Joint Multi-Condition Representation Modelling via Matrix Factorisation for Visual Place Recognition
Authors: Timur Ismagilov, Shakaiba Majeed, Michael Milford, Tan Viet Tuyen Nguyen, Sarvapali D. Ramchurn, Shoaib Ehsan
Organizations: School of Electronics and Computer Science, University of Southampton, SO17 1BJ Southampton, U.K. · Department of Computer Science and Engineering, Hanyang University, Seoul, South Korea. · QUT Centre for Robotics, School of Electrical Engineering and Robotics, Brisbane, QLD 4000, Australia. · School of Computer Science and Electronic Engineering, University of Essex, Colchester, CO4 3SQ, U.K.
We address multi-reference visual place recognition (VPR), where reference sets captured under varying conditions are used to improve localization performance. While deep learning with large-scale training improves robustness, increasing data diversity and model complexity incur extensive computational cost during training and deployment. Descriptor-level fusion via voting or aggregation avoids training, but often targets multi-sensor setups or relies on heuristics with limited gains under appearance and viewpoint change. We propose a training-free, descriptor-agnostic approach that jointly models places using multiple reference descriptors via matrix decomposition into basis representations, enabling projection-based residual matching. We also introduce SotonMV, a benchmark for multi-viewpoint VPR captured over time. Our method improves Recall@1 by up to ≈18% over conventional single-reference set matching on appearance-change datasets, and matches or outperforms existing multi-reference baselines across most appearance-change settings. For viewpoint-change datasets, our method supports matching when viewpoint is distributed among the references, with the benefit depending on reference set coverage of the query conditions. On unconstrained reference sets, we additionally observe gains of ≈5%, demonstrating strong generalization while remaining computationally lightweight.
Figures & tables
Fig. 1: Top : Each place is represented by multiple images with different conditions, whose descriptors are stacked and subspace is formed using matrix factorization, here QR decomposition, to construct the map. Bottom : At inference, the query is projected and reconstructed using each subspace, with the smallest residual determining the match.
Fig. 2: Example places from datasets we evaluate. These datasets cover multi-reference sets with varying viewpoints, illumination, seasons and weather. Notably, the GLDv2-micro dataset contains irregular reference imagery with varying conditions, size and channel compositions.
Fig. 3: (Left) The area considered in Southampton. Each point is a cluster of scenes across multiple dates, with each scene encompassing four images at regularly spaced orientation. (Right) Example places from SotonMV. Each place is observed multiple times (appearance change), and captured at four headings (viewpoint change). The query images are taken at intermediate headings, representing VPR under challenging viewpoint with non-trivial spatial overlap.
Fig. 4: (Left, Middle) Distribution of scenes in SotonMV over years and months. (Right) Number of unique dates per place. Each place is composed of multiple scenes at different dates, where each scene consists of 4 images at different orientations.
Queries →
Fall
Spring
Summer
Winter
Fall
Spring
Summer
Winter
Fall
Spring
Summer
Winter
Fall
Spring
Summer
Winter
↓ References
EigenPlaces (2048D)
CosPlace (2048D)
MixVPR (4096D)
NetVLAD (4096D)
Fall
—
88.9
93.3
62.7
—
85.9
91.2
58.4
—
95.3
95.9
81.6
—
47.2
69.0
17.4
Spring
80.6
—
73.5
67.6
76.4
—
68.3
68.2
88.0
—
83.2
89.6
42.5
—
37.4
21.0
Summer
94.3
82.6
—
59.3
92.8
79.5
—
53.8
95.8
91.6
—
78.4
69.1
42.9
—
18.2
Winter
51.8
70.7
47.1
—
46.4
66.2
41.4
—
70.2
86.9
64.8
—
14.3
22.8
13.9
—
Pooling
96.3
92.9
93.6
70.1
95.1
90.7
92.1
70.5
97.7
98.4
96.4
91.8
72.0
54.6
70.2
23.9
TABLE I: Recall@1 with Nordland retrieval using CNN-based Models.
