ECCV 2026

Trajectory-aware Cross-view Geo-localization with Sequential Observations

Localizing route-level videos and natural-language trajectory descriptions by matching them to geo-tagged satellite imagery.

Tianyi Gao1, Jiayu Lin1, Danielle Beaulieu1, and Nathan Jacobs1

1Washington University in St. Louis, Saint Louis MO 63130, USA

ArXiv Supplementary Material Code πŸ€— Dataset BibTeX
TrajLoc motivation: trajectory video or route narrative both retrieve the matching satellite tile Hover to read the abstract

Abstract

Cross-view geo-localization matches ground-level observations against geo-tagged satellite imagery. Recent methods show that sequential queries such as video clips yield richer spatiotemporal cues than single images, yet they overlook a complementary sequential modality: route descriptions—which capture the same trajectory at a higher level of abstraction and are often the only input available (e.g., a user directing an autonomous vehicle to a pickup point). To bridge this gap, we introduce SeqGeo-VL, a dataset of ∼39K video–text–satellite triplets, and TrajLoc, a unified framework capable of processing both video clips and route descriptions. By leveraging both dense visual and abstract linguistic semantics, TrajLoc enables these modalities to mutually reinforce cross-view matching. We further propose TrajMod, a lightweight module that conditions query embeddings on trajectory geometry, yielding spatially-aware representations. Experiments show that TrajLoc achieves substantial gains over state-of-the-art methods on both video and text geo-localization.

A trajectory can be queried as dense video or as abstract language — both should retrieve the same satellite tile.

Highlights

01 · Benchmark

A Setting Prior Datasets Missed

Cross-view datasets are single-image, video-only, or text-only β€” never sequential and linguistic together. SeqGeo-VL pairs route video, route narrative, and satellite imagery: 38,863 triplets, 91.8% human-verified.

Jump to section 02 · Method

Our Unified Approach

TrajLoc is adapted directly from the pretrained CLIP β€” no bespoke temporal architecture. Query with a trajectory, video frames, plain language, or any mix.

Jump to section 03 · Looking Ahead

From Spatial-Reasoning Benchmark to Agentic Tool

Egocentric-to-allocentric matching in outdoor scenes still stumps strong MLLMs, and explicit geometric conditioning beats prompting by a wide margin (+19.33 vs. +1.34 R@1%) β€” positioning TrajLoc as a specialized tool that agentic MLLMs can call.

Jump to section
Interactive · Try the Task

Can You Localize the Route?

Watch a route play out at street level while its narrative unfolds line by line, then pick the satellite tile it traverses. Every route is drawn from SeqGeo-VL's held-out validation split.

Ground-level view along the route
1 / 7

Route narrative

Which satellite tile does this route traverse?

MLLM agent

TrajLoc tool

Explore more: TrajLoc as an agentic tool

01 · Benchmark

A Setting Prior Datasets Missed

Existing cross-view datasets are single-image, sequential-video, or single-image-plus-text β€” never sequential and linguistic together. We extend SeqGeo with a progressive VLM+LLM annotation pipeline to close that gap.

DatasetTextSequentialView TypeQueriesReferencesObs.
CVUSAβœ—βœ—Panoramic~1.5M~1.5M~1
CVACTβœ—βœ—Panoramic137K137K~1
VIGORβœ—βœ—Panoramic105K90K~1
CVGlobalβœ—βœ—Panoramic134K134K~1
University-1652βœ—βœ—Limited FOV5.5K90K~1
GeoText-1652βœ“βœ—Limited FOV14K90K~1
CVG-Textβœ“βœ—Limited FOV30K60K~1
KITTI-CVLβœ—βœ“Limited FOV28K41K~4
GAMaβœ—βœ“Limited FOV~15M†1.9M—
SeqGeoβœ—βœ“Limited FOV118K38K~7
SetVL480Kβœ—βœ—Limited FOV480K16K~4
SeqGeo-VL (ours)βœ“βœ“Limited FOV118K38K~7

Obs. is the average number of observations (frames) per query. † As reported by SetVL480K.

Four-stage VLM and LLM annotation pipeline for SeqGeo-VL

38,863 video–narrative–satellite triplets

Built from SeqGeo with a four-stage pipeline: per-frame semantic captioning, motion integration from trajectory metadata, LLM summarization into a route-level narrative, and human quality control.

