Integrating Orthoimagery into Feed-Forward 3D Reconstruction

Develop feed-forward 3D reconstruction methods that effectively integrate orthorectified satellite imagery (orthoimagery) alongside ground-level perspective imagery to enable consistent cross-view geometry estimation within a unified coordinate system.

Background

Feed-forward 3D reconstruction has rapidly advanced using ground-level perspective imagery, enabled by large datasets and strong backbones such as DUSt3R and VGGT. However, these models and datasets are predominantly designed for perspective cameras with 6DoF poses and aligned depths.

Orthorectified satellite imagery lacks perspective and parallax, making traditional multi-view stereo and structure-from-motion approaches unsuitable. Moreover, satellite imagery differs in resolution, appearance, and availability of camera parameters, complicating its integration with ground-level methods. The paper motivates bridging this gap and proposes Cross-View Splatter as a step toward leveraging satellite priors with ground imagery.