Queries →
Fall
Spring
Summer
Winter
Fall
Spring
Summer
Winter
Fall
Spring
Summer
Winter
Fall
Spring
Summer
Winter
↓ References
MegaLoc (8448D)
SALAD (8448D)
CricaVPR (10752D)
BoQ (12288D)
Fall
—
97.2
97.1
94.9
—
97.5
97.3
88.9
—
98.6
98.4
96.1
—
97.9
97.2
93.1
Spring
96.6
—
93.8
98.1
95.2
—
93.2
92.3
98.1
—
95.8
98.1
96.4
—
94.8
95.9
Summer
97.0
94.3
—
91.3
97.6
95.1
—
87.0
98.1
96.9
—
94.5
97.4
95.7
—
90.0
Winter
93.1
97.9
90.1
—
86.6
93.9
84.7
—
93.5
97.4
92.7
—
89.8
96.9
88.7
—
Pooling
99.0
99.4
97.9
98.5
99.0
99.0
97.8
94.0
99.4
99.4
98.6
99.0
99.2
99.3
97.7
97.0
TABLE II: Recall@1 with Nordland retrieval using Transformer-based Models.
Queries →
Sun
Dusk
Rain
Overcast
Night
Snow
Sun
Dusk
Rain
Overcast
Night
Snow
Sun
Dusk
Rain
Overcast
Night
Snow
Sun
Dusk
Rain
Overcast
Night
Snow
↓ References
EigenPlaces (2048D)
CosPlace (2048D)
MixVPR (4096D)
NetVLAD (4096D)
Sun
—
61.3
94.4
96.7
44.2
86.7
—
52.4
90.0
93.7
23.0
81.7
—
59.6
96.0
96.8
70.5
89.0
—
22.9
74.5
86.4
14.4
68.2
Dusk
57.9
—
54.6
58.8
35.0
62.0
49.0
—
42.4
49.6
17.7
50.7
51.8
—
49.2
52.9
55.8
57.0
16.3
—
14.6
17.8
16.5
25.6
Rain
95.2
57.7
—
94.7
44.9
84.8
89.7
48.2
—
89.9
25.8
78.5
95.8
55.7
—
95.4
68.2
85.2
78.5
23.4
—
80.4
15.9
57.8
Overcast
96.3
63.6
94.2
—
49.2
89.7
94.0
55.3
90.6
—
24.4
85.4
96.8
61.8
96.0
—
75.4
90.5
89.3
25.4
78.4
—
19.2
69.1
Night
62.4
47.2
60.9
66.6
—
60.5
41.9
32.2
38.4
45.2
—
37.4
69.6
58.2
70.2
74.5
—
69.9
14.8
17.0
11.4
15.0
—
17.7
TABLE III: Recall@1 with Oxford RobotCar retrieval using CNN-based Models.
Queries →
Sun
Dusk
Rain
Overcast
Night
Snow
Sun
Dusk
Rain
Overcast
Night
Snow
Sun
Dusk
Rain
Overcast
Night
Snow
Sun
Dusk
Rain
Overcast
Night
Snow
↓ References
MegaLoc (8448D)
SALAD (8448D)
CricaVPR (10752D)
BoQ (12288D)
Sun
—
69.9
94.2
95.4
77.2
84.6
—
72.4
95.5
96.2
83.0
89.1
—
71.4
96.1
97.2
79.8
89.0
—
77.0
97.4
97.5
88.9
90.5
Dusk
64.7
—
59.8
67.2
68.3
64.6
66.7
—
60.6
68.1
68.9
65.4
65.2
—
61.7
66.0
68.8
66.4
74.4
—
67.3
75.3
74.5
72.2
Rain
93.6
62.6
—
93.0
76.2
81.1
94.8
63.4
—
95.1
81.8
87.1
96.4
65.6
—
96.2
79.5
86.0
97.0
70.5
—
96.8
88.8
88.5
Overcast
95.3
69.0
93.0
—
77.6
86.7
96.2
71.8
94.4
—
84.4
90.8
96.7
71.4
95.0
—
82.6
90.1
97.3
79.0
96.3
—
90.2
92.4
Night
72.5
67.7
73.1
74.0
—
67.9
79.9
70.1
79.1
80.2
—
77.8
79.8
69.9
80.4
83.0
—
75.7
87.4
77.0
86.1
88.3
—
82.9
TABLE IV: Recall@1 with Oxford RobotCar retrieval using Transformer-based Models.
