AutoCompass

Accurate Visual Localization on Public Maps
by Learning from Weak Labels

ECCV 2026

Javier Tirado-Garín1,*   Alan Savio Paul2   Shuai Chen2   Axel Barroso-Laguna2   Tommaso Cavallari2   Daniyar Turmukhambetov2   Victor Adrian Prisacariu2,3   Eric Brachmann2

1I3A, Universidad de Zaragoza    2Niantic Spatial    3University of Oxford   
*Work done during an internship at Niantic Spatial.
Comparison of standard pose supervision and AutoCompass supervision under inaccurate geo-referenced training poses.
Geo-localization datasets can be noisy, such as in this example, where images captured on a sidewalk have poses inside a building. Methods trained with strong supervision learn and replicate these errors, while AutoCompass is robust to them. We treat the same GT poses as weak labels by: 1. modeling the unknown true pose as lying anywhere within a neighborhood, and 2. supervising with relative poses, which are robust to shared shift errors such as the one in this example.

Abstract

Neural map matchers estimate an image's 3-DoF pose relative to a 2D map. These models are trained on large-scale datasets of geo-referenced images, whose position and heading labels often contain noise that affects the trained models. To address this, we present AutoCompass, a supervision approach for training neural map matchers from inaccurate absolute pose labels. First, we show that heading labels are unnecessary: trained from raw GPS labels, models learn to predict accurate headings, automatically. Second, defining a tolerance region around raw GPS improves positional accuracy. Third, if available, our supervision uses relative poses between training images, obtained via SLAM or SfM, which provide a more accurate training signal. Across driving and egocentric benchmarks, AutoCompass consistently outperforms counterparts trained with the usual strong reliance on absolute pose labels.

Method

Overview of neural map matching, conventional strong pose supervision, and AutoCompass supervision using GPS neighborhoods and relative poses.
(a) AutoCompass adopts neural-map matching to estimate a distribution over camera poses. (b) Current methods supervise this distribution under the assumption of accurate 3-DoF pose labels (Opt.). (c, d) We instead propose weak-supervision strategies that are robust to such inaccuracies. The simplest strategy, shown in (c), requires only noisy 2-DoF GPS labels, as it maximizes the probability assigned to a GPS-centered spatial chunk, thus only assuming that the true pose lies within this chunk. (d) We also learn from relative poses obtained from SfM, which are robust to offsets in geo-referenced poses

AutoCompass Learns More Detailed Features

AutoCompass is more robust to the presence of noise in the ground-truth pose labels. This robustness is reflected in the sharper features it learns. In contrast, existing approaches, such as OrienterNet, rely strongly on these labels. Therefore, their learned representations become spatially blurred, as the model must explain these errors on average.

This difference is much more striking when comparing the norms of these learned features. Intuitively, the cross-correlation scores are proportional to these norm values, and, as such, they are useful for understanding the regions of the map that the model focuses on for localization. AutoCompass focuses on distinctive cues that are commonly visible in the query images, such as building corners or lane intersections. In contrast, OrienterNet focuses on much larger regions, such as entire buildings, which are only useful for coarse localization.

2D map from OpenStreetMap
Neural Map

Quantitative Results

AutoCompass sets a new state of the art. It outperforms OrienterNet, the strongest alternative that uses the same 2D maps, as well as the best-performing approach that instead relies on aerial imagery as its map representation. AutoCompass is also significantly more accurate than GPS.

This is especially the case when fusing multiple single-view estimates in order to geo-reference an input sequence with known relative poses. This confirms that the more accurate single-view estimates are also benefitial for sequential fusion as the accuracy of AutoCompass improves more rapidly tan that of OrienterNet and GPS.

Legend for the quantitative results

Single-view

KITTI
Single-view recall at one meter lateral error on KITTI
Recall at 1 m lateral error.
Oxford Day-and-Night
Single-view recall at one meter position error on Oxford Day-and-Night
Recall at 1 m position error.
LaMAria
Single-view recall at one meter position error on LaMAria
Recall at 1 m position error.

Sequential

KITTI
Recall at one meter position error by sequence length on KITTI
Recall at 1 m versus number of images.
Oxford Day-and-Night
Recall at one meter position error by sequence length on Oxford Day-and-Night
Recall at 1 m versus number of images.
LaMAria
Recall at one meter position error by sequence length on LaMAria
Recall at 1 m versus number of images.

BibTeX

If you find this work useful for your research, please consider citing our paper:

@inproceedings{tirado2026autocompass,
    title={{AutoCompass: Accurate Visual Localization on Public Maps by Learning from Weak Labels}},
    author={Tirado-Garín, Javier and Paul, Alan Savio and Chen, Shuai and Barroso-Laguna, Axel and Cavallari, Tommaso and Turmukhambetov, Daniyar and Prisacariu, Victor Adrian and Brachmann, Eric},
    booktitle={ECCV},
    year={2026}
}