38,863video–text–satellite triplets
≈7frames per route
101.3avg. words per description
91.8%human QC pass rate

πŸ€— View SeqGeo-VL on Hugging Face →

02 · Method

Our Unified Approach

TrajLoc turns one pretrained CLIP ViT-L/14 into video, text, and satellite encoders with a two-stage curriculum: Stage 1 grounds the satellite encoder in video semantics; Stage 2 aligns route descriptions while regularizing against satellite-feature drift.

TrajMod, plug and play. In a final stage the encoders freeze, and a lightweight MLP maps route geometry Ο„ = {(Ξ”xα΅’, Ξ”yα΅’, ΞΈα΅’)} — waypoint offsets and headings, no map or POI metadata — to scale/shift parameters that modulate the query embedding, fβ€² = Ξ³ βŠ™ f + Ξ².
TrajLoc architecture: shared satellite encoder with video/text encoders and TrajMod modulation
Shared satellite encoder, independent video/text encoders, and TrajMod's trajectory-conditioned modulation.

State of the Art on Both Query Types

TrajLoc outperforms prior video- and text-based methods alike using the same ViT-L/14 backbone as its closest competitors, and beats an MLLM-based embedding model fine-tuned for this task.

Video β†’ Satellite

MethodBackboneR@1R@5R@10R@1%
SeqGeoVGG161.806.4510.3634.38
SeqGeo†ViT-L/148.1424.4234.0266.17
GARetDeiT-m3.3411.1917.1844.39
FlexGeoConvNeXt-B3.5114.2220.7748.53
Qwen3-VL-Embedding*Qwen3-VL-2B5.4419.2128.1561.16
TrajLocViT-L/1412.0935.7747.6880.21

Text β†’ Satellite

MethodBackboneR@1R@5R@10R@1%
CLIPViT-L/140.803.326.0122.99
SigLIPViT-L/16@3840.853.135.2421.33
EVA2-CLIPViT-L/14@3360.893.786.7026.99
Qwen3-VL-Embed.*Qwen3-VL-2B0.933.385.5520.91
CrossText2LocViT-L/14@3360.984.087.0825.45
TrajLocViT-L/142.529.8216.1145.48

* LoRA fine-tuned. † Reimplemented with a ViT-L/14 backbone.

Qualitative retrieval comparison with and without TrajMod
Top-3 retrievals with and without TrajMod. TrajMod yields layout-aligned candidates that match the route's actual turns and landmarks.
Multimodal Querying

Query with Video, Language, or Both

A simple blend of video and text query embeddings consistently beats the stronger single modality, with language helping most when visual evidence is sparse.

VisualVideo + SemanticLanguage → Ξ± = 0.4Blend
+9.49 R@10with only one video frame
R@10
FramesVideoTextBlendGain
130.1115.0839.60+9.49
344.9815.9750.31+5.33
548.0215.9952.79+4.77
747.7516.0652.50+4.75

Gain is measured over the stronger single modality.
f = (1−Ξ±)fvideo + Ξ±ftext · inference only, no extra parameters.

Recall vs. number of frames, with and without TrajMod
TrajMod's margin widens with more observations: Ξ”R@1 grows from 2.6 to 5.6 points as frames go from 1 to 6.

Geometry Makes the Modalities Cooperate

Anchoring video and text to the same trajectory geometry lets them reinforce each other instead of competing for capacity.

  • Co-training amplifies both tasks — with TrajMod, adding text co-training lifts video R@1 from 9.69 to 12.09, and video co-training lifts text R@1 from 1.90 to 2.52.
03 · Looking Ahead

From Spatial-Reasoning Benchmark to Agentic Tool

SeqGeo-VL · Benchmark

A spatial-reasoning testbed: can an MLLM connect egocentric route observations with their allocentric satellite footprint?

TrajLoc · Agentic Tool

The same capability, packaged as an API-callable service: an agent sends a route description and inferred trajectory, then folds the retrieved satellite locations into its broader reasoning.

Loading the agent result…

BibTeX

@inproceedings{gao2026trajloc,
  title     = {Trajectory-aware Cross-view Geo-localization with Sequential Observations},
  author    = {Gao, Tianyi and Lin, Jiayu and Beaulieu, Danielle and Jacobs, Nathan},
  booktitle = {European Conference on Computer Vision},
  year      = {2026}
}