Approach
CosPlace (2048D)
Eigenplaces (2048D)
MixVPR (4096D)
NetVLAD (4096D)
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
Pooling
34.2
50.7
57.3
64.3
70.7
82.0
88.0
91.3
72.0
84.0
88.0
92.3
32.3
55.0
61.7
73.0
dMat. Avg.
20.7
38.7
47.0
60.7
54.7
71.3
79.3
86.3
59.3
77.7
82.0
89.3
26.7
44.0
52.3
65.3
LogSumExp
36.7
54.0
59.0
64.3
75.7
85.7
90.0
91.3
72.7
86.3
90.3
92.3
33.7
56.3
64.3
73.0
Sum Descriptors
32.3
53.0
60.7
71.3
67.3
83.7
88.0
92.3
67.3
82.3
88.3
91.3
28.7
54.7
63.7
74.7
QR-FullVP (Ours)
43.7
61.7
67.0
75.3
80.7
90.7
93.0
94.7
76.3
89.7
93.0
95.7
40.3
62.3
70.0
80.0
TABLE V: Retrieval performance on SotonMV using CNN-based Models.
Approach
MegaLoc (8448D)
SALAD (8448D)
CricaVPR (10752D)
BoQ (12288D)
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
Pooling
78.7
91.3
94.7
96.7
76.3
88.3
93.0
96.7
57.7
74.7
81.3
90.0
85.7
94.3
97.3
98.7
dMat. Avg.
74.0
86.3
91.7
95.3
67.3
82.0
88.3
94.0
32.0
50.3
60.3
71.3
84.0
94.7
96.3
98.3
LogSumExp
78.3
92.0
95.3
96.7
77.7
89.0
92.3
96.7
56.0
73.3
82.0
90.0
87.3
94.7
98.7
98.7
Sum Descriptors
76.0
89.7
93.3
97.3
74.0
88.0
92.7
94.3
42.7
62.3
72.7
82.0
85.0
94.3
97.7
98.7
QR-FullVP (Ours)
81.7
93.7
96.3
97.3
81.3
91.3
95.3
97.7
60.3
78.0
85.3
90.0
87.7
96.7
98.7
99.0
TABLE VI: Retrieval performance on SotonMV using Transformer-based Models.
Approach
CosPlace (2048D)
Eigenplaces (2048D)
MixVPR (4096D)
NetVLAD (4096D)
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
Pooling
74.2
92.2
95.5
98.6
89.0
97.7
98.9
99.8
89.2
96.3
98.2
99.3
55.6
76.4
84.4
91.0
dMat. Avg.
54.1
77.5
86.3
96.3
73.6
89.9
93.9
98.0
77.7
90.8
95.2
98.4
37.0
59.1
71.5
84.8
LogSumExp
76.4
92.9
95.8
98.6
90.7
96.5
98.4
99.8
88.8
96.8
97.5
99.3
55.4
73.6
83.1
91.0
Sum Descriptors
62.9
85.8
93.3
97.9
80.3
94.1
97.4
98.8
81.2
93.7
96.8
98.7
42.9
67.1
77.2
87.5
QR-FullVP (Ours)
82.7
95.9
98.0
99.3
93.5
98.6
99.3
99.8
91.4
97.9
99.0
99.6
60.3
81.8
88.6
94.0
TABLE VII: Retrieval performance on Pitts250k-Test using CNN-based Models.
Approach
MegaLoc (8448D)
SALAD (8448D)
CricaVPR (10752D)
BoQ (12288D)
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
Pooling
94.4
99.6
99.8
99.9
90.3
97.9
99.0
99.4
79.1
95.8
98.0
99.8
94.9
99.7
99.9
99.9
dMat. Avg.
