Organizations: College of Systems Engineering and the Laboratory for Big Data and Decision, National University of Defense Technology, Changsha, Hunan 410073, China
Visual place recognition (VPR) aiming at predicting the location of an image based solely on its visual features is a fundamental task in robotics and autonomous systems. Domain variation remains one of the main challenges in VPR and is relatively unexplored. Existing VPR models attempt to achieve domain agnosticism either by training on large-scale datasets that inherently contain some domain variations, or by being specifically adapted to particular target domains. In practice, the former lacks explicit domain supervision, while the latter generalizes poorly to unseen domain shifts. This paper proposes a novel query-based domain-agnostic VPR model called QdaVPR. First, a dual-level adversarial learning framework is designed to encourage domain invariance for both the query features forming the global descriptor and the image features from which these query features are derived. Then, a triplet supervision based on query combinations is designed to enhance the discriminative power of the global descriptors. To support the learning process, we augment a large-scale VPR dataset using style transfer methods, generating various synthetic domains with corresponding domain labels as auxiliary supervision. Extensive experiments show that QdaVPR achieves state-of-the-art performance on multiple VPR benchmarks with significant domain variations. Specifically, it attains the best Recall@1 and Recall@10 on nearly all test scenarios: 93.5%/98.6% on Nordland (seasonal changes), 97.5%/99.0% on Tokyo24/7 (day-night transitions), and the highest Recall@1 across almost all weather conditions on the SVOX dataset. Our code will be released at https://github.com/shuimushan/QdaVPR.
Figures & tables
Fig. 1: Overview of the domain-agnostic VPR models. (a) Most models acquire some domain agnosticism through training on large-scale datasets. (b) Some VPR models are specially designed to be robust to target domains but generalize poorly to unseen domains. (c) Our model achieves domain agnosticism via a novel dual-level adversarial learning framework, which makes it more robust to VPR tasks under domain shift.
Fig. 2: Overview of the QdaVPR model. (a) The base architecture. Regardless of whether the input image is from the original GSV-cities or the generated six-domain synthetic GSV-cities datasets, this architecture outputs a global descriptor and Nc query combinations. (b) The framework for dual-level adversarial learning. Each of the L×M query features and the L domain features (extracted from the L image features) is fed into a domain discriminator, producing output logits with six values. These values predict the specific domain of the input image. Note that output logits are generated only when the input image is from the generated six-domain dataset; otherwise, the outputs are as shown in (a). During inference, the red blocks are discarded, and only the global descriptor is output. The channel dimension is omitted from the figure for clarity. See text for details.
Fig. 3: Mutual reinforcement in the dual-level adversarial learning framework. Query features improve image features’ domain agnosticism through negative gradient flow, and image features reciprocally produce domain-agnostic query features via the framework flow. The symbols are defined in the text.
Dataset
Inquiry/Database
Scenery
Domain Shift
Pitts30k-test [ 51 ]
6816/10000
Urban
None.
Tokyo24/7 [ 24 ]
315/75984
Urban
Day/Night.
MSLS-val [ 36 ]
740/18871
Urban, Suburban
Day/Night, season and weather.
AmsterTime [ 52 ]
1231/1231
Urban
Long-term, modalities.
SVOX-night [ 16 ]
823/17166
Urban
Day/Night.
SVOX-overcast [ 16 ]
872/17166
Urban
Weather.
TABLE I: Overview of the test VPR datasets
Method
Image size
Nordland [ 23 ]
Tokyo24/7 [ 24 ]
MSLS-val [ 36 ]
AmsterTime [ 52 ]
Pitts30k-test [ 51 ]
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
SALAD [ 28 ] , CVPR2024
224×224
79.8
90.2
93.4
91.7
96.5
96.5
91.1
95.5
96.4
53.0
74.4
79.5
91.7
95.7
97.0
BoQ [ 21 ] , CVPR2024
83.9
92.4
95.0
94.3
97.5
97.8
90.9
95.7
96.4
53.6
72.5
76.9
93.0
96.4
97.3
EDTFormer [ 35 ] , TCSVT2025
80.9
91.2
94.2
96.5
97.8
98.1
91.9
96.1
96.6
58.9
80.3
84.7
92.8
96.8
97.8
ImAge [ 31 ] , NIPS2025
88.2
95.5
97.1
92.1
96.2
97.8
90.7
95.9
96.6
56.3
76.1
81.6
93.2
96.8
97.8
CerfeVPR [ 18 ] , CMC2025
92.5
96.8
97.8
\
87.7
94.2
95.3
\
94.9
97.5
98.0
TABLE II: R@N comparison to SOTA methods on benchmark datasets. The best results are in bold and the second best underlined .
