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

Zheyuan Zhang; Jiwei Zhang; Boyu Zhou; Linzhimeng Duan; Hong Chen

cs.CVNov 16, 2025

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

Authors: Zheyuan Zhang, Jiwei Zhang, Boyu Zhou, Linzhimeng Duan, Hong Chen

Organizations: 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

Abstract

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 $D^{2}$ -VPR, a $D$ istillation- and $D$ eformable-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.

Figures & tables

Figure 1: The comparison of average R@5 against parameter count on Pitts30k, MSLS-val, and SPED (with image size of 224×224) demonstrates that our model achieves competitive performance despite significantly reduced parameter count, striking an effective trade-off.

Figure 2: Two training stages of our VPR model.

Figure 3: Top-down-attention-based deformable aggregator.

Dataset	Description	Number
Dataset	Description	Database	Queries
Pitts30k	urban, panorama	10,000	6,816
Pitts250k	urban, panorama	83952	8280
MSLS-val	urban, suburban	18,871	740
Nordland	natural, seasonal	27,592	27,592
AmsterTime	very long-term	1,231	1,231

Table 1: Summary of the test datasets in experiments.

Method	Pitts30k			MSLS-val			SPED			Pitts250k			AmsterTime			Nordland
Method	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
Cosplace	90.9	95.7	96.7	87.4	94.1	94.9	-	-	-	92.3	97.4	98.4	47.7	69.8	75.8	71.9	83.8	88.1
MixVPR	91.5	95.5	96.3	87.2	93.1	94.3	-	-	-	94.1	98.2	98.9	40.2	59.1	64.6	76.2	86.9	90.3
EigenPlaces	92.5	96.8	97.6	89.1	93.8	95.0	-	-	-	94.1	97.9	98.7	48.9	69.5	76.0	71.2	83.8	88.1
LPS-VPR	91.8	96.0	96.8	89.9	94.2	95.0	84.8	93.9	95.7	94.1	98.0	98.9	-	-	-	-	-	-
$\text{BoQ}^{\dagger}$	92.0	95.6	96.6	86.6	92.3	93.5	82.7	91.3	94.2	94.4	97.9	98.7	42.0	60.5	66.5	78.9	88.5	91.5

Table 2: Comparison with similar-scale SOTA methods on popular benchmarks. The best results are highlighted in bold and the second best are underlined .

\text{BoQ}^{\dagger}

is re-evaluated using the officially provided weights. The input image size is set to 224×224, and the feature dimension is reduced from 16384 to 4096 using PCA for a fairer comparison. - for not reported. Results of Cosplace and MixVPR are reported from EigenPlaces. Here,

D^{2}

-VPR does not use cross-image encoder.

Method	Dim.	Backbone	Image	Param.
Method	Dim.	Backbone	Size	(M)
Cosplace ${}_{{\text{CVPR' 2022}}}$	2048	ResNet50	No Resize	27.70
MixVPR ${}_{{\text{WACV' 2023}}}$	4096	ResNet50	No Resize	10.88
EigenPlaces ${}_{{\text{ICCV' 2023}}}$	2048	ResNet50	No Resize	27.70
LPS-VPR ${}_{{\text{RAL' 2024}}}$	2048	ResNet50	640×480	29.71
BoQ ${}_{{\text{CVPR' 2024}}}$	4096	ResNet50	224×224	23.84

Table 3: Detailed information of the similar-scale SOTA methods in the comparison. Clus-VPR and EPSA-VPR report only the parameters of their aggregators. Therefore, we use their full ResNet50 version and report 23.51M parameters for the backbone. Note that MixVPR and BoQ utilize a cropped ResNet-50 as their backbone, with the parameter count of the backbone network being less than 23.51M.

Method	Pitts30k			MSLS-val			SPED			Pitts250k			AmsterTime			Nordland
Method	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
$\text{CricaVPR}^{*}$	94.9	97.3	98.2	90.0	95.4	96.4	91.9	95.7	96.7	97.5	99.4	99.7	64.7	82.8	87.5	90.7	96.3	97.6
SALAD	91.6	95.7	97.1	90.5	95.5	96.2	90.8	95.2	96.9	94.6	98.1	98.9	53.4	75.0	79.9	81.2	91.1	94.0
BoQ	93.1	96.5	97.5	91.6	95.9	96.8	91.8	95.6	96.5	95.9	98.8	99.4	57.6	76.9	81.9	85.8	93.6	95.9
$\text{SuperVLAD}^{*}$	94.1	97.3	98.0	90.7	95.9	96.8	90.9	95.6	96.5	96.1	99.0	99.5	60.0	80.3	84.4	88.6	94.7	96.5
Sela-global	90.2	96.1	97.1	87.7	95.8	96.6	84.5	91.8	93.9	92.8	98.0	98.9	41.5	62.1	69.3	72.3	89.4	94.4

Table 4: We compare our method with larger-scale SOTA methods on popular benchmarks. To ensure a fair comparison, the evaluation resolution is set to 224×224. Except for CricaVPR, all other methods are originally evaluated at higher resolutions, so we re-evaluate them using their source code and provided weights. Therefore, the results may differ from those reported in the original papers. Methods with the suffix ‘-global’ correspond to the first stage of the two-stage approach.* for using cross-image encoder.

Method	Dim.	Dinov2	Image	Param.
Method	Dim.	Dinov2	Size	(M)
CricaVPR ${}_{{\text{CVPR' 2024}}}$	4096	Base	224×224	106.76
SALAD ${}_{{\text{CVPR' 2024}}}$	8448	Base	224×224	87.99
BoQ ${}_{{\text{CVPR' 2024}}}$	12288	Base	224×224	95.21
SuperVLAD ${}_{{\text{NIPS' 2024}}}$	3072	Base	224×224	97.61
Sela-global ${}_{{\text{ICLR' 2024}}}$	1024	Large	224×224	357.43

Table 5: Detailed information of the larger-scale SOTA methods in the comparison.