88.4
98.9
99.8
99.8
83.0
95.6
97.8
99.0
62.6
84.6
90.1
94.7
92.9
98.7
99.5
99.9
LogSumExp
93.8
98.8
99.6
99.9
91.5
97.6
98.5
99.4
81.8
94.0
97.3
99.8
96.2
99.0
99.4
99.9
Sum Descriptors
90.6
99.4
99.8
99.8
87.2
97.5
98.5
99.1
68.5
88.5
93.7
97.4
94.2
99.3
99.8
99.9
QR-FullVP (Ours)
95.6
99.6
99.9
99.9
93.7
98.6
99.1
99.7
82.3
97.1
98.6
99.8
97.0
99.6
99.8
99.9
TABLE VIII: Retrieval performance on Pitts250k-Test (non-default split, as elsewhere - see Section V ) using Transformer-based Models.
Approach
CosPlace (2048D)
Eigenplaces (2048D)
MixVPR (4096D)
NetVLAD (4096D)
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
Pooling
53.3
69.2
75.2
84.4
71.8
83.2
88.3
93.7
75.9
87.3
90.8
93.0
22.9
40.6
54.6
63.8
dMat. Avg.
23.2
47.0
58.1
71.1
39.4
57.8
65.1
77.5
42.5
65.4
70.5
81.0
16.2
28.3
36.2
49.8
LogSumExp
53.0
69.8
74.9
84.4
70.5
83.2
87.9
93.7
71.4
83.2
88.9
93.0
22.5
37.5
49.8
63.8
Sum Descriptors
37.5
60.6
66.4
78.4
47.3
68.3
76.2
82.9
51.4
70.8
77.8
83.8
19.1
34.3
40.0
58.4
QR-FullVP (Ours)
55.9
73.7
81.3
89.5
75.6
85.7
90.8
94.9
76.5
87.0
89.2
92.4
25.4
44.8
56.5
67.0
TABLE IX: Retrieval performance on Tokyo24/7 using CNN-based Models.
Approach
MegaLoc (8448D)
SALAD (8448D)
CricaVPR (10752D)
BoQ (12288D)
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
Pooling
95.9
98.4
99.1
99.7
88.9
96.5
97.1
99.1
80.0
90.2
93.3
95.6
94.0
97.8
98.4
98.7
dMat. Avg.
85.1
94.3
98.4
99.4
61.3
83.5
89.2
93.7
32.4
53.7
64.8
75.9
87.3
97.1
98.7
98.7
LogSumExp
94.0
98.4
99.1
99.7
87.9
94.6
97.1
99.1
71.8
86.4
90.8
95.6
94.3
98.1
98.4
98.7
Sum Descriptors
88.3
96.2
98.7
99.1
74.6
90.2
93.3
97.1
39.7
61.3
72.1
81.0
89.8
97.1
98.7
98.7
QR-FullVP (Ours)
94.8
99.1
99.7
99.7
91.4
96.8
97.5
98.4
76.5
88.9
92.1
94.9
94.6
97.8
98.4
98. 7
TABLE X: Retrieval performance on Tokyo24/7 (non-default split, as elsewhere - see Section V ) using Transformer-based Models.
Approach
Eigenplaces (2048D)
CosPlace (2048D)
MixVPR (4096D)
NetVLAD (4096D)
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
Pooling
53.6
64.7
69.6
75.3
48.1
60.0
65.2
72.1
59.7
71.0
75.4
80.9
41.1
54.9
61.0
69.6
dMat. Avg.
35.2
51.3
58.0
66.7
28.4
45.0
52.8
62.4
45.2
60.6
66.7
74.6
29.1
46.6
53.9
64.7
LogSumExp
52.6
65.5
70.7
75.3
47.2
60.9
67.1
72.1
57.1
71.0
76.5
80.9
41.1
56.5
63.8
69.6
Sum Descriptors
46.4
61.7
67.3
74.2
39.5
56.3
62.4
70.7
54.8
67.9
73.1
79.3
40.4
55.8
62.2
71.1
QR-FullVP (Ours)
58.8
70.4
74.9
80.3
53.3
66.0
71.0
77.7
64.9
76.4
79.8
84.7
48.7
62.7
68.4
75.4
TABLE XI: Retrieval performance on the GLDv2-micro dataset using CNN-based Models.
