Conceptual

Robust Scene Coordinate Regression for Large-Scale Visual Localization

Improving scene coordinate regression (SCR) so it matches feature-matching pipelines for city-scale 6-DoF visual localization while keeping a compact map. Covers a covisibility-graph-based global encoding that conditions the regressor on jointly-visible map regions, augmentation with multiple global hypotheses for robustness to appearance and viewpoint change, and a depth-adjusted reprojection loss that supervises purely in image space to implicitly triangulate scene geometry without any 3D model or depth ground truth.