to generate the ground-truth information in VPR benchmarks. Instead of defining that a query and datbase image show the same place if and only if their camera poses are within fixed position and viewing direction thresholds, we propose to make the decision dependent on whether the match would be valuable in the downstream SLAM task. To avoid the overhead and restrictions of evaluating VPR in-the-loop within SLAM systems, we provide VPR ground-truth (GT) matrices based on the performance of an ensemble of state-of-the-art Relative Pose Estimators (RPE). Under this GT we require a VPR method to identify exactly those image pairs for which a subsequently applied RPE is expected to provide a valuable pose estimate in a SLAM context. (3) We evaluated nine state-of-the-art VPR approaches. The results show considerable room for improvement and indicate directions for future research. We provide a pre-release version of the dataset on github: $\texttt{https://github.com/rfechner/VPR-Space}$

Figures & tables

Fig. 1 : Image descriptor similarity histograms for Oxford RobotCar day–night [ 25 ] (left) and SPICE-HL³ noon–night [ 28 ] (right). We used Bag-of-Queries [ 2 ] , a SotA VPR method, to compute the descriptors. The distributions in the space analog domain are even less separable than in this challenging urban domain (i.e., there is no similarity threshold which approximately separates the positive and negative matchings).

Dataset	Grayscale	RGB	LiDAR	IMU/Gyro	GNSS/Pos	hard VC	hard IC	Mars analog	Moon analog
Devon Island [ 9 ]	✗	✓	✗	✗	✓	✓	✗	✓	✗
MADMAX [ 26 ]	✗	✓	✗	✓	✓	✓	✗	✓	✗
S3LI Vulcano [ 12 ]	✗	✓	✓	✗	✓	✓	✗	✓	✗
S3LI Etna [ 13 ]	✓	✗	✓	✓	✓	✓	✗	✓	✗
Erfoud [ 21 ]	✓	✗	✓	✓	✓	✓	✗	✓	✗
ROBEX [ 35 ]	✓	✗	✗	✓	✓	✓	✗	✓	✓

TABLE I : Sensor modalities relevant to VPR and attributes of SLAM datasets in space domains. The properties Viewpoint Change (VC) and Illumination Change (IC) denote the existence of hard viewpoint- and illumination changes within and between two sequences of the dataset respectively.

Fig. 4 : An overview of the database (blue) and query (red) sequences of the VPR-Space benchmark. We show example matches from both sequences to illustrate the shape of the task. For SPICE-HL³, database and query sequence lie on top of each other.

Fig. 5 : GT matrices (hard GT in dark, soft GT in light color) and Precision-Recall Curves for VPR-GTs of different constructions. To enable comparison, the number of positive matches is equalized across constructions. This example illustrates that established GT constructions tend to request harder matches. This pattern repeats across the dataset.

Panel of Figure 6 — Fig. 6 : Examples of hard queries in VPR-Space, illustrating the challenges of space analog domains. A query image (dark blue) has highest similarity to a false positive image (orange) outside its VPR-GT. The best correct image (green) inside the hard ground truth is displayed in the center. The margin between the similarities indicates the severity of the example.

VPR chall.	DB/Q pair	MeLo [ 5 ]	BoQ [ 2 ]	SLD [ 16 ]	AnyL [ 18 ]	DenV [ 34 ]
Viewpoint change	MM A0 – A2	0.67	0.70	0.65	0.68	0.69
	MM A2 – A4	0.36	0.41	0.39	0.29	0.32
	MM C0 – C1	0.51	0.52	0.41	0.33	0.31
	MM E0 – E1	0.33	0.30	0.23	0.26	0.30
	MM F0 – F3	0.79	0.71	0.68	0.63	0.57
	MM G0 – G2	0.57	0.49	0.48	0.50	0.41

TABLE II : AUPRC of SotA VPR methods on all sequence comparisons. Best per row in bold, runner-up underlined.

Case	Method	R@1	R@5	R@10	R@100p
Worst case	MeLo	0.24	0.46	0.53	0.00
	BoQ	0.38	0.59	0.69	0.00
	SLD	0.25	0.56	0.66	0.00
	AnyL	0.31	0.54	0.64	0.001
	DenV	0.06	0.23	0.33	0.00
Avg. $\pm$ std	MeLo	0.59 $\pm$ 0.25	0.75 $\pm$ 0.19	0.81 $\pm$ 0.15	0.07 $\pm$ 0.07

TABLE III : Recall@k and recall@100% precision of SotA VPR methods. Best per row in bold, runner-up underlined

DB/Q pair	DB	Q	w/ match	matches
	# Images		% queries	% of pos.
MM A0 – A2	171	180	92.8	8.55
MM A2 – A4	180	412	20.9	0.74
MM C0 – C1	502	463	78.6	1.89
MM E0 – E1	517	507	43.2	0.62
MM F0 – F3	117	229	89.1	5.14

TABLE IV : Additional dataset statistics. For each database/query pair we report the number of images, the percentage of queries which have at least one match (% queries w/ match) and the percentage of positive matches (% of pos. matches) given the proposed VPR-GT-RPE-0.