Approach
MegaLoc (8448D)
SALAD (8448D)
CricaVPR (10752D)
BoQ (12288D)
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
R@1
R@5
R@10
R@25
Pooling
86.8
92.3
94.0
95.9
71.8
82.2
85.3
89.1
70.2
81.4
84.8
88.8
74.2
83.6
86.5
89.9
dMat. Avg.
80.5
89.4
92.6
94.8
59.1
74.3
79.0
85.2
52.1
67.4
73.3
80.1
67.2
80.2
84.4
88.4
LogSumExp
85.1
92.7
94.6
95.9
70.2
82.5
86.5
89.1
66.4
79.6
84.5
88.8
71.7
83.1
86.9
89.9
Sum Descriptors
84.6
92.3
94.4
95.8
68.7
79.9
84.3
88.9
65.9
79.3
83.7
88.2
71.9
83.4
86.7
90.2
QR-FullVP (Ours)
89.4
94.3
95.7
97.0
76.8
86.2
89.1
92.3
76.2
86.4
88.9
92.8
77.5
86.7
89.4
92.4
TABLE XII: Retrieval performance on the GLDv2-micro dataset using Transformer-based Models.
Nordland
Queries →
Fall
Spring
Summer
Winter
↓ References
SeqVLAD (24576D)
Fall
—
85.7
95.0
68.4
Spring
78.9
—
80.0
77.3
Summer
92.2
83.6
—
69.4
Winter
57.9
74.2
60.9
—
TABLE XIII: Retrieval performance using SeqVLAD across datasets.
Fig. 5: QR-FullVP residuals using Eigenplaces for correct (green) and incorrect (red) top-1 matches across all datasets. Correct matches show lower residuals across all conditions. Appearance-challenging datasets (ORC Night, Nordland Winter), show larger overall residuals, reflecting a larger domain gap between query and reference basis. Multi-viewpoint datasets exhibit high residuals (from non-overlapping viewpoints), yet still improve retrieval over pooling, suggesting the subspace is sufficiently discriminative.
Fig. 6: Retrieval comparison on Oxford RobotCar (overcast query) using Eigenplaces. Left shows retrieval outcomes for single reference matching (dusk reference) and QR-FullVP (all references), with green indicating correct and red incorrect top-1 matches. The difference map (center) highlights where QR-FullVP succeeds over single-reference (green), the reverse (red) or both agree (grey). As expected, QR-FullVP resolves ambiguity across large portions of the traverse. (Retrieved images for QR-FullVP show a randomly sampled image from the matched place’s reference collection).
Fig. 7: Retrieval comparison on Oxford RobotCar (snow query) using Eigenplaces. Left shows retrieval outcomes for reference pooling and QR-FullVP, with green indicating correct and red incorrect top-1 matches. The difference map (center) highlights where QR-FullVP succeeds over pooling (green), the reverse (red), or both agree (grey). Overall, QR-FullVP outperforms pooling, and qualitative examples suggest this is often a fine disambiguation, such as the third example where the pooling match falls just outside the ground truth tolerance while QR-FullVP recovers the correct place. (Retrieved images for QR-FullVP show a randomly sampled image from the matched place’s reference collection).
Fig. 8: Retrieval comparison on Oxford RobotCar (night query) using Eigenplaces. Left shows retrieval outcomes for reference descriptor summing and QR-FullVP, with green indicating correct and red incorrect top-1 matches. The difference map (center) highlights where QR-FullVP succeeds over summing (green), the reverse (red), or both agree (grey). Although outperforming pooling, summing produces a more robust aggregate under night conditions, likely because this severe appearance shift is not as well represented by the span of the reference basis, and benefiting from reinforcement of low-variance structure. However, summing relies on viewpoint consistency across references, with performance degrading when viewpoints differ within a place’s collection or where outlier effects such as light refraction are present, as in the third example. (Retrieved images for Summed & QR-FullVP show a randomly sampled image from the matched place’s reference collection).
