An open computer-vision benchmark built from 617,388 ground-level and aerial image pairs reported lower errors on its in-the-wild test when Mapillary imagery was included in training. In the final OpenCVL setting, mean localization error was 7.90 metres and mean orientation error was 26.07 degrees, compared with 8.61 metres and 35.59 degrees for ZOD-only training. Those figures come from a computational comparison, not a randomized test, so they show an association rather than a causal effect.
The paper addresses the stated lack of an open, large-scale and diverse dataset for fine-grained cross-view localization. In plain terms, the task is to use a ground-level image and an aerial image to estimate where the scene is and which way the camera was facing. OpenCVL draws on permissive sources for both kinds of imagery, including ZOD, Mapillary and national mapping agencies.
A benchmark built from two views
OpenCVL contains 617,388 pairs spanning 41 cities across four European countries. Of those, 579,752 were used for training, including 238,212 ZOD pairs and 341,540 Mapillary pairs. Validation contained 14,756 pairs. The cross-area, snowy and in-the-wild test sets contained 18,504, 3,015 and 1,361 pairs, respectively.
Reliable poses, the camera positions and directions attached to images, were a central part of the Mapillary curation. The framework matched local ZOD-Mapillary visual features and checked them geometrically, then used Perspective-n-Point estimation followed by Levenberg-Marquardt refinement of the pose and camera parameters. Estimates with high reprojection error, meaning the projected points fit the images poorly, were discarded. Computed poses were then manually checked using LiDAR projections in the ground image and pose checks in aerial imagery.
A smaller weight for Mapillary
The training experiments also varied the influence of Mapillary examples. Every Mapillary sample received the same global loss weight, reducing how strongly that example contributed to the training objective. Factors of 0.1, 0.3, 0.5 and 0.7 were tested, and 0.3 was selected as slightly better. With that setting, mean localization error was 6.72 metres on the cross-area split, versus 6.87 metres with no weighting, and 7.80 metres on the snowy split, versus 8.04 metres with no weighting.
Across the final OpenCVL tests, the cross-area split produced mean and median localization errors of 6.72 and 4.30 metres, with mean and median orientation errors of 11.52 and 5.24 degrees. The snowy split recorded 7.80 and 5.48 metres for localization, and 5.09 and 3.52 degrees for orientation. On the in-the-wild split, the corresponding figures were 7.90 and 6.19 metres, and 26.07 and 12.01 degrees.
Pose labels and transfer tests
A separate in-the-wild validation comparison gave different benchmark outputs for different pose labels. With no fine-tuning, mean and median localization errors were 8.05 and 5.75 metres, while orientation errors were 26.73 and 10.57 degrees. Raw GNSS labels produced 7.41 and 5.73 metres, and 22.09 and 7.70 degrees. OpenSfM labels produced 7.20 and 5.23 metres, and 23.44 and 9.27 degrees. The corrected pose produced the lowest errors among the tested labels: 6.92 and 5.14 metres for localization, and 17.98 and 7.01 degrees for orientation.
Transfer to KITTI was mixed. On the KITTI cross-area split without an orientation prior, full OpenCVL training had a lower mean localization error than KITTI-only training, 11.25 versus 11.71 metres, but a higher median, 9.40 versus 9.11 metres. Orientation errors were also higher with OpenCVL: 71.17 and 46.96 degrees for the mean and median, compared with 55.18 and 33.41 degrees for KITTI-only training.
The picture changed when the setup included a prior restricting orientation to plus or minus 10 degrees. OpenCVL pretraining followed by KITTI fine-tuning was reported as the best overall KITTI strategy. It reached 0.93 metres mean and 0.61 metres median localization error on same-area data, with orientation errors of 1.62 and 1.19 degrees. On cross-area data, it recorded 5.07 and 2.84 metres, and 3.30 and 2.15 degrees.
What the numbers leave open
Loc2 also compared favorably with the listed baselines on the three OpenCVL test splits. Its mean localization error was 6.72 metres on cross-area, versus 13.87 metres for HC-Net and 11.95 metres for CCVPE; on snowy data, 7.80 versus 12.82 and 10.81; and in-the-wild, 7.90 versus 12.66 and 16.24. These are benchmark results from Loc2 experiments, not evidence that every cross-view model would perform the same way.
The evidence has a defined boundary. OpenCVL covers four European countries, the main experiments center on Loc2, and the computational comparisons were not randomized. The paper reports no confidence intervals or formal significance tests, leaving broader geographic and model generalization unresolved.
The manuscript is an arXiv version 1 preprint dated 26 Aug 2026. The authors acknowledge support from the Swiss Open Research Data Fund under project PgB_25-28_674_A1_19.
Paper data and sources
Original title: OpenCVL: An Open, Diverse, and Large-Scale Dataset for Fine-Grained Cross-View Localization
Authors: Zimin Xia, Mubariz Zaffar, Junsheng Fu et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text