Figure 4: Qualitative VPR comparison results. Our method demonstrates competitive performance compared to these DINOv2-based SOTA models under these challenging cases: long-term appearance changes (first row), drastic lighting variations (second row), perceptual aliasing (third row), and viewpoint changes (fourth row). Green indicates the right match while red is for the wrong one. Key matching regions are highlighted with red dashed boxes.

Figure 5: Method comparison of inference and computational speed on AmsterTime. The inference time includes the inference of both the database and the query set. The batch size is 32. SelaVPR performs calculations in a two-stage manner. PCA is not used here.

Configurations	MSLS-val
Configurations	R@1	R@5	R@10
Baseline	85.1	93.5	95.3
$+\text{distillation}$	88.4	95.0	96.4
$+\text{finetune}$	91.6	96.1	97.2
Baseline	89.9	95.0	95.7
$+\text{DRM}$	90.5	95.5	96.2

Table 6: Ablation study on the MSLS-val benchmark.

Figure 6: Visualizations of attention maps from the knowledge distillation ablation study are presented below, with rows corresponding to the student model (Ours), teacher model (CricaVPR), and a baseline without distillation. The non-distilled baseline exhibits the strongest focus on irrelevant regions (e.g., ground and sky), followed by the student model, while the teacher model best suppresses such spurious attention.

Batchsize	MSLS-val
Batchsize	R@1	R@5	R@10
1	90.9	95.8	96.6
16	90.5	96.5	96.9
32	91.1	96.5	96.8
64	90.9	96.2	96.8
128	91.6	96.1	97.2

Table 7: Influence of bathsize on

D^{2}

-VPR.

Method	Dinov2	Pitts30k		MSLS-val		SPED		Pitts250k		ArmsterTime		Nordland
Method	Dinov2	R@1	R@5	R@1	R@5	R@1	R@5	R@1	R@5	R@1	R@5	R@1	R@5
CricaVPR ${}_{{\text{encoder, batchszie=1}}}$	Base	91.8	95.8	89.1	95.1	87.6	93.1	94.6	98.3	49.0	69.3	80.5	90.7
$D^{2}$ -VPR ${}_{{\text{encoder, batchszie=1}}}$	Small	91.8	96.2	91.1	96.1	83.2	91.4	94.9	98.4	48.7	69.6	77.6	88.5
SuperVLAD ${}_{{\text{no encoder}}}$	Base	92.6	96.2	90.7	95.5	89.1	94.4	95.2	98.4	50.6	73.4	74.6	85.7
$D^{2}$ -VPR ${}_{{\text{no encoder}}}$	Small	91.6	95.8	90.9	95.5	86.2	93.2	94.6	98.4	49.1	70.3	78.0	89.1

Table 8: Performance comparison with methods that use cross-image encoders (CricaVPR and SuperVLAD) when the cross-image encoder is not employed. Here,

D^{2}

-VPR and CricaVPR do not use PCA.

Method	Dim.	Pitts30k		MSLS-val		SPED		Pitts250k		Amstertime		Nordland
Method	Dim.	R@1	R@5	R@1	R@5	R@1	R@5	R@1	R@5	R@1	R@5	R@1	R@5
$D^{2}$ -VPR ${}_{{\text{encoder}}}$	10752	94.9	97.3	91.6	96.5	91.3	96.2	97.8	99.4	62.9	80.7	87.2	94.5
$D^{2}$ -VPR ${}_{{\text{encoder}}}$ +PCA	4096	94.9	97.3	91.6	96.1	90.9	96.0	97.8	99.4	62.9	80.8	86.6	94.1
BoQ	16384	92.1	95.8	87.8	92.3	84.2	93.6	94.7	98.2	43.5	62.3	79.5	88.9
BoQ +PCA	4096	92.0	95.6	86.6	92.3	82.7	91.3	94.4	97.9	42.0	60.5	78.9	88.5

Table 9: Comparison of feature dimensions vs. performance (R@1 and R@5) across six datasets.

Component	Param. (M)
Adapter	2.30
Backbone	22.16
Distill Recovery Module	1.17
Deformable Aggregator	1.58
Cross-image Encode	11.03
All	38.24

Table 10: Analysis of model parameter quantity.

Figure 7: Qualitative attention map of our model. The images are sourced from Pitts30k, Pitts250k, MSLS, Nordland, SPED, and AmsterTime. Their content ranges from natural scenes to urban environments, and they include both black-and-white and color images.

Visual Place Recognition

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

See all 84 papers

Region Matters: Efficient and Reliable Region-Aware Visual Place Recognition
Apr 24, 2026Shunpeng Chen, Yukun Song, Changwei Wang +6Visual Representation Learning Geospatial Representation Learning
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
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
Selectivity Drives Efficiency: Dataset Pruning for Visual Place Recognition
Jul 16, 2026Tong Jin, Yunpeng Liu, Shuyu Hu +3Dataset Pruning Visual Place Recognition
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

Apr 24, 2026J/K move · Enter open · S save

Region Matters: Efficient and Reliable Region-Aware Visual Place Recognition

Shunpeng Chen, Yukun Song, Changwei Wang, Rongtao Xu+5

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

Visual Representation Learning Geospatial Representation Learning Visual Place Recognition

Open paper AskPDF arXiv

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.

Abstract

Visual Place Recognition

Explore similar work

Region Matters: Efficient and Reliable Region-Aware Visual Place Recognition

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

Selectivity Drives Efficiency: Dataset Pruning for Visual Place Recognition