Fig. 9: Retrieval comparison on Pitts250k-Test using Eigenplaces. Left shows retrieval outcomes for reference pooling and QR-FullVP, with green indicating correct and red incorrect top-1 matches. The difference map (center) highlights where QR-FullVP succeeds over pooling (green), the reverse (red), or both agree (grey). QR-FullVP outperforms pooling, using complementary viewpoint information, and qualitative examples suggest that QR-FullVP improves disambiguation, retrieving places closer to the true location. (Retrieved images for QR-FullVP show a randomly sampled image from the matched place’s reference collection).
Fig. 10: Recall@1 versus descriptor size. Our QR-based approach using smaller descriptor sizes often outperforms other multi-reference approaches at maximum size, offering optimization with map memory. Note: for each plot the y-axis is scaled respectively for clear visualization.
Fig. 11: Change in Recall@1 against pooling ( with the same descriptor size ) using SVD decomposition instead of QR. At full rank (top rows), the performance is equivalent to QR-FullVP.
Fig. 12: Average cumulative variance of SVD-derived principal components for reference representations across the datasets. Nordland and Oxford RobotCar exclude Fall and Sun conditions, respectively.
Method
Dim
Approach
Extraction Time (ms/query)
SotonMV (300 Query, 63872 Ref. Images)
Pitts250k Test (1000 Query, 20988 Ref. Images)
Match Time (ms/query)
Total Map Time (s)
Map Mem. (GB)
R@1
Match Time (ms/query)
Total Map Time (s)
Map Mem. (GB)
R@1
CosPlace
2048
Pooling
5.90
0.43
-
0.523
34.2
0.15
-
0.172
74.2
CosPlace
2048
QR-FullVP
5.57
1.41
3.22
0.523
43.7
0.48
0.88
0.172
82.7
CosPlace
2048
SVD (k-1)
4.78
1.20
9.25
0.392
36.9
0.39
2.50
0.143
83.5
CosPlace
256
Pooling
6.21
0.12
-
0.0654
32.7
0.05
-
0.0215
69.1
CosPlace
256
QR-FullVP
5.61
0.26
1.66
0.0654
42.3
0.11
0.45
0.0215
81.9
TABLE XIV: Computational overhead of conventional VPR vs QR and SVD-based approach using an Nvidia A100 GPU with a batch size of 1. QR uses all ( k ) references for the subspace while SVD truncates to k−1 . Note that matching and map construction have asymptotic complexity across approaches, however, measured differences reflect our implementation choices and kernel behavior, with pooling executed as a matrix–vector multiplication, subspace methods using einsum-based projections with reductions, map construction performed via full per-place decompositions and explicit synchronizations, affecting the practical runtime.
Fig. 13: Match and map construction latency for pooling, QR-FullVP, and SVD (k-1) evaluated on an Apple M4 CPU using SciPy linear algebra with synthetic descriptors. Queries are processed sequentially (batch size 1) with single-threaded BLAS, and all approaches use identical matrix-vector multiplication for matching, differing only in the post-multiplication aggregation. Map construction is performed per place via QR or truncated SVD. Results are reported as median latency over 300 queries across sweeps in descriptor dimension, number of places, and references per place, which, when non-swept are fixed to 512, 500, and 4 respectively.
Fig. 14: Evaluation of orientation estimates with SotonMV. ( Left ) Distribution of absolute angular errors. ( Right ) Predicted vs ground-truth headings.
Methods →
EigenPlaces (2048D)
CosPlace (2048D)
MixVPR (4096D)
Queries →
Fall
Spring
Summer
Winter
Fall
Spring
Summer
Winter
Fall
Spring
Summer
Winter
References →
Sp+Su
Sp+Wi
Su+Wi
Fa+Su
Fa+Wi
Su+Wi
Fa+Sp
Fa+Wi
Sp+Wi
Fa+Sp
Fa+Su
Sp+Su
Sp+Su
Sp+Wi
Su+Wi
Fa+Su
Fa+Wi
Su+Wi
Fa+Sp
Fa+Wi
Sp+Wi
Fa+Sp
Fa+Su
Sp+Su
Sp+Su
Sp+Wi
Su+Wi
Fa+Su
Fa+Wi
Su+Wi
Fa+Sp
Fa+Wi
Sp+Wi
Fa+Sp
Fa+Su
Sp+Su
Pooling
96.4
80.7
94.4
91.4
91.4
87.2
93.6
93.3
73.7
69.1
67.3
69.5
95.1
76.8
93.0
88.4
89.0
85.0
92.1
91.2
68.7
69.8
61.7
69.8
97.7
88.1
95.8
97.0
97.6
95.0
96.4
95.9
83.3
90.7
85.4
91.4
dMat. Avg.
