GeoXel addresses two challenges in long-duration aerial reconstruction: cumulative trajectory drift and the lack of geographic alignment. It uses geo-referenced maps to supervise chunk-wise inference, test-time token adaptation, and trajectory refinement.
Abstract
GeoXel is a framework for geo-referenced 3D reconstruction from long-duration aerial video. Given a video stream and geographic priors including Digital Orthophoto Maps (DOM), Digital Elevation Models (DEM), and OpenStreetMap (OSM), GeoXel processes overlapping chunks with a transformer-based visual geometry model. Geo-referenced anchors connect visual reconstruction to global coordinates. A segment-local test-time training strategy optimizes detached aggregator tokens while keeping the pretrained visual backbone frozen, followed by intra-chunk and global pose-graph optimization for local and global spatial consistency.
Method Overview
Frozen Forward
StreamVGGT processes overlapping video chunks.
Geo Anchors
DOM matching and DEM elevation provide ENU constraints.
Token TTT
Segment-local aggregator tokens adapt to geography.
PGO
Intra-chunk and global graphs refine the trajectory.
Quick Start
Install the dependencies, prepare UAVScenes and a local StreamVGGT checkpoint, then run the public AMtown entry point.
UAVSCENES_ROOT=/path/to/uavscene \
CHECKPOINT=/path/to/checkpoints.pth \
bash scripts/run_uavscenes_amtown.shFrame index 200 · stride 5 · 32-frame segments · 8-frame overlap · TTT layers 17 and 23
BibTeX
Bibliographic details will be added when the paper is publicly available.
@article{geoxel,
title={GeoXel: Geographic Per-pixel 3D Mapping from Aerial Video in the Wild},
author={GeoXel Authors},
year={2026}
}