Method
Nordland [ 23 ]
SVOX [ 16 ]
Summer/Spring
Summer/Fall
Rain
Sun
Snow
Night
Overcast
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
R@1
R@5
R@10
R@1
R@5
R@10
SALAD [ 28 ]
96.2
98.3
98.9
99.4
99.7
99.7
98.6
99.7
99.8
97.3
99.1
99.5
98.6
99.5
99.7
95.5
99.3
99.4
98.4
99.3
99.3
BoQ [ 21 ]
97.3
99.1
99.4
99.5
99.8
99.8
98.5
99.6
99.7
97.4
99.4
99.4
99.3
99.8
99.8
97.2
98.9
99.4
98.3
99.2
99.4
EDTFormer [ 35 ]
97.3
98.9
99.4
99.6
99.8
99.8
98.3
99.8
99.8
98.4
99.5
99.8
98.5
99.7
99.8
95.7
98.8
99.5
98.1
99.2
99.4
ImAge [ 31 ]
97.6
99.2
99.5
99.4
99.7
99.8
98.0
99.6
99.6
97.4
99.4
99.5
98.7
99.5
99.7
97.3
99.1
99.6
97.6
99.1
99.2
TABLE III: R@N comparison to SOTA methods on the challenging benchmark datasets under varied domain conditions. Inference images are resized to 322×322 . The best results are in bold and the second best underlined .
λadvq
λadvx
λlocal
Train set
MSLS-val
Nordland
Tokyo24/7
0
0
0
original
92.7
88.0
97.1
0
0
0
augmented
92.7
89.3
96.8
0
0.05
0
93.0
93.5
97.1
0.01
0.05
0
92.3
92.9
96.2
0.1
0.05
0
92.2
89.0
96.2
0.5
0.05
0
92.6
92.8
96.5
TABLE IV: Ablation on loss weights. ’Original’ and ’Augmented’ refer to the GSV-cities dataset used for training. The best R@1 results are in bold and the second best underlined .
α
G
H
MSLS-val
Nordland
Tokyo24/7
0.05
10
4
92.8
93.3
97.1
0.05
10
8
93.6
93.5
97.5
0.05
10
12
93.0
94.7
96.5
0.05
10
16
92.4
92.9
96.8
0.05
10
20
93.1
92.8
97.1
0.05
10
24
93.1
91.0
96.5
TABLE V: Ablation on hyperparameters of the query-combination-based triplet supervision. The best R@1 results are in bold and the second best underlined .
Method
Dimension
MSLS-val
Nordland
Tokyo24/7
AmsterTime
BoQ [ 21 ]
12288
93.8
90.6
96.5
63.0
4096
93.0
88.8
96.5
59.5
2048
92.2
85.4
95.2
54.8
512
85.8
68.4
88.6
34.9
QdaVPR(ours)
12288
93.6
94.1
97.1
64.0
4096
93.6
93.5
97.5
61.1
TABLE VI: Ablation on descriptor dimensionality with the best R@1 results in bold .
Fig. 4: Average attention map visualization for (b) QdaVPR and (c) BoQ [ 21 ] . In the input images (a), green blocks indicate regions containing buildings. In (b) and (c), green blocks denote consistently high attention to the buildings shown in (a) across different weather conditions, whereas red blocks indicate high attention under only one specific weather condition and low attention under the other.