95.9
76.2
90.7
90.0
91.8
89.1
91.1
87.5
68.7
76.6
66.9
75.8
92.7
69.7
83.6
87.3
86.2
83.6
87.2
80.2
62.7
73.4
61.1
71.6
97.6
90.1
96.6
96.2
98.1
96.6
95.3
94.8
85.2
93.2
84.2
92.9
LogSumExp
97.0
81.7
95.1
90.9
94.1
90.5
93.5
93.5
74.8
74.8
67.6
74.3
95.2
77.6
93.5
88.7
90.4
87.0
91.6
91.8
69.3
73.0
62.3
73.1
98.1
89.9
96.6
96.7
98.3
96.6
96.2
95.9
84.9
93.3
85.2
93.4
Sum Descriptors
97.8
80.4
95.4
91.5
95.8
93.9
93.3
92.3
73.2
75.2
67.6
75.0
96.3
75.3
91.5
88.7
92.6
90.7
91.4
87.7
68.2
72.4
61.5
71.8
98.6
90.9
97.7
97.3
98.5
97.7
96.5
96.1
86.3
92.7
86.4
92.9
TABLE XV: Recall@1 with Nordland retrieval using pairwise reference-season combinations for place representations.
Methods →
MegaLoc (8448D)
CricaVPR (10752D)
BoQ (12288D)
SALAD (8448D)
Queries →
Sun
Overcast
Night
Snow
References →
Du+Ra
Du+Ov
Du+Ni
Du+Sn
Ra+Ov
Ra+Ni
Ra+Sn
Ov+Ni
Ov+Sn
Ni+Sn
Su+Du
Su+Ra
Su+Ni
Su+Sn
Du+Ra
Du+Ni
Du+Sn
Ra+Ni
Ra+Sn
Ni+Sn
Su+Du
Su+Ra
Su+Ov
Su+Sn
Du+Ra
Du+Ov
Du+Sn
Ra+Ov
Ra+Sn
Ov+Sn
Su+Du
Su+Ra
Su+Ov
Su+Ni
Du+Ra
Du+Ov
Du+Ni
Ra+Ov
Ra+Ni
Ov+Ni
Pooling
93.1
95.8
73.2
88.8
97.0
93.6
94.3
95.4
96.4
89.4
96.6
98.2
97.2
97.8
95.5
81.0
94.0
96.3
97.9
95.1
83.7
91.4
91.3
90.5
82.8
85.2
84.3
92.1
90.2
91.2
88.3
90.4
92.5
89.6
87.1
91.1
78.2
91.8
88.0
91.4
dMat. Avg.
91.5
93.6
75.6
87.8
97.4
93.1
95.5
95.2
96.0
90.2
94.9
98.1
97.4
98.1
93.1
84.0
90.9
96.4
98.0
95.6
88.7
91.8
92.7
92.5
89.3
88.7
88.5
92.9
92.5
93.0
86.2
90.5
92.3
89.2
86.0
87.8
79.3
91.4
88.5
90.3
LogSumExp
93.3
96.0
76.7
89.7
97.7
94.0
95.7
95.8
96.7
90.4
96.7
98.3
97.4
98.0
95.7
83.5
92.7
97.0
98.1
95.6
87.2
92.4
92.9
92.5
87.0
88.1
87.4
93.5
92.5
93.2
89.3
91.0
92.9
90.8
88.8
91.9
80.4
92.2
89.4
91.5
Sum Descriptors
93.2
94.9
77.8
89.1
97.7
94.7
96.4
96.1
96.4
91.4
96.8
98.5
97.9
98.9
95.6
87.0
93.4
97.2
98.7
96.9
89.7
92.5
93.1
93.5
90.0
90.3
90.1
93.7
92.8
93.7
89.4
91.4
93.0
90.5
88.5
90.8
82.0
92.6
90.5
91.5
TABLE XVI: Recall@1 with Oxford Robot Car retrieval using pairwise reference-weather combinations for place representations.
