TrajLoc Framework Boosts Geo-Localization with Route Descriptions

Cross-view geo-localization matches ground-level observations (like a photo or video) against geo-tagged satellite imagery to determine location. Existing methods have used sequential queries such as video clips to capture richer spatiotemporal cues than single images. However, they overlook another complementary sequential modality: route descriptions — text that describes the same trajectory at a higher level of abstraction and is often the only input available (e.g., a user directing an autonomous vehicle to a pickup point). To bridge this gap, researchers from Tianyi Gao, Jiayu Lin, Danielle Beaulieu, and Nathan Jacobs introduce TrajLoc, a unified framework capable of processing both video clips and route descriptions for cross-view geo-localization. They also release the SeqGeo-VL dataset, containing ~39K video-text-satellite triplets.

TrajLoc leverages both dense visual semantics from video and abstract linguistic semantics from route text, enabling these modalities to mutually reinforce cross-view matching. A key innovation is TrajMod, a lightweight module that conditions query embeddings on trajectory geometry, producing spatially-aware representations that significantly improve accuracy. Experiments show that TrajLoc achieves substantial gains over state-of-the-art methods on both video and text geo-localization tasks. The work has been accepted to ECCV 2026. By combining visual and textual route information with trajectory-aware conditioning, TrajLoc could enable more robust geo-localization for autonomous systems when only verbal route guidance is available, or enhance drone navigation using natural language commands.