Organizations: State Key Laboratory of AI Safety, Institute of Computing Technology, Chinese Academy of Sciences · University of Chinese Academy of Sciences, Beijing, China
Visual place recognition (VPR) faces critical challenges in handling extreme environmental variations while meeting the computational constraints of practical applications. Current methods predominantly address these challenges by either scaling up model capacity or employing computationally intensive reranking stages, creating a significant efficiency bottleneck. To overcome this limitation, we propose EfficientVPR, a lightweight one-stage framework that achieves speed-accuracy trade-offs through two key innovations: i) a scene-aware visual prompt tuning method which adapts pretrained features with less parameters while dynamically adjusting to sample-specific characteristics, and ii) an instance-dependent key local feature enhancement module that further reinforces discriminative regions. Comprehensive evaluations on Pitts250k, MSLS, Eynsham, AmsterTime and SVOX demonstrate that our method establishes a new SOTA for DINOv2-small models by outperforming all same-scale competitors, and delivers a 73× speedup with 60% lower-dimensional features while maintaining competitive (within 2.5% average R@1 gap) against the SOTA DINOv2-large-based two-stage method. Our code will be available at https://github.com/WiniTang/EfficientVPR.
Figures & tables
Figure 1: Comparison with SOTA methods : average R@1 over 7 datasets and descriptor dimensionality. Our method achieves the best performance with lower feature dimensions compared to approaches with similarly-scale.
Figure 2: The overall architecture of our one-stage method . To maintain strong discriminative power while ensuring lightweight efficiency, we employ DINOv2-small as the backbone with SceneVPT for fine-tuning, preserving generalizable features while adapting to instance-specific characteristics. During feature enhancement, EfficientVPR uses MsIA and OLE to extract multi-scale global features and orientation-aware local features respectively, while SGFS performs instance-specific local feature enhancement of key regions based on sample-dependent semantic cues obtained from backbone.
Figure 3: SceneVPT Overview . Our SceneVPT extends VPT-deep via adaptive prompt selection, dynamically adjusting to input characteristics based solely on the backbone’s CLS token.
Figure 4: Visualization of viewpoint-induced matching failure. The first two images depict the same location, while the third shows a different but similar place. Dramatic viewpoint shifts and seasonal appearance variations cause trees in the query image (red border in the first picture) to become non-overlapping regions absent in the correct reference match. Current one-stage methods tend to be misled by these irrelevant features, producing false matches to incorrect locations with visually similar non-overlapping elements (red border in the third picture). By enhancing the discriminative representation of task-relevant and sample-specific key local regions and structural information, our method effectively mitigates this issue.
Method
Backbone
Dim (k)
Pitts250k -test
MSLS -val
Amster Time
Eynsham
SVOX Night
SVOX Overcast
SVOX Snow
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
NetVLAD [ 7 ]
VGG-16
4
90.5
96.2
53.1
66.5
16.3
-
77.7
-
8.0
-
66.4
-
54.4
-
SFRS [ 18 ]
VGG-16
4
90.7
96.4
69.2
80.3
29.7
48.5
72.3
-
28.6
40.6
81.1
-
76.0
-
CosPlace [ 9 ]
ResNet-50
2
92.3
97.2
87.4
93.8
47.7
-
90.0
-
50.7
-
92.2
-
92.0
-
MixVPR [ 3 ]
ResNet-50
4
94.6
98.3
88.2
93.1
40.8
58.9
89.6
93.2
62.9
79.8
96.2
98.2
97.0
98.3
EigenPlaces [ 11 ]
ResNet-50
2
94.1
98.0
89.1
93.8
48.9
69.5
90.7
94.4
58.9
76.9
93.1
-
93.1
-
Table 1: Comparison with SOTA methods on seven benchmarks. The best results are highlighted in bold and the second best are underlined . † Since the paper of SALAD, BoQ and FoL did not report their results on DINOv2-S, we additionally reproduced these results. ‡ CricaVPR utilizes a cross-image encoder architecture that exhibits instability with varying inference batch sizes. Consequently, we additionally provides results under the single query image scenario.
Method
Backbone
Dim
Train Time (min)
Latency per Query (ms)
Parameters(M)
Memory (GB)
Peak GPU Memory (MiB)
Avg. Acc.
Extrac.
Rerank.
