Evaluation of Visual Place Recognition Methods for Image Pair Retrieval in 3D Vision and Robotics
Authors: Dennis Haitz, Athradi Shritish Shetty, Michael Weinmann, Markus Ulrich
Organizations: Karlsruhe Institute of Technology, Institute of Photogrammetry and Remote Sensing, Karlsruhe, Germany · Delft University of Technology, Department of Intelligent Systems, The Netherlands
Visual Place Recognition (VPR) is a core component in computer vision, typically formulated as an image retrieval task for localization, mapping, and navigation. In this work, we instead study VPR as an image pair retrieval front-end for registration pipelines, where the goal is to find top-matching image pairs between two disjoint image sets for downstream tasks such as scene registration, SLAM, and Structure-from-Motion. We comparatively evaluate state-of-the-art VPR families - NetVLAD-style baselines, classification-based global descriptors (CosPlace, EigenPlaces), feature-mixing (MixVPR), and foundation-model-driven methods (AnyLoc, SALAD, MegaLoc) - on three challenging datasets: object-centric outdoor scenes (Tanks and Temples), indoor RGB-D scans (ScanNet-GS), and autonomous-driving sequences (KITTI). We show that modern global descriptor approaches are increasingly suitable as off-the-shelf image pair retrieval modules in challenging scenarios including perceptual aliasing and incomplete sequences, while exhibiting clear, domain-dependent strengths and weaknesses that are critical when choosing VPR components for robust mapping and registration.
Figures & tables
Figure 1: Example images from sets A (left) and B (right), taken from the Caterpillar scene of the Tanks and Temples dataset ( Knapitsch et al., 2017 ) . Images including overlapping scene area are framed in red, the objective in our contribution is to test the ability of VPR methods to retrieve these pairs. Challenging for VPR models in such scenes is the perceptual aliasing, e.g. indicated by the wheels, which are visible on both sides and even appear in similar relative positions.
Method
Backbone
Descriptor size
CosPlace512
ResNet18
512
CosPlace2048
ResNet101
2048
EigenPlaces512
ResNet18
512
EigenPlaces2048
ResNet101
2048
MixVPR512
ResNet50
512
MixVPR4096
ResNet50
4096
Table 1: Different backbone and descriptor size configurations used for the methods CosPlace ( Berton et al., 2022 ) , EigenPlaces ( Berton et al., 2023 ) , MixVPR ( Ali-Bey et al., 2023 ) , and SALAD ( Izquierdo and Civera, 2024 ) .
Figure 2: Qualitative top-5 retrieval results for different VPR methods on the T&T scene Barn . This scene is especially prone to perceptual aliasing, as indicated by the brown door, which is at different positions on both sides and the color of the floor in front of the house. Overlap areas are the back of the house, which is completely correctly retrieved by SALAD8192 and MixVPR4096. Columns correspond to methods (a–f) and, within each column, images are shown from rank k=1 (top) to k=5 (bottom).
Figure 3: Qualitative Top-5 retrieval results from CosPlace512 (a) and MegaLoc (b) for KITTI scene 07_02. The challenge here is that some images were left out between A and B . This is especially important for tasks such as SLAM relocalization caused from missing frames. CosPlace ( Berton et al., 2022 ) completely fails here, likely because of perceptual aliasing, indicated by two different black cars in A and B and other similarities, whereas MegaLoc ( Berton and Masone, 2025 ) yields only true positives. The cue for correct results is the houses in the center of the A images, which appear towards the left border of the B images. Images are shown from k=1 (top) to k=5 (bottom) per method.
Dataset
Method
P@1
P@5
P@10
R@1
R@5
R@10
mAP@5
mAP@10
μt
σt
PatchNetVLAD
25.00
21.25
17.50
25.00
37.50
43.75
25.10
26.27
1.102
0.105
CosPlace512
56.25
43.75
36.88
56.25
56.25
62.50
55.62
52.69
0.133
0.014
CosPlace2048
56.25
40.00
32.50
56.25
56.25
62.50
53.33
51.54
0.645
0.070
EigenPlaces512
56.25
47.50
36.88
56.25
56.25
62.50
54.37
54.35
0.134
0.016
EigenPlaces2048
56.25
45.00
35.62
56.25
62.50
62.50
57.19
53.67
0.649
0.070
T&T
MixVPR512
18.75
6.25
5.00
18.75
25.00
31.25
21.88
21.56
0.130
0.012
Table 2: Image pair retrieval across different datasets, consisting of multiple scenes (quantity in braces) for each dataset: T&T (16), SN-GS (26), and KITTI (26). Each dataset is evaluated on seven VPR methods, split into eleven configurations. Metrics are calculated for the k best ranked image pairs, leading to Precision@k (P@k), Recall@k (R@k), and mean-Average-Precision@k (mAP@k), all given in percent. Average time in seconds for all retrievals per scene is included as μt with σt as the standard deviation, respectively.
Figure 4: Overview of P@10 (left column) and R@10 (right column) for all VPR methods on T&T (top row), SN-GS (middle row) and KITTI (bottom row).
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