Visual Place Recognition (VPR) localizes a query image by retrieving database images of the same or nearby place, yet its robustness is often degraded by domain shifts arising from illumination, weather, seasonal changes, and dynamic occlusions. One contributing factor is the limited appearance diversity of the same place in existing training data. To address this issue, we propose AdaptVPR, a route-aware generative augmentation framework that constructs same-place hard positives for robust VPR training. AdaptVPR first uses a vision language model to parse scene attributes and estimate editing feasibility, while a rule-based scheduler determines the generation route according to editability scores and risk constraints. The generation process is decomposed into three complementary routes: the Global Appearance Route introduces global scene changes in weather, illumination, and time of day; the Local Occlusion Route inserts plausible dynamic occluders; and the Dual Route combines both types of perturbations to produce more challenging appearance shifts. Each generated candidate is evaluated using a VPR-oriented verification scheme based on geometric consistency and appearance diversity, reducing the risk of structural drift while ensuring sufficient appearance variation. Global candidates are generated once and rejected if verification fails, while Local Occlusion and Dual candidates use verification feedback for limited prompt refinement and regeneration. Using this framework, we construct AdaptCities, containing 160K verified synthetic same-place hard positives. Experiments across multiple VPR baselines and vision foundation backbones show consistent gains on standard benchmarks and substantial improvements under challenging domain shifts, with R@1 gains of up to 9.2%. The source code and data resources are publicly available at https://github.com/chenshunpeng/AdaptVPR.
Shunpeng Chen, Jingyi Zhang, Changwei Wang +6
School of Artificial Intelligence, Beijing University of Posts and Telecommunications · Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center, Qilu University of Technology
Visual Place Recognition (VPR) aims to determine the geographic location of a query image by retrieving its most visually similar counterpart from a geo-tagged reference database. Recently, the emergence of the powerful visual foundation model, DINOv2, trained in a self-supervised manner on massive datasets, has significantly improved VPR performance. This improvement stems from DINOv2's exceptional feature generalization capabilities but is often accompanied by increased model complexity and computational overhead that impede deployment on resource-constrained devices. To address this challenge, we propose D2-VPR, a Distillation- and Deformable-based framework that retains the strong feature extraction capabilities of visual foundation models while significantly reducing model parameters and achieving a more favorable performance-efficiency trade-off. Specifically, first, we employ a two-stage training strategy that integrates knowledge distillation and fine-tuning. Additionally, we introduce a Distillation Recovery Module (DRM) to better align the feature spaces between the teacher and student models, thereby minimizing knowledge transfer losses to the greatest extent possible. Second, we design a Top-Down-attention-based Deformable Aggregator (TDDA) that leverages global semantic features to dynamically and adaptively adjust the Regions of Interest (ROI) used for aggregation, thereby improving adaptability to irregular structures. Extensive experiments demonstrate that our method achieves competitive performance compared to state-of-the-art approaches. Meanwhile, it reduces the parameter count by approximately 64.2% (compared to CricaVPR).Code is available at https://github.com/tony19980810/D2VPR.
Zheyuan Zhang, Jiwei Zhang, Boyu Zhou +2
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
Long-term Visual Place Recognition (VPR) is typically evaluated by matching queries from one condition against a database from another. Crowdsourced map databases, however, may mix conditions and include images that resemble the query in condition but depict different places. In the presence of these distractors, a method may retrieve by condition similarity rather than place identity. We argue that this susceptibility arises because the discriminability of VPR methods allows them to encode information such as illumination, weather, and seasonal appearance in their descriptors. We therefore introduce Distractor-Augmented Recall (DAR) to isolate and quantify the effect of distractors, and propose condition suppression to remove condition information from VPR descriptors. Across eleven methods and six datasets, method rankings under DAR@1 differ from those under Recall@1 (R@1), while applying INLP and LEACE as condition suppression methods generally improves DAR@1 without reducing R@1. Thus, distractor robustness is distinct from standard retrieval performance and can be improved by suppressing condition information.
Beomsu Kim, Minwoo Jung, Giseop Kim
Department of Robotics and Mechatronics Engineering, DGIST, Daegu, Republic of Korea · Department of Mechanical Engineering, Seoul National University, Seoul, Republic of Korea