VectorReLoc: Reliable Vectorized SD Map Visual Re-localization with Contrastive Feature Alignment

Ziming Liu¹ · Quanjie Xiang^1,3 · Yun Wang² · Yiting Wang² · Wen Chao² · Zhuanjian Xu² · Leichen Wang¹ · Hao Sun¹ · Guangyu Gao³

¹Bosch Research, Shanghai · ²Bosch Cross-domain Computing Solutions (XC), Shanghai · ³Beijing Institute of Technology

ECCV 2026 · to appear

Overall architecture of VectorReLoc — Overall structure of VectorReLoc. Blue: RTK SD map (used only for the feature alignment loss during training). Orange: GNSS SD map — the two SD map encoders share parameters. Green: visual map derived from the camera images. The online map construction network is frozen.

Abstract

Standard-definition (SD) maps are widely used in autonomous driving, but accurate ego-centric SD map retrieval typically relies on costly RTK-GNSS joint localization. In contrast, low-cost GNSS introduces meter-level 3-DoF pose offsets with high variance, yielding misaligned local maps that can degrade downstream planning and control. Prior visual re-localization methods mostly rely on dense rasterized BEV representations, which are computation-heavy and easily distracted by scene content irrelevant to road geometry.

We propose VectorReLoc, the first sparse re-localization framework that directly aligns vectorized SD maps with online-constructed vectorized visual maps to estimate the 3-DoF pose offset. VectorReLoc further introduces a feature alignment objective that structures the embedding space for accurate offset regression, together with a reliability predictor to identify potentially unreliable corrections and mitigate silent failures. To enable realistic training and evaluation, we additionally present a large-scale dataset with paired RTK-retrieved and GNSS-retrieved SD maps, providing real pose-offset labels rather than the simulated noise used in prior work. Experiments on public benchmarks and the proposed dataset demonstrate improved accuracy, robustness and efficiency over previous approaches.

Contributions

Sparse vectorized re-localization. The first SD map visual re-localization method that bypasses dense BEV grids and operates directly on vectorized map representations, achieving the best accuracy at significantly reduced computational cost.
Contrastive map feature alignment. A bidirectional InfoNCE objective with a per-sample margin loss that explicitly structures the embedding space for reliable offset regression.
Reliability prediction. A per-dimension confidence scoring module; filtered reliable samples show higher recall and lower error, letting downstream modules selectively accept corrections.
New dataset and benchmark. A large-scale dataset with paired real RTK-GNSS and low-cost GNSS SD maps, plus the GNSS Offset Distribution Hypothesis: real GNSS offsets are heavy-tailed and well captured by Student-t distributions rather than Gaussian noise.

VectorReLoc Dataset

Roughly 6,534 scenes and 170,533 frames recorded at 1 Hz across Shanghai, Jiangsu and Anhui, covering urban, suburban, highway, night and rain conditions. Each frame provides 8 surround-view images at 1024×768, low-cost GNSS/IMU measurements, dual RTK poses and the corresponding SD maps from a commercial map provider. Unlike prior benchmarks that perturb a reference trajectory with artificial noise, every offset label here comes from real paired RTK and low-cost GNSS retrievals.

Comparison of datasets used in SD map visual re-localization
Dataset	Scenes	Frames	Freq.	RTK-GNSS loc.	GNSS loc.
nuScenes	~1,000	40,157	2 Hz	✗	✓
Argoverse2	~850	27,000	2 Hz	✗	✓
VectorReLoc (ours)	6,534	170,533	1 Hz	✓	✓

Histogram of GNSS position offsets — Position offset histogram

Histogram of GNSS yaw offsets — Yaw offset histogram

Contrastive Feature Alignment

The visual map is built under the true vehicle pose while the SD map is queried under an inaccurate GNSS pose. A well-behaved embedding space therefore requires the visual-map feature to align with the RTK SD-map feature, while the GNSS SD-map feature is pushed away from both. Plain regression does not enforce this geometry in the common small-offset regime, so we optimise a bidirectional InfoNCE loss with a per-sample margin term jointly with the regression loss.

t-SNE of map features for the baseline model — Baseline

t-SNE of map features with joint contrastive training — Joint training with feature alignment

Qualitative Results

Ground-truth local visual map
Ground-truth RTK SD map
GNSS-retrieved SD map (before correction)
VectorReLoc-corrected SD map

Qualitative SD map re-localization results, first sequence

Qualitative SD map re-localization results, second sequence

BibTeX

@inproceedings{vectorreloc-eccv-2026,
  title     = {VectorReLoc: Reliable Vectorized SD Map Visual Re-localization with Contrastive Feature Alignment},
  author    = {Liu, Ziming and Xiang, Quanjie and Wang, Yun and Wang, Yiting and Chao, Wen and Xu, Zhuanjian and Wang, Leichen and Sun, Hao and Gao, Guangyu},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}

Acknowledgements

Supported by the National Natural Science Foundation of China (Grant No. 62472033) and the Beijing Natural Science Foundation (Grant No. L2607023), and by the Bosch High Performance Computing Cluster.