Total
Trainable
Total
FoL †
DINOv2-L
5×105
74.7
102.3
124.8
227.1
56.9
308.8
33.0
10944.0
92.8
SALAD †
DINOv2-B
8448
28.6
6.9
0
6.9
29.6
88.0
0.6
1762
90.0
BoQ †
DINOv2-B
12288
12.4
7.6
0
7.6
28.3
100
0.9
2516
91.6
SALAD
DINOv2-S
8448
14.1
2.4
0
2.4
11.9
22.9
0.6
1474
87.9
BoQ
DINOv2-S
12288
7.2
4.2
0
4.2
15.9
34.4
0.9
2112
89.3
Table 2: Comprehensive comparison with SOTA methods across model scales . † denotes the original version of the method reported by the authors. ”Train Time” denotes train time per epoch. Memory footprint (Memory) is calculated on the MSLS-val dataset, which includes 18871 database images. Latency is measured on Pitts250k-test, using the same CPU and GPU (RTX A800) and is averaged over 5 identical runs. ”Avg. Acc.” denotes the average R@1 across seven datasets.
Figure 5: Qualitative comparison on challenging cases . Green boxes: correct matches; red boxes: errors. Drastic viewpoint changes along with domain shift (1th row), renovation-induced visual changes (2nd row), extreme lighting variations (3rd row), and severe occlusion coupled with time shift (4th row).
Version
#Params. (M)
Pitts250k -test
Amster Time
SVOX Overcast
R@1
R@5
R@1
R@5
R@1
R@5
Frozen
0
89.8
96.3
43.1
64.8
89.1
96.8
Full-tuning
21.7
93.7
98.0
36.8
56.9
95.3
98.4
Adapter [ 22 ]
1.8
94.1
98.6
47.7
69.4
96.4
98.5
LoRA [ 23 ]
0.06
94.0
98.1
47.0
67.4
95.5
98.5
IAPT [ 56 ]
1.38
92.7
97.5
46.8
67.9
94.2
98.2
Table 3: Ablation on SceneVPT . Given that SGFS relies on prompts peocessed by the backbone, it is replaced with a linear layer in all experiments of this set. All test images are resized to 266×266 .
Version
Pitts250k -test
Amster Time
SVOX Overcast
R@1
R@5
R@1
R@5
R@1
R@5
Random
94.4
98.2
52.1
74.1
97.1
99.0
top-k
94.7
98.3
54.5
75.8
97.8
99.1
adaptive selection
95.0
98.4
55.2
77.0
97.9
99.2
Table 4: Ablation on token selection strategy . Random: randomly select prompts. Top-k: preserve only the top-k prompts with the highest attention scores (k=4, optimized through multiple trials). In contrast, our approach dynamically weights all tokens based on their attention values, effectively preserving more comprehensive information. Results demonstrate the superiority of our adaptive method across all datasets. All test images are resized to 266×266 .
Version
Pitts250k -test
Amster Time
SVOX Overcast
R@1
R@5
R@1
R@5
R@1
R@5
MS-fc
94.2
98.1
47.1
69.2
96.1
98.6
MS+OLE-fc
94.4
98.2
48.0
69.2
96.3
98.7
MsIA+OLE-fc
94.5
98.2
48.4
69.5
96.7
99.0
MsIA+OLE-SGFS
95.0
98.4
55.2
77.0
97.9
99.2
Table 5: Ablation on instance-dependent key local feature enhancement module . We analyzed the role of each sub-module in the instance-dependent key local feature enhancement module by removing or replacing key components. All test images are resized to 266×266 .
Version
Pitts250k -test
Amster Time
SVOX Overcast
Eynsham
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
with BoQ block
94.6
98.2
51.1
72.3
96.3
99.0
91.2
94.9
with SGFS
95.0
98.4
55.2
77.0
97.9
99.2
91.6
95.1
Table 6: Ablation on SGFS. ”with BoQ block” refers to the variant where the SGFS module is replaced with a BoQ block, while ”with SGFS” means our full model that retains SGFS. All test images are resized to 266×266 .
Figure 6: Visualization of fine-tuning methods . The first and third row visualizes features extracted by backbones fine-tuned with different methods. The second and fourth row shows their retrieval results (green boxes: correct matches; red boxes: errors). Frozen backbones attend to excessive task-irrelevant information. VPT-shallow improves but still has this issue. In the first group, street features are non-discriminative, whereas in the second group, they become discriminative features. Our SceneVPT enables dynamic adjustment of attention regions based on input images, while both VPT-deep and adapter-based methods fail to achieve this adaptability effectively.
Size
Pitts250k -test
Amster Time
SVOX Overcast
R@1
R@5
R@1
R@5
R@1
R@5
224×224
94.4
98.3
52.2
74.6
96.7
98.9
266×266
95.0
98.4
55.2
77.0
97.9
99.2
322×322
95.4
98.8
60.2
78.4
98.4
99.4
504×504
95.8
98.7
61.4
82.0
98.5
99.4
Table 7: Ablation on inference image size . While keeping the training image size fixed at 266×266 , we test with images resized to varying sizes (in pixel).