TABLE XVII: Recall@1 comparison of our QR approach, creating one place representation (FullVP) vs multiple representations using pairwise adjacent viewpoints (2VP), on SotonMV and Pitts250k.
TABLE XVIII: Recall@1 using query images aligned with reference map headings. Despite the additional non-overlapping viewpoint information, our approach suffers little to no degradation.
A key challenge in Visual Place Recognition (VPR) is matching query images against reference maps captured under diverse environmental conditions and viewpoints. While multiple reference traversals improve robustness, existing fusion strategies either aggregate references uniformly or rely on heuristic selection, without distinguishing descriptor variations that preserve stable place identity from those caused by changing conditions or viewpoints. In this paper, we propose DisPlace, a multi-reference VPR framework that fuses multiple reference descriptors into a single compact and discriminative place representation. DisPlace formulates descriptor fusion as a generalized eigenvalue problem that maximizes between-place separability while suppressing within-place variation across references, rather than preserving overall descriptor variance. Unlike existing multi-reference fusion methods, DisPlace exploits variation across reference traversals to identify which linear combinations of descriptor dimensions preserve place identity and which capture condition- or viewpoint-specific variation. We evaluate DisPlace on Oxford RobotCar, Nordland, Pittsburgh30k, and Google Landmarks v2 across six state-of-the-art VPR descriptors. DisPlace outperforms seven multi-reference baselines in 49 out of 54 appearance-varying conditions, consistently improves descriptor-level fusion performance under viewpoint and unstructured settings, and requires less storage during inference than all compared fusion methods.
Dhyey Manish Rajani, Michael Milford, Tobias Fischer
QUT Centre for Robotics, School of Electrical Engineering and Robotics at the Queensland University of Technology.
Modern Visual Place Recognition (VPR) methods excel on standard benchmarks yet remain brittle in feature-poor environments. By treating each query image in isolation, they discard the sequential context in any real trajectory. We formalize a task that exploits this context: given a query sequence, localize the final image against an unordered reference database -- which, unlike sequence-to-sequence methods, requires no sequential structure in the database. We propose TRAIL (TRajectory-Aware Image Localization), a principled framework based on Conditional Random Fields (CRF) that combines learned functions for visual similarity and for camera-motion consistency, refining a distribution over candidate references as each query arrives. A lightweight post-processing layer atop any pre-trained VPR backbone, TRAIL improves a state-of-the-art baseline by up to 8.3 percentage points on our primary benchmark, transfers to unseen datasets without retraining, and delivers its largest gains where visual cues are scarce.
Dominik A. Kloepfer, Patrick Wenzel
Visual Geometry Group (VGG), University of Oxford · Helsing
Visual Place Recognition (VPR) has been traditionally formulated as a single-image retrieval task. Using multiple views offers clear advantages, yet this setting remains relatively underexplored and existing methods often struggle to generalize across diverse environments. In this work we introduce UniPR-3D, the first VPR architecture that effectively integrates information from multiple views. UniPR-3D builds on a VGGT backbone capable of encoding multi-view 3D representations, which we adapt by designing feature aggregators and fine-tune for the place recognition task. To construct our descriptor, we jointly leverage the 3D tokens and intermediate 2D tokens produced by VGGT. Based on their distinct characteristics, we design dedicated aggregation modules for 2D and 3D features, allowing our descriptor to capture fine-grained texture cues while also reasoning across viewpoints. To further enhance generalization, we incorporate both single- and multi-frame aggregation schemes, along with a variable-length sequence retrieval strategy. Our experiments show that UniPR-3D sets a new state of the art, outperforming both single- and multi-view baselines and highlighting the effectiveness of geometry-grounded tokens for VPR. Our code and models will be made publicly available on Github https://github.com/dtc111111/UniPR-3D.
Tianchen Deng, Xun Chen, Ziming Li +5
Shanghai Jiao Tong University · Nanyang Technological University · University of Macau +1