	MADMAX		S3LI Etna		LunarLoc		SPICE-HL 3
	soft	hard	soft	hard	soft	hard	soft	hard
$d_{\max}$ [m]	35	35	35	35	35	35	1	1
$\tau_{t}$ [m]	4	3	4	3	4	3	0.4	0.2
$\tau_{R}$ [ ∘ ]	30	20	30	20	30	20	20	10
$n$	2	4	2	4	2	4	1	1

TABLE V : Hyperparameters used for VPR-GT-RPE-0.

Visual Place Recognition

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

See all 84 papers

Breaking Déjà Vu: Independent Auditing of Visual Place Recognition through Vision-Language Reasoning
Jul 14, 2026Sania Waheed, Michael Milford, Sarvapali D. Ramchurn +1Visual Place Recognition Robot Localization
School of Electronics and Computer Science, University of Southampton, Southampton SO17 1BJ, U.K. · School of Electrical Engineering and Computer Science, Queensland University of Technology, Brisbane, QLD 4000, Australia · School of Computer Science and Electronic Engineering, University of Essex, CO4 3SQ Colchester, U.K.
QdaVPR: A novel query-based domain-agnostic model for visual place recognition
Mar 8, 2026Shanshan Wan, Lai Kang, Yingmei Wei +3Visual Place Recognition Domain Generalization
College of Systems Engineering and the Laboratory for Big Data and Decision, National University of Defense Technology, Changsha, Hunan 410073, China
SLAM: Structured and Localized Analytic Manifold Adaptation for Lifelong VPR
Jul 6, 2026Kenta Tsukahara, Kanji Tanaka, Rai HisadaVisual Place Recognition Continual Learning
Department of Mechanical Engineering, Faculty of Engineering, University of Fukui, Fukui 910-8507, Japan

Jul 14, 2026J/K move · Enter open · S save

Breaking Déjà Vu: Independent Auditing of Visual Place Recognition through Vision-Language Reasoning

Sania Waheed, Michael Milford, Sarvapali D. Ramchurn, Shoaib Ehsan

School of Electronics and Computer Science, University of Southampton, Southampton SO17 1BJ, U.K. · School of Electrical Engineering and Computer Science, Queensland University of Technology, Brisbane, QLD 4000, Australia · School of Computer Science and Electronic Engineering, University of Essex, CO4 3SQ Colchester, U.K.

Visual Place Recognition Robot Localization

Open paper AskPDF arXiv

Visual place recognition (VPR) is a key enabler of accurate localization and long-term autonomous navigation in robotics applications, such as loop closure detection for simultaneous localisation and mapping (SLAM). However, real-world VPR deployment relies on selecting an image matching threshold that balances precision and recall. These thresholds are typically tuned using labeled validation data and fixed during deployment, making them unreliable under environmental changes where ground truth is unavailable. This is particularly problematic in safety-critical robotics, where accepting a false loop closure can corrupt the estimated trajectory and map. In this work, we introduce Visual Place Recognition Auditing, an independent post-retrieval verification framework that leverages Vision-Language Models (VLMs) to assess retrieved matches by reasoning jointly over query and candidate images. Unlike conventional verification methods, our approach performs instance-level verification without requiring architecture-specific confidence measures, dataset-dependent thresholds, or prior knowledge of the deployment environment. We evaluate our method on six benchmark datasets using five state-of-the-art VPR methods and four VLMs. Results show that VLM-based auditing improves recall@1 by 13.6% on average as compared to state-of-the-art methods while reducing false acceptance rates to 12%, maintaining precision above 95% and coverage above 75%.

Abstract

Visual Place Recognition

Explore similar work

Breaking Déjà Vu: Independent Auditing of Visual Place Recognition through Vision-Language Reasoning

QdaVPR: A novel query-based domain-agnostic model for visual place recognition

SLAM: Structured and Localized Analytic Manifold Adaptation for Lifelong VPR