Figure 7: Parameters and detailed timing breakdown by sub-module . Results are based on the average of 5 identical runs, measured on Pitts250k using the same CPU and GPU (RTX A800). Latency ( 0.1 ms) and Parameters (Millions).
Size
Pitts250k -test
Amster Time
SVOX Overcast
R@1
R@5
R@1
R@5
R@1
R@5
224×224
94.8
98.5
57.8
78.8
97.7
98.9
266×266
95.4
98.8
60.2
78.4
98.4
99.4
280×280
95.2
98.6
61.4
82.0
98.2
99.2
Table 8: Ablation on training image size . While training with varying sizes, we test with images fixed at 322×322 (in pixel).
Figure 8: Qualitative comparison on challenging cases under various viewpoint changes . Green boxes: correct matches. Red boxes: errors. Viewpoint changes with rotation (1th row), viewpoint changes coupled with dynamic occlusions in suburban(2nd row), viewpoint changes coupled with dynamic occlusions in urban (3rd row). Under significant viewpoint shifts and temporal occlusions, competing methods consistently return images with transient elements (e.g., trees) or non-discriminative structures (incorrect matches). Unaffected by these factors, our method consistently retrieving the correct reference image from the same location.
Method
Back -bone
Pitts250k -test
MSLS -val
Amster Time
Eynsham
SVOX Night
SVOX Overcast
SVOX Snow
Avg. Acc.
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
SALAD
B
95.0
98.5
91.5
96.4
55.5
78.2
91.6
95.0
95.5
99.1
98.4
99.3
98.5
99.5
89.4
95.1
BoQ
B
94.5
98.0
92.4
95.8
55.2
74.2
91.4
94.9
96.6
98.8
98.4
99.7
99.0
99.7
89.6
94.4
SALAD
S
94.0
98.2
89.6
95.4
54.1
75.2
90.6
94.4
88.7
96.2
97.7
98.6
98.4
99.1
87.6
93.9
BoQ
S
94.6
98.3
91.2
95.8
52.2
73.5
90.9
94.5
93.2
97.6
97.9
99.2
99.0
99.6
88.4
94.1
EfficientVPR (Ours)
S
95.4
98.8
92.8
96.6
60.2
78.4
92.0
95.3
94.4
97.9
98.4
99.4
99.1
99.7
90.3
95.2
Table 9: Comparison with SOTA methods across different model scales with 3456 feature dimensions on seven benchmarks. B: DINOv2-base, S: DINOv2-small. The best results are highlighted in bold and the second best are underlined .
Method
Back -bone
Pitts250k -test
MSLS -val
Amster Time
Eynsham
SVOX Night
SVOX Overcast
SVOX Snow
Avg. Acc.
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
R@1
R@5
SALAD
B
94.8
98.4
91.4
96.4
55.9
77.2
91.5
95.1
95.0
98.2
98.2
99.1
98.5
99.4
89.3
94.8
BoQ
B
94.3
97.8
91.8
95.5
51.1
69.6
91.2
94.8
95.7
98.4
97.8
99.1
98.7
99.7
88.7
93.6
SALAD
S
93.9
98.2
89.5
95.0
53.5
74.5
90.6
94.4
87.5
96.0
97.6
98.6
98.3
99.2
87.3
93.7
BoQ
S
93.0
97.8
90.4
95.1
48.8
69.0
90.7
94.4
92.0
96.5
97.5
98.6
98.6
99.2
87.3
92.9
EfficientVPR (Ours)
S
94.8
98.5
92.2
96.4
58.5
77.6
91.7
95.2
93.4
97.8
98.2
99.2
98.7
99.7
89.6
94.9
Table 10: Comparison with SOTA methods across different model scales with 2048 feature dimensions on seven benchmarks. B: DINOv2-base, S: DINOv2-small. The best results are highlighted in bold and the second best are underlined .
Figure 9: Qualitative comparison on challenging cases under domain shift . Green boxes: correct matches. Red boxes: errors. Drastic viewpoint changes (1th row), viewpoint changes coupled with time shift (2nd row), viewpoint changes coupled with dynamic occlusion in urban (3rd row), viewpoint changes in natural (4th row). In the first and third row, our method successfully retrieves the correct image containing the main building with non-overlapping area, while all other methods fail. In the second and fourth row, under conditions of extreme illumination change and color loss, our method correctly handles the viewpoint variations and returns accurate matches, while all other methods fail.
Num.
#Params.
Pitts250k -test
Amster Time
SVOX Overcast
R@1
R@5
R@1
R@5
R@1
R@5
8
0.04
94.0
98.2
53.8
75.8
97.5
99.3
16
0.07
94.7
98.5
54.5
75.5
97.6
99.2
32
0.15
95.0
98.4
55.2
77.0
97.9
99.2
64
0.29
94.9
98.5
54.6
75.7
97.9
99.0
Table 11: Ablation on the number of prompts . We adjust the number of prompts used in SceneVPT and record the corresponding fine-tuning parameter count along with R@1 and R@5 over Pitts250k-test, AmsterTime and SVOX-overcast. To ensure consistency with the ablation studies in the main paper, all test images are resized to 266×266 .
Figure 10: Additional visualization of fine-tuning methods . Consistent with the body of the text, the first and third row visualizes features extracted by backbones fine-tuned with different methods. The second and fourth row shows their retrieval results (green boxes: correct matches; red boxes: errors). In both cases, while the sky and clouds occupy most of the area, they lack distinctiveness. The core discriminative features lie in the smaller objects, such as street lamps, signs, and building details. Fine-tuned by SceneVPT, the backbone effectively learns to focus on image-specific discriminative regions, such as the signs and street lamps in the first query and the architectural details in the second, while suppressing interference from irrelevant areas like the clouds. In contrast, with other approaches the backbone is prone to either missing the core discriminative features or paying excessive attention to non-essential parts of the image.
Figure 11: An enlarged version of Fig. 6 .
Version
Pitts250k -test
Amster Time
SVOX Overcast
R@1
R@5
R@1
R@5
R@1
R@5
concat
94.6
98.1
53.2
75.6
97.1
98.7
nonlinear orthogonal fusion mechanism
95.0
98.4
55.2
77.0
97.9
99.2
Table 12: Ablation on OLE. ”nonlinear orthogonal fusion mechanism” refers to our proposed OLE, i.e. the variant that contains nonlinear orthogonal fusion mechanism. ”concat” means using direct feature concatenation in place of nonlinear orthogonal fusion mechanism. All test images are resized to 266×266 .
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
Visual Place Recognition (VPR) determines a query image's geographic location by matching it against geotagged databases. However, existing methods struggle with perceptual aliasing caused by irrelevant regions and inefficient re-ranking due to rigid candidate scheduling. To address these issues, we introduce FoL++, a method combining robust discriminative region modeling with adaptive re-ranking. Specifically, we propose a Reliability Estimation Branch to generate spatial reliability maps that explicitly model occlusion resistance. This representation is further optimized by two spatial alignment losses (SAL and SCEL) to effectively align features and highlight salient regions. For weakly supervised learning without manual annotations, a pseudo-correspondence strategy generates dense local feature supervision directly from aggregation clusters. Our Adaptive Candidate Scheduler dynamically resizes candidate pools based on global similarity. By weighting local matches by reliability and adaptively fusing global and local evidence, FoL++ surpasses traditional independent matching systems. Extensive experiments across seven benchmarks demonstrate that FoL++ achieves state-of-the-art performance with a lightweight memory footprint, improving inference speed by 40% over FoL. Code and models will be released (and merged with FoL) at https://github.com/chenshunpeng/FoL.
Shunpeng Chen, Yukun Song, Changwei Wang +6
School of Artificial Intelligence, Beijing University of Posts and Telecommunications, Beijing 100876, China. · Key Laboratory of Computing Power Network and Information Security, Ministry of Education, Shandong Computer Science Center, Qilu University of Technology. · Spatialtemporal AI. +1
Recent visual place recognition (VPR) studies have increasingly relied on large-scale datasets to train more robust and discriminative models. Although this trend significantly improves recognition performance, it also introduces substantial storage and training costs, especially when new architectures or training strategies need to be repeatedly developed and evaluated. Dataset pruning (DP) provides a promising way to improve data efficiency by retaining only informative training data. However, conventional DP methods mainly follow the sample-wise classification paradigm, which overlooks the relation-dependent training nature of VPR, where supervision is typically formed by image pairs rather than independent images. To address this issue, we propose a place-wise dataset pruning framework tailored for VPR. Instead of pruning individual images, our method treats each place as the basic pruning unit and introduces two complementary novel metrics, i.e., intra-place diversity (IPD) and inter-place similarity (IPS), to evaluate the training value of each place. By jointly considering these two metrics, our method ranks all places and constructs a compact yet informative coreset, thereby allowing the pruned dataset to still support the training of robust and discriminative VPR models. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art DP baselines under different pruning ratios while reducing selection and training costs. Moreover, by pruning a merged dataset roughly 3.5× the size of GSV-Cities to a comparable scale, our coreset maintains highly competitive performance, achieving 94.5% R@1 on MSLS-val and 97.0% R@1 on Nordland with only NetVLAD. Codes will be made publicly available.
Tong Jin, Yunpeng Liu, Shuyu Hu +3
Shenyang Institute of Automation, Chinese Academy of Sciences, Shenyang 110016, China · University of Chinese Academy of Sciences, Beijing 100049, China · Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen 518071, China +1