TL;DR
GOOSELoc is the first multimodal benchmark specifically designed for ground-to-aerial localization in unstructured off-road environments, featuring 9,356 scenes with synchronized 360° RGB imagery and 3D LiDAR. The paper demonstrates that while state-of-the-art urban localization methods fail in the woods, a multimodal approach—using LiDAR for geometric positioning and cameras for orientation—significantly narrows the performance gap.
What Problem Does This Paper Address?
Most autonomous navigation systems rely on GNSS, but signals often degrade under dense forest canopies or in uneven terrain. Cross-view geo-localization—matching ground-level sensor data to overhead aerial imagery—serves as a vital backup.
However, existing benchmarks like VIGOR or CVUSA are almost entirely urban or road-centric. They rely on stable, human-made "anchors" like buildings and lane markings. In off-road settings, these anchors disappear. Instead, you have self-similar vegetation that changes drastically with the seasons (the "appearance gap"). This paper addresses the lack of data and specialized methods for metric localization in these challenging, unstructured environments.
Methods
The authors first adapted four state-of-the-art image-only baselines (CCVPE, Shi et al., HC-Net, and FG²) to the off-road setting. Finding their performance lacking, they proposed a new multimodal baseline.
The Multimodal Baseline: * Encoders: It leverages frozen foundation models: DINOv3 for RGB images and Utonia (with a LoRA adapter) for LiDAR point clouds. * Position Branch: The model generates separate score maps for RGB and LiDAR by correlating ground features with aerial patches. A novel "evidence gate" (a small MLP) analyzes the entropy and peakiness of these maps to dynamically weight their contribution to the final fused position estimate. * Heading Branch: To resolve orientation, the model uses a learned "generalized Hough vote." It correlates features from the four surround cameras and eight LiDAR sectors against the aerial tile to find the rotation that best aligns the ground observations with the map.
Datasets
The core contribution is the GOOSELoc dataset.
- Composition: 9,356 scenes recorded around Munich, Germany, covering all four seasons.
- Sensors: Each scene includes four 360° surround-view RGB images and a 128-beam LiDAR sweep.
- Reference: Ground data is paired with 20 cm/pixel geo-referenced aerial orthophotos (captured in summer).
- Labels: Includes paved/unpaved terrain labels and high-accuracy ground truth from an RTK-GNSS/INS system.
Key Insights and Ideas
- The Division of Labor: One of the most striking findings is the complementary nature of the sensors. LiDAR is the "where" sensor: its 3D geometry provides a strong position cue against the overhead map. Cameras are the "which way" sensor: because forest geometry is often front/back symmetric, LiDAR struggles with orientation, whereas camera appearance breaks this 180° ambiguity.
- Urban Methods Don't Travel Well: Methods that perform excellently on urban datasets (like FG²) see their errors roughly double when moved off-road. The lack of structured geometry makes the matching problem much more ambiguous.
- Seasonal Sensitivity: Localization is hardest in "leaf-on" seasons (summer/autumn) when dense canopy occludes the ground from the aerial view. Interestingly, the LiDAR advantage over RGB grows in summer because geometry remains more stable than foliage appearance.
- Perceptual Aliasing: The authors quantified "look-alike" positions in aerial tiles. Unpaved, off-road tiles have ~28% more perceptual aliasing than paved ones, explaining why models often get "lost" in similar-looking patches of forest.
Key Results
The multimodal fusion (RGB+L) consistently outperformed single-modality baselines across all settings.
- North-up (No heading prior): The multimodal baseline achieved a median localization error of 22.2m, compared to 37.0m for the best-adapted urban baseline (CCVPE).
- IMU-prior (±10° heading prior): With a coarse heading, the error dropped to 14.1m.
- Terrain Impact: On unpaved off-road terrain, the median error for the RGB+L model was 15.4m (with IMU prior), whereas the image-only FG² baseline struggled at 27.0m.
Limitations and Future Work
The authors acknowledge several constraints:
- Geographic Scope: The data is confined to a specific region in Upper Bavaria; generalization to different biomes (e.g., deserts or tropical forests) remains untested.
- Aerial Mismatch: The aerial imagery is summer-only. While this reflects real-world availability, it creates a permanent "appearance gap" for winter ground data.
- Single-Frame Reasoning: The current baselines process single snapshots. Incorporating temporal information (multi-frame reasoning) could likely smooth out the "jumps" seen in the qualitative results.
The team has released the data and baselines at gooseloc.github.io to encourage further research into robust, off-road navigation.
Summary updated