Empowering Feed-Forward Reconstruction Models with Metric Scale via Satellite Images

A team of researchers has proposed a method that uses satellite imagery to resolve a persistent scale ambiguity in feed-forward 3D reconstruction models, enabling them to recover metric-scale geometry without costly calibration or annotations. Feed-forward 3D reconstruction models can generalize across diverse scenes but typically recover geometry only up to an unknown global scale, a limitation that restricts their use in applications requiring metric environmental understanding [1]. The new framework, described in a paper submitted on 6 Jun 2026, integrates readily available satellite imagery as a global metric reference [1]. Given a coarse camera pose, the method retrieves a local satellite patch and fuses it with a reconstruction backbone through bidirectional cross-view interaction, enforcing consistency to infer absolute scale, refine scene geometry, and estimate camera pose in a metric coordinate frame [2]. Existing approaches to metric reconstruction often depend on large-scale metric annotations or accurate camera calibration, both of which are costly or unreliable in many real-world settings [3]. The proposed method relaxes the requirement for precise GPS localization, assuming only a coarse estimate sufficient to retrieve a satellite patch covering the surrounding area [3]. From this initialization, the framework jointly aligns ground-view observations with the satellite patch to infer absolute camera pose and metric scene geometry [3]. Experiments on the KITTI, nuScenes, and Oxford RobotCar datasets demonstrated consistent improvements in metric depth estimation, multi-view point-cloud reconstruction, and cross-view camera localization [2]. The model preserved strong generalization across datasets and geographic regions [2]. The work addresses a long-standing challenge by exploiting widely available satellite observations to provide complementary geometric cues, rather than depending on metric annotations or calibrated cameras [3]. This research builds on a growing body of work exploring satellite-ground cross-view geometry. A related feed-forward model, Cross3R, ingests a satellite tile alongside UAV or ground images to recover cross-view 3D point clouds and 6-DoF camera poses in a single forward pass [4]. Another recent effort, Sat3R, adapted a monocular depth foundation model to the satellite domain using RPC-aware metric depth fine-tuning, reducing mean absolute error by 38 percent over zero-shot baselines on the DFC2019 benchmark while delivering a speedup of more than 300 times over optimization-based methods [5].