Organizations: Department of Precision Instrument, Tsinghua University, Beijing 100084, China · School of Aerospace Engineering, Beijing Institute of Technology, Beijing 100081, China · School of Instrumentation Science and Opto-electronics Engineering, Beijing Information Science and Technology University, Beijing 100192, China · School of Intelligent Systems Engineering, Sun Yat-sen University, Guangzhou 510275, China
In this work, we propose HE-VPR, a visual place recognition (VPR) framework that incorporates height estimation. Our system decouples height inference from place recognition, allowing both modules to share a frozen DINOv2 backbone. Two lightweight bypass adapter branches are integrated into our system. The first estimates the height partition of the query image via retrieval from a compact height database, and the second performs VPR within the corresponding height-specific sub-database. The adaptation design reduces training cost and significantly decreases the search space of the database. We also adopt a center-weighted masking strategy to further enhance the robustness against scale differences. Experiments on two self-collected challenging multi-altitude datasets demonstrate that HE-VPR achieves up to 6.1% Recall@1 improvement over state-of-the-art ViT-based baselines and reduces memory usage by up to 90%. These results indicate that HE-VPR offers a scalable and efficient solution for height-aware aerial VPR, enabling practical deployment in GNSS-denied environments. All the code and datasets for this work have been released on https://github.com/hmf21/HE-VPR.
Figures & tables
Fig. 1 : Overview of the proposed HE-VPR, an aerial VPR pipeline with height estimation.
Fig. 2 : Illustration of the adapter network in ViT.
Fig. 3 : The proposed HE-VPR pipeline. A height estimation branch is added for sub-database selection such that the proposed pipeline is robust to height variance. It requires only a single forward pass of the model for both parts.
Dataset
Query Image
Height Database
VPR Database
Height Range ( m )
GEStudio
1200
102
29768
100 ∼ 1200
MHFlight
1970
850
36400
200 ∼ 640
TABLE I : The statistics of GEStudio and MHFlight.
Fig. 4 : Overviews of two evaluation datasets.
Method
R @ 1
R @ 5
R @ 10
E avg
threshold at 50 m and 100 m ( % / % )
( m )
GEStudio dataset
UniDepth V2
1.25/8.50
N/A
N/A
470.83
DA V2
5.58/15.00
N/A
N/A
550.03
HE-VPR
63.08/91.75
92.33/99.92
98.25/100.00
43.63
TABLE II : Evaluation for the height estimation. We present the performance with height thresholds of 50m and 100m to define a correct retrieval. We compare with two MMDE methods, namely UniDepth V2 and Depth Anything (DA) v2. The best result is highlighted in bold.
Fig. 5 : Qualitative results for height estimation. The top two rows are from GEStudio dataset and the bottom two rows are from MHFlight dataset. The height label for top-5 database image is annotated at the top and all the samples are correctly retrieved at the top-1 candidate.
Dataset
Database
R @ 1
R @ 5
R @ 10
GEStudio
M
36.50/54.25
73.33/91.08
91.58/98.33
G
63.08/91.75
92.33/99.92
98.25/100.00
∗ G ∗
61.58/91.33
90.25/99.75
95.58/99.92
MHFlight
G
29.04/49.75
52.08/67.26
66.29/72.13
M
34.01/60.96
76.95/91.83
89.85/96.80
TABLE III : Height estimation performance with different databases. In this evaluation, G and M denote denote databases sharing similar scene domains (e.g., urban and rural environments) with the GEStudio and MHFlight datasets, respectively. And the superscript ∗ indicates a database that is larger than the original database.
Method
Backbone
Dim
Param.
GEStudio dataset
MHFlight dataset
( MB )
R @ 1
R @ 5
R @ 10
R @ 1
R @ 5
R @ 10
GeM [ 23 ]
ResNet50
1024
8.5
20.08/24.00
35.75/42.33
43.08/51.67
37.31/63.81
61.88/75.69
69.34/78.88
CosPlace [ 31 ]
ResNet50
2048
27.7
47.92/51.50
61.42/66.50
66.17/71.83
55.89/83.35
67.61/84.62
72.99/84.87
MixVPR [ 27 ]
ResNet50
4096
10.9
50.75/53.92
61.08/65.67
66.00/70.92
58.63 /84.47
71.83/84.82
77.26/85.03
CricaVPR [ 6 ]
DINOv2-B
10752
106.8
61.17/65.58
73.25/76.75
76.42/79.67
57.21/83.20
70.00/84.37
76.19/84.47
TABLE IV : VPR comparison of different methods. We only use the VPR adapter branch for evaluation. Here we also present the performance under different thresholds with 100 m and 200 m . The best results are shown in bold.
Method
GEStudio dataset
MHFlight dataset
R @ 1
R @ 5
R @ 10
R @ 1
R @ 5
R @ 10
Adapter(full)
69.50
76.42
79.08
57.61
72.49
77.82
HE-VPR(1)
57.25
66.50
70.00
49.14
68.07
74.06
HE-VPR(5)
69.92
76.17
78.67
56.80
72.94
77.72
HE-VPR(10)
70.42
76.83
79.00
57.41
71.93
77.46
TABLE V : Comparison between the standalone VPR adapter branch and the proposed HE-VPR system.
Method
Full
HE-VPR(1)
HE-VPR(5)
HE-VPR(10)
GEStudio dataset
Memory Usage ( % )
100
12.50
38.71
58.51
Performance Ratio ( % )
100
86.11 -13.89
99.89 -0.11
100.56 +0.56
MHFlight dataset
Memory Usage ( % )
100
7.14
27.18
42.66
TABLE VI : Comparison for memory usage and performance